I didn’t get rich inventing Ethernet. I got rich selling it!
-Bob Metcalfe

Models are moving faster than customers can change.

Some expect AI to fully transform the global economy in a matter of months. The technology is here. The narrative is here. Massive capex is being allocated throughout the stack. Yet, much of the real world will take a decade—or more—to meaningfully integrate AI.

Why?

Invention and diffusion are separate processes. Diffusion only happens after performance and economics clear a bar and when pressure to change exceeds friction.

Lasers & Inertia

A little while back, I stumbled across the story of how handheld laser welding came to SpaceX, as told by a former welding engineer.

The company had long used automated laser welding, but many use cases still relied on traditional handheld welding (e.g. small welds, welds on or inside large structures, in-place repairs).

One day, Elon saw a Chinese handheld laser on Twitter and bought one on Amazon. It sat unused. Even when a major equipment vendor released a handheld laser, SpaceX bought one and it still gathered dust.

Eventually one welding process had a high rejection rate. They tried it with the handheld laser and it worked perfectly, but it still wasn’t generalizable for longer welds until a key advancement arrived: automated wire feeding. Now, handheld laser welding was suitable for many more processes.

At this point, it’s tempting to assume adoption is assured because the tech works and the math mathed.

Yet, adoption stalled. The welding engineer had working prototypes and begged other teams to adopt them, but “everybody was too apprehensive to use it in production”.

Then one part’s weld was so difficult that only a single person at the company could complete it, creating massive backlog risk (a career-limiting move in Elon companies). Handheld laser welding, performed by a broader team, was the only plausible way to hit deadlines. The team tried the new process and successfully built their first production work cell.

Even then, the organization still wasn’t on board. The engineer continued his begging campaign to no avail. Nobody wanted to be the person whose new process caused a vehicle failure—even if that process was objectively superior on many metrics.

One day, Elon walked the line, looked at the welds and reportedly said, “This looks like garbage”.

Immediately, perceived risk shifted from “What if this new process fails?” to “Elon will totally fire me if I don’t fix these ugly welds.”

Suddenly, the welding engineer had a line out the door of teams ready to adopt the process: “Please. Let’s do this. Right now.”

Even though deploying handheld laser welding required upstream changes (including updates to design engineering assumptions), urgency overcame inertia. A number of bespoke laser-welding processes made it to production.

The results were better than anyone expected. Compared to prior processes, handheld laser welding reportedly reduced distortion by a factor of 10 and increased throughput by 5-6x.

Demand kept growing, so the welding engineer created a chart of pre-qualified procedures so he didn’t have to be involved with each new use case. Within a month, it exploded and they lost track of how many procedures were in production.

Handheld laser welding had won over traditional welding techniques.

Throughout all of this, notice that sufficient laser welding technology was merely the start of the battle. The larger blockers were a superposition of individual career risk and organizational inertia, only overcome with overwhelming external pressure.

Good market timing requires the combination of economically viable technology and where pressure overcomes organizational friction.

Trader vs Maker Market Timing

When people refer to “market timing,” they often conflate two different things:

• Trader Market Timing: whether it’s the right time to buy or sell an asset.

• Maker Market Timing: whether a specific customer set actually adopts a product.

Trader Market Timing matters for revenue multiples, but Maker Market Timing determines whether there’s any revenue at all.

Diffusion

Four factors control when customers actually buy products:

Performance: Does it work in the real world? Is it reliable enough across edge cases? Can it be operationalized at scale?

Economics: Is it viable when you incorporate all costs: compute, integration, compliance, training, and ongoing maintenance?

Friction: How hard is it to discover, decide, and deploy? Is it complex to communicate? Easy to try? Easy to implement at scale? Are results measurable and attributable?

Pressure: What causes urgent effort for or against change? Are there economic, regulatory, competitive or political forces pushing urgency? Will people lose their jobs if the product replaces labor? What’s the balance of career upside vs downside amongst decision makers?

Diffusion happens when net pressure overcomes friction, but only after performance and economics are viable.

Friction vs Pressure

Performance and Economics are relatively legible. Failures are easy to diagnose.

Friction and Pressure are illegible and can nuke even sophisticated teams. Determining whether a complex group of humans will actually alter their workflow to adopt the product requires deep industry context and more than a little wisdom.

Most AI “slow adoption” stories today are SpaceX handheld lasers. Model performance and price issues are easy to see, but products fail through some combination of Friction and Pressure (the latter often in the form of active resistance).

Asking ChatGPT to plan a trip is low-risk, reversible, and lives outside systems of record. Adoption is instant.

Automating health insurance claims adjudication touches payments, compliance, appeal rights, edge cases, and liability. It requires deep integration into brittle workflows and threatens roles in the existing claims ops team. Adoption is a slog.

Founders often assume that access to intelligence is sufficient. It is not. Simply granting an LLM access to a database does not fix a poorly documented workflow that intersects with three different compliance departments.

Ethernet and PCs

Consider 3Com, a company founded in 1979 to commercialize Ethernet. They began by building networking hardware for large UNIX systems. The technology worked but the business was meh.

In 1981, the launch of the IBM PC changed everything. Suddenly, millions of identical computers could share expensive peripherals (e.g. laser printers) rather than requiring separate equipment per computer. ROI was obvious. 3Com effectively “bet the company” on this new segment and rode the wave for years.

Metcalfe didn’t get rich because Ethernet suddenly got better in 1981. He got rich by paying attention to an increase in Pressure (strong motivation to avoid buying too many peripherals) paired with low Friction (low switching costs on a new platform) that made Ethernet obvious to a particular set of customers.

Maker Market Timing in AI

We are entering a period where AI model capabilities will outpace organizational absorption. Inevitably, there will be failures.

Blanket pessimism is a lazy coping strategy. Successful founders and investors will distinguish between “wrong tech” and “wrong time”:

• Wrong tech: pilots fail on usability/accuracy/latency/reliability/unit economics even with friendly customers.

• Wrong time: top of the funnel is slow, procurement stalls, integration stalls, users resist, renewals fail, and incentives misalign.

Startup success won’t come from watching model progress alone. It will come from identifying which handheld lasers are gathering dust, and what specific catalysts create extreme market pull that forces the organization to pick them up.

If you’re new, start with the first post and subscribe to be notified as new posts are published.

Applications are the most important area of the AI stack because products must win before other layers of the stack can become durable businesses. If no applications exist, other layers will have no long-term customers; strategic innovation spend can only carry a market for so long.

Recalling the central topic of this letter, it’s not a question of whether AI will benefit customers, but “How will AI companies achieve durable value?”. Put more bluntly: who will achieve power and how will they achieve it?

In past market cycles, defensibility has been a theoretical exercise for startups; competition was rarely the cause of premature mortality. As is often said in the Y-Combinator community, “more startups die from suicide than homicide”.

Given the speed of advancement and intense competition in the market, I believe homicide—death by competition—will be a significant cause of AI application startup failure.

With AI, this time is different: incumbents are stronger and disruption is more frequent

The “invest in products that are 10x better than alternatives” investment algorithm has historically made a lot of money for founders and investors for two major reasons, as we discussed in the opening of this letter:

1. Incumbents were either slow or nonexistent. The taxi industry was ill-prepared for Uber. Symantec was digesting Veritas as CrowdStrike rose to prominence. Neither Friendster or MySpace were entrenched incumbents during Facebook’s rise.

2. New disruptive tech stacks occurred more slowly than startup lifecycles. Historically, startups could build structural power before they faced their first Seldon crisis: the first time they need to catch a new technology wave. Netflix’s switch from DVD to streaming occurred long after their IPO. Facebook’s focus on mobile (via their Instagram acquisition) happened right before their IPO.

Today, neither are true.

“Big tech” companies are much better at execution than incumbents from a decade ago and have quickly integrated AI into their products. For the most part, tech leaders have strong leadership, strategies that embrace disruptive innovation, and many talented individuals at all levels of the organization.

In their 12.3 update, Tesla was able to migrate their autonomy architecture from a traditional computer vision model with “300k lines of explicit C++ code” to a modern AI stack without needing to deploy any major hardware changes.1 Google has neatly integrated AI results at the top of their search results page. Microsoft has added co-pilot features to their enterprise productivity suite. OpenAI (which I think now counts as big tech) has been able to successfully incorporate major AI advancements into ChatGPT for several years without changing their interaction model.

Some investors believe the vast majority of the value created by AI will go to incumbents, with little whitespace left over for startups. Regardless of who accumulates more value, I believe there are ripe opportunities for startup founders and investors to seize, as we’ll cover below.

Aside from incumbent competition, disruption timelines are compressed in this market cycle. AI is moving so fast that there may be a new tech stack every few years. As quickly as we remember technology moving in the past, AI is moving even faster.

Today, many AI products are “10x better” than the human equivalent. Enterprise co-pilots vs. legacy manual processes. AI marketing tools vs. agencies. AI-powered industrial robotics vs. manual work cells and expensive system integrators.

A 10x product today might take the form of a strong technical team building a carefully fine-tuned model or exceptionally engineered RAG system that provides far better results than an “off the shelf” models. A better AI legal assistant, a better AI software engineer, etc.

Here’s the problem: wave after wave of new AI tech stacks make today’s 10x into tomorrow’s table stakes. The next iteration of foundation models might be better than current 10x products out of the box. Once that happens, today’s mind-blowing products become commonplace. What happens when the Adobe, Microsoft, Google, and OpenAI build the feature into their leading suites? What was once 10x value where users would deal with the UX and procurement inconvenience of a new AI-first application turns into 1.5x value that doesn’t clear a customer’s activation energy. Why leave Gmail, Photoshop, ChatGPT, or Excel when you don’t have to?

To reiterate a point from earlier in this letter:

“If everyone else can use the component that enables 10x better product, there will be no resultant power.”

One way to avoid the trap of “point in time thinking” is to think backward, not forward. What does the world look like when all companies can build their products on top of superhuman AGI?

Founders and investors must be prepared for a world where AI products are not compared with humans, but with other AI products. In enterprise knowledge work, for example, we’re seeing the first wave of AI replace manual processes (“human is the loop”) with co-pilots (“human in the loop”). Soon, co-pilots will be the status quo and fully autonomous agents (“human on the loop”) will be the disruptive new stack. Over time, AI super-intelligence may be able to handle long-duration strategic projects that would have taken a whole department of humans. Super-intelligence would be yet another disruptive stack, replacing singular agent products.

Assuming a company can get enough momentum to be able to build structural power, what form should they invest in? Below, we’ll look at how AI impacts two common forms of power in technology companies: switching costs that disincentivize customers from leaving a product and data network effects that allow datasets to provide increased customer value with increased scale.

AI reduces switching costs.

Switching costs have been strong moats for technology companies: data, integration and process lock-in make it difficult for customers to leave. Historically, when customers store a large amount of data in one vendor’s product it was expensive to “port” that data to a competitor’s. Large hospital systems spend years and millions of dollars customizing new electronic health record (EHR) systems. Enterprises of all sorts incur tremendous costs integrating ERP systems (e.g. SAP, Oracle, Microsoft, NetSuite) such that switching becomes a prominent feature of employees’ nightmares.2 CRMs, similarly, contain a lot of historical data and a multitude of integrations accumulated over the years. Adobe’s PDF format and Microsoft’s various Office document formats create both switching costs for preexisting data and a network effect for sharing data between individuals and organizations.

AI makes everything fungible, reducing structural switching costs.

AI makes it much easier and more reliable to migrate data, integrations, and processes between platforms. What used to be a slow, manual, and error-prone process necessitating a small army of consultants can now be fully automated. There’s one exception: any time a user interface is tightly integrated into peoples’ daily work life, switching times are longer because people change more slowly than computers. The product design principle of MAYA—“most advanced yet acceptable”—is a speed limit to the rate of change groups of humans can tolerate.

Incumbents fight back. We should expect to see legacy vendors put up technical and contractual barriers to customers slurping data out of their systems as new entrants build AI-driven migration tools. I speak with countless startups who, when integrating with legacy enterprise software products, have tough negotiations to access platform data or API functionality. Sometimes barriers take the form of prohibitive “API access” fees or even blanket prohibition on anything other than manual “point-and-click” usage. Fortunately for startups, customers have leverage and can push legacy vendors to open up. Some large vendors have recognized this effect and embraced openness. Salesforce executed on an “open system of record” strategy effectively in the last market cycle: instead of trying to own all use cases, they became the integration point other sales and marketing tools depended on.

AI can have strong or weak data network effects, depending on the use case.

Questioning the orthodoxy for data always being a strong network effect seems like heresy given how often companies talk about their “data moat”. This is applicable both to application vendors who customize models as well as model-only vendors.

To understand why data is valuable for AI, it helps to first classify data into two key flavors:

• World data is used to train models to better understand how our universe works. Internet content, experimental protein folding data, emails, labeled images, vehicle driving sessions, and human feedback on prior AI generated responses all help AI models get more intelligent: to make their output less random and more desirable.

• Context data is used to provide background information about a particular user, organization, or situation to produce better results than “one-size-fits-all” output. A user’s current location and historical reservations when asking for a restaurant recommendation. Prior emails when composing a new message. Existing marketing collateral when designing a new ad campaign.

To distinguish between world and context data, we must ask whether the data contains “new information” about the world as a whole or whether it tells us new information about a particular person, organization, or situation.

Data vs. new information is a subtle, yet important, concept. Any form of raw input is data, e.g. text, images, audio and video. Information is the meaning extracted from that raw data. Consider the following simple dataset:

1. Sally owns a red convertible.

2. Sally owns a black truck.

3. Sally owns a red convertible and a black truck.

If we have data point #1, then adding #2 tells us new information: that Sally also has a truck. If we add data point #3, we’re adding new data but no new information. Data points #1 and #2 improve the model’s understanding of the world, but #3 does not. When we add data but no new information, the additional data holds no value.3

Put another way, new information is the only form of valuable data. If data tells us something we already know: it’s likely useless.

As this applies to world data: when we already know a lot about the world, it becomes increasingly difficult to find things we don’t already know. For randomly-gathered datasets, novelty becomes increasingly hard to come by. Such data has “decreasing marginal value”—each additional piece of data is less valuable than the last. We’ll get into a specific example as this applies to robotics below.

To understand context data, let’s consider another data point:

4. Sally is currently driving her red convertible with the top down.

While #4 doesn’t tell us too much new information about the world, it does tells us what Sally is doing right now. When asked for a restaurant recommendation, a model can infer that it’s probably warm outside and Sally is more likely to want to eat outdoors than if the convertible’s top were up. This contextual information is very valuable to help a model produce the right results without Sally having to explicitly ask for restaurants with outdoor seating.

Context data is only valuable in a limited scope. In a month from now, knowing that at one point Sally had her convertible’s top down doesn’t tell us about what sort of restaurants she might want then. Similarly, knowing what kind of car Sally is driving is unlikely to help a model make recommendations for Sam, a completely unrelated person in a different part of the world. In order for a source of context data to remain valuable, it must be fresh and relevant to the situation a model is reasoning about.

Both world and context data produce value for customers because they improve the quality of a model’s output. This implies that there is some form of network effect, but the network effect plays out differently for each:

• World data has a global network effect: all users benefit from data generated by all other users. This can lead to power at a global scale: early leaders have the best data and therefore an opportunity to build best product. They can attract more users, generate more data, build an even better product, and so forth.

• Context data has a local network effect: only the user whose context is present benefits from the data. This tends to advantage incumbents who have existing customer relationships where they can build new products on top of existing context data. New entrants must invest not only in a new product but also a source of context data; however, when context can be gathered via API (e.g. Email, Calendar, financial transactions) or via a mobile device (e.g. location, activity tracking) that incumbent “cross-sell” data advantage becomes smaller.

In both cases, each additional data point generally has diminishing new information and therefore, diminishing additional value. The most important question when analyzing data defensibility is: when is a dataset sufficient? When do we reach a point where we stop caring about more data.

Let’s look at another example of AI for robotics where data directly translates into improved operating characteristics. In systems engineering, one way to measure reliability is the “uptime percentage”. The shorthand engineers use is to “count the 9s”:

• 99% (two nines) =~ 3.7 days of downtime per year

• 99.9% (three nines) = ~8.8 hours of downtime per year

• 99.99% (four nines) = ~53 minutes of downtime per year

• 99.999% (five nines) = ~5.3 minutes of downtime per year4

• 99.9999% (six nines) = ~32 seconds of downtime per year

In AI for robotics, at what point does the reliability of a model reach sufficiency such that further improvements no longer meaningfully impact customer satisfaction? Assuming there are no safety issues5, the number of nines to reach sufficiency depends on the use case:

• Fully autonomous (level 5) passenger vehicles probably need 5-6 nines. We’re willing to tolerate a few minutes of stoppage during a snowstorm, but hours would be infuriating.

• Other robots out in the world probably need probably 2-3 nines. It’s easy for such robots to stop while remote pilots can take over difficult situations: 0.1-1% human piloted time adds very little to a cost structure.

• Factory automation needs perhaps 3-5 nines, depending on the criticality of the process. Manufacturing lines are expensive to stop, but if stoppage is infrequent it doesn’t meaningfully impact economics of the factory as a whole.

Now, given we have reliability targets, how much data do we need to achieve each? It depends on the complexity of the problem and how good our data is. If we’re taking a random sample of data from the real world, adding additional 9s requires increasing amounts of data. For example, moving from 99.9% to 99.99% might require 10x the data because adding that extra 9 requires training on very rare circumstances that, by definition, are infrequent in naturally occurring datasets.

Because data and training cost money, at some point there’s a breakeven point where the cost of training on additional data is more than the customer value unlocked by improved model reliability. If we’re in a circumstance where models can get to sufficiency with just few 9s, this might not be much data at all!

Let’s not stop there: not all data is created equal. In circumstances where a system’s failure modes are predictable, we can supply carefully-chosen or synthetically-generated training data to capture edge cases and achieve many nines of reliability with even less data than a random sampling of real world events would require. With carefully chosen training data, data network effects in narrow applications might be even smaller than most would predict.

If a model can get to sufficiency with little data, then there’s no data network effect.

On the other hand, general purpose AI systems must know about everything in the world. Getting to “good enough” requires a colossal amount of data. General purpose AI applications seem to have a large data network effect.

For founders and investors, we better know how much data we need before reaching sufficiency. Betting on a data network effect where none exists instead of investing in other forms of power can be a terminal error. Once models reach sufficiency, power is not based on data, but other effects: e.g. economies of scale, cost, distribution, brand, and execution.

$100 trillion of global GDP is up for grabs outside big tech’s blast radius.

I believe AI will be a disruptive tech stack in every industry. This is great news for startup founders and investors because it means there will be opportunities to build durably valuable companies outside of the immediate roadmaps of incumbent tech leaders.

Large technology companies who are very good at incorporating new advancements are scary competitors. In contrast, companies in “sleepy” industries accustomed to only gradual technological change present softer targets for startups to overtake.

When entering traditionally “non-tech” markets, startups need to decide how they approach the market and what type of offering to sell.

Sell to or compete with?

When entering an existing market, startups must decide whether incumbents are customers or enemies. Sometimes the answer is obvious because there’s only one viable path. Much of the time, though, there’s a complex series of tradeoffs. Some key questions to ask include:

• What are sales, marketing and implementation costs of selling to incumbents? Do they have experience integrating technology? If not, a startup will probably face long sales and integration cycles leading to poor unit economics. It might be easier to compete.

• Are there regulatory barriers? Healthcare, legal, and financial services, for example, are all markets that have so much regulation that it’s not realistic for a single company to dominate the entire market with a disruptive product. Unsurprisingly, startups in these sectors tend to either sell to incumbents when their product is broadly applicable (e.g. generalist AI legal tools) or compete with incumbents when it is narrowly focused (e.g. AI-powered specialty law firms).

• Is the market fragmented or consolidated? At one extreme, selling into a consolidated monopoly or oligopoly market is dangerous because there are a tiny number of customers. Economists call such a monopoly of demand a “monopsony”. Airbus and Boeing are unlikely to cede power to suppliers because they’re the only games in town for large commercial passenger planes. Unsurprisingly, autonomous aircraft startups are competing with incumbents, starting with smaller aircraft. A fully fragmented market can also be difficult to sell disruptive products into because deal sizes are so small—often better to compete there as well. There’s a goldilocks zone for “sell to” where deal sizes are large but no individual customer can push us around.

• Scale of customer impact? In industries that require hundreds of millions or billions of dollars of capital expenditures (e.g. steel, semiconductors, mining, aviation, railroads, telecom, chemicals) a company must have a transformative tech stack that massively impacts economics to raise the required capital. Startups increasing profit by a small amount in chemical manufacturing might have a large market when selling across segments, but the impact is probably not a large enough difference to justify the risk of building their own plants.

• What are end customers’ switching costs? Historically, these are high for any product with workflow integration or lots of data. AI lowers these costs, especially if users are willing and able to learn a new system. It’s an especially interesting opportunity when existing users abhor the user experience of legacy products.

• How will we eventually gain power? Startups should expect incumbent customers to either try to duplicate their products internally or switch to a cheaper competitor if possible. Is there a path to power?

“Compete with” is a viable option in a lot of markets because AI can substantially reduce costs while providing more value to customers. Expect large rewards for startups outside of what we traditional consider “technology” for those that successfully execute and build durable power.

Sell outcomes, not software.

While there are still some a number of recalcitrant industries that use pen and paper or ancient software, many markets are software-saturated.

It’s tempting to dismiss new software companies as squeezing water from a rock. After years of skilled teams looking for opportunities in every corner of the economy, many buyers have software fatigue. Overbuying during the good times is leading to consolidation and cost cutting in the lean times.

At the same time, many companies are still a mess. Internal processes cost too much, take too long, fail frequently and provide little differentiated value to the organization. Instead of selling software to support these workflows, AI allows startups to sell workflows as a whole that adapt to each customer’s unique environment.

Just as the global economy benefits from division of labor of humans, the unbundling of companies will lead to division of labor amongst organizations. There’s no reason most companies need to run undifferentiated workflows when they can pay someone else to run the process better and cheaper. An early version of this is the migration from internally-run datacenters to cloud computing. Very few organizations can run a datacenter as effectively as Amazon, Google, or Microsoft so they shouldn’t bother. As a result, companies can focus on applications “at the top of the stack” where they can provide differentiated value.

Already, there are a number of startups successfully selling AI-enabled outcomes across functions: IT, security, marketing, HR, legal, finance, sales, and engineering.

In consumer markets, I suspect a similar specialization of labor will occur: more situations where we can pay someone to do something we don’t want to do.

When selling outcomes, companies can create high switching costs by investing in human relationships (i.e. sales) and integrate into multiple workflows. Once embedded at multiple points with a customer, a company ceases selling software and starts selling “forever-ware”.

When selling outcomes, improvements in AI technology are likely “evolutionary” because the product we deliver the customer remains the same. We can cleanly integrate AI advancement into our offering. Transitioning from co-pilots to agents doesn’t change what we sell, it just improves our gross margin.

Will AI application revenue justify the hype?

Unlike some technology booms where the “story:reality ratio” never gets right side-up, I believe AI will deliver on its promise to transform the global economy. Even today’s embryonic AI products deliver tangible benefits (making things better, faster, cheaper, and/or more reliable).

Consider use cases where both startups and incumbents are successfully employing AI either in production or soon-to-be-production environments:

• Industrial robotics

• Autonomous vehicles

• Software engineering

• Legal services

• Customer support

• Travel

• Marketing

• Healthcare

• Biology

• Media

• PsyOps

The AI cynics point to all sorts of use cases where AI isn’t yet good enough to supplant current solutions. Of course, there’s a very reasonable clear-eyed cynical argument against the blinded optimism that some AI boosters profess: AI investments are currently outstripping sustainable revenue.

Ultimately, I believe it’s a matter of timing. Given the trajectory of progress, I believe there will be no such thing as an AI industry. All companies will need to embrace AI or they will get smoked by others that do.

To support the optimistic case, very few markets have reached “sufficiency” with respect to foundation models, so every model improvement makes AI products even better.

That said, value is created by particularly talented teams who are dedicated to serving their customers. Not by simply showing up and waving our hands, shouting “(something something) AI”.

“…rapid puncturing of a share price bubble is typical of what happens during a phase in which market valuations are driven more by themes and concept stocks than by profits, dividends and other fundamental considerations. It has been repeated many times…”
- Alasdair Nairn

The best companies will build something people want—catch a wave and keep surfing it—until there’s another wave. Then use their execution power to catch that one.

Speed, rainmaking, and taste. That’s power.

Previous Post

Defensibility
A series of posts dedicated to answer the question: Where will value accrue in AI?

1. Introduction: Why this time is different.

2. Technology markets are organized around “stacks”.

3. "Execution power” is becoming more important than classical “structural power”.

4. Market revolutions occur when “critical" technology makes a new stack “viable”.

5. When multiple stacks become viable in rapid succession, companies must “AND” or “OR”.

6. Power within the AI stack—hardware, hosting, models, and infrastructure.

7. Power in AI applications—big tech, switching costs, network effects, and the $100 trillion of global GDP up for grabs.

If you’re new, start with the first post and subscribe to be notified as new posts are published.

Enough with principles. Let’s now examine each layer of the AI stack to ask whether customers will be able to switch vendors freely (a commodity market) or where they might prefer a premium vendor. We’ll bring in historical examples that can inform market dynamics; however, we should always be wary of overfitting.

“History doesn’t repeat itself, but it often rhymes.”
- Mark Twain (attributed)

To have a chance at predicting how things play out, we have to pay close attention to:

• Technology markets—where snowflakes cluster into snowballs.

• Leaky abstractions—where build vs. buy decisions require detailed technical knowledge of the component in question and the psychology around it.

• Power—where a company maintains margin amidst competition because of something they control while excluding others from the same advantage.

• Waves of disruptive technology—where a new stack becomes viable to deliver superior customer value, while watching out for additional waves of disruption to come.

• Sufficiency in existing stacks—when advancement in a technology no longer meaningfully impacts the market.

Compute Hardware

In modern AI systems, there are two phases that each require compute: training and inference. 

Training is the process of converting a model from a blank slate—a child—into something useful. Currently training is done in datacenters. Much like the education of human children, training state of the art models is really expensive1 because it requires a colossal amount of data and compute. Algorithmic enhancements will likely decrease compute requirements for a given level of model performance. however, the demand for better models will lead to a net increasing demand for training compute. We’ll dig a bit deeper into the economics behind model performance demand in the foundation model section.

Once a model is ready to do work—inference—it still requires compute. In many cases, inference requires less compute and can be done both in datacenters and the field; however, recent advancements in “chain of thought” models require much more compute at inference time leading to datacenters being a primary execution venue.

I believe datacenter vs. local compute will be the primary segments of the compute hardware market. Let’s dig into each individually:

Datacenter—Training and Inference

NVIDIA is currently king with the best hardware and a software network effect (e.g., CUDA). Compute, however, is a commodity there are many working to improve the software of other hardware vendors. Prolific engineer George Hotz, originally known for his iOS jailbreak work, is actively building software tooling to enable modern AI stacks to run on non-NVIDIA GPUs, as are many other developers. Similarly, Mark Zuckerberg recently made a not-so-subtle remark about their hybrid GPU infrastructure plans:

“By the end of this year, we’re going to have around 350,000 NVIDIA H100s. Or around 600,000 H100 equivalents of compute if you include other GPUs.”
-Mark Zuckerberg, 1/18/2024 

While NVIDIA’s lead is formidable, I believe it would be incorrect to apply the analogy of Intel’s longtime desktop processor dominance to AI models. Unlike compiled software, in which executables target a specific CPU architecture (e.g. Intel’s x86), models are generally portable across GPU architectures. Additionally, Intel had a very long tail of CPU buyers, while with datacenter AI compute, there are a few very large buyers who make up the majority of the market—a much weaker network effect.

“Sales to one direct customer, Customer A, represented 13% of total revenue and sales to a second direct customer, Customer B, represented 11% of total revenue for the first quarter of fiscal year 2025, both of which were attributable to the Compute & Networking segment.”
NVIDIA 10-Q, filed 5/29/2024

While other customers are disclosed to be under 10% in the quarterly report, Bloomberg and Barclays Research are reported to have estimated that for the 2024 fiscal year, NVIDIA’s top four customers (Microsoft, Meta, Amazon, and Alphabet) represented approximately 40% of total revenue. It’s probably a reasonable bet that all of them are looking furiously for ways to reduce their datacenter GPU costs.

How do datacenter GPU users make build vs. buy decisions? In some cases, developer familiarity with a platform (network effect) may drive product selection. In other cases a team may carefully evaluate price/performance of alternatives, where price is a “total cost of ownership” (TCO) calculation including both hardware and energy. Viewed through the lens of the stack, datacenter GPU buyers are “above” compute hardware vendors and can “swap out” the compute component in their stacks by directing bucks at any GPU vendor they choose. Either way, datacenter compute hardware is likely to remain a critical technology in many stacks because model compute requirements keep going up.

In the case where GPU buyers are cost-optimizing over all else, can NVIDIA remain the only game in town? I suspect the answer will be a function of the durability of process power given NVIDIA’s deep experience in GPU architecture and execution speed in adapting to changing technical requirements driven by changes in underlying AI model architecture.

Local—Inference

Locally-hosted AI will become a significant market that plays out differently between end-user devices (e.g. PCs/laptops, tablets and mobile phones) and embedded devices (e.g. cars, household and industrial robots).

Regarding end-user devices, consider the recent Apple Intelligence announcement showcasing the power of models running on existing hardware devices. At the time of writing, I can run a 70B parameter language model on my Apple laptop’s M3 chip perfectly. While it’s possible for a new category of AI-first end-user hardware (e.g. wearables) to emerge, I suspect local AI to primarily drive existing hardware refresh cycles. There will be no distinct end-user AI hardware market unless some new device form factor takes off.

In the pre-AI world, embedded devices generally had minimal compute requirements and used low-cost commodity processors or microcontrollers. Self-driving vehicles, humanoid robots, and manufacturing automation robots that utilize modern AI models need a lot of compute. I believe most of these devices will locally host their models given requirements for low-latency inference (self-driving cars can’t wait to brake because of a slow connection) or where they cannot rely on continuous internet connectivity (factories can’t afford to shut down their production lines if the Internet goes out).

In embedded computing, I believe the market will bifurcate depending on whether the use case has a fixed requirements (e.g., manufacturing) or whether it has induced demand for additional intelligence (e.g., real world robotics).

In fixed-requirement embedded use cases, device makers are “above” compute hardware on the stack. These device makers will choose a hardware vendor that provides enough compute to deliver value to a device buyer even further up the stack. Because the device buyer in a fixed-requirement use case may only care about a narrow set of operating characteristics (e.g. cost, speed), compute hardware may end up as a complete commodity.

Consider the case of a manufacturing robot that uses AI to perform manufacturing steps previously performed by a human with minimal configuration. A factory owner is unlikely to pay extra for a robot with the ability to compose a sonnet worthy of Shakespeare vs. one that just does the job sans poetry.

If the requirements for a job-to-be-done are clear and fixed, device makers will run a simple cost calculation and the AI hardware vendor that provides sufficient compute at the lowest price is likely to get the sale. In fixed-requirement embedded stacks, compute hardware will eventually become sufficient.

In embedded use cases with induced demand for intelligence, requirements will rapidly change for the foreseeable future as models become increasingly capable. Because of this flux, embedded device makers need the ability to respond to changing customer demands. 

For example, once humanoid robots can vacuum our floors, we’ll immediately want them to do far more complex physical and intellectual tasks such as re-organizing the closet, cooking dinner, or hand-washing delicate crystal wine glasses.

I believe that embedded device makers in “infinite intelligence” markets will evaluate price/performance tradeoffs between compute hardware vendors but will pay a premium for anything that helps them iterate faster. Compute hardware advancement massively impacts how they can deliver customer value and will likely remain critical for some time to come.

Hosting—Who runs the hardware?

If a company runs its own models2, they have to run them somewhere. After buying hardware, they need to keep it at a reasonable temperature and feed it a lot of electrons. This is a lot of work. Some companies will just have somebody else deal with it (e.g. paying AWS, GCP or Azure to host hardware for them) while others will choose to build out their own datacenters if they can justify it for cost or political reasons.

Given the capital required, I believe only cloud providers, sovereign nations, and other large organizations (e.g. xAI, Tesla) will be able to build datacenters while almost everyone else rents hardware from public cloud providers. Given these forces, I believe AI will serve as just another factor in choosing public cloud providers for the vast majority of companies and there will be no change in market structure.

There’s one exception: local inference. I believe two models will emerge: per-device compute where each hardware device is self contained and onsite “micro-datacenters” that aggregate compute loads from multiple devices. So long as compute hardware remains a relatively costly component of embedded systems, allocating compute from a large pool is more cost effective than having a lot of expensive, partially idle hardware sitting around.

Foundation Models—What runs on the hardware?

Foundation models—the configuration of parameters that transform input into hopefully useful output—are one of the most interesting areas of the stack. At time of writing, the most advanced foundation models contain trillions of parameters that are set during training.

OpenAI with its state of the art foundation models is generating billions in revenue. At the same time, as Gavin Baker of Atreides Management has quipped, “Foundation models… are likely the fastest depreciating assets in human history”.

When a new foundation model comes out that’s superior in terms of absolute capability or capability/cost ratio, then all existing models become less valuable. Models that can cost over a billion dollars to train can become nearly worthless in a single competitive release.

How can we figure out which foundation model vendors will win? First we must dig into what investments improve foundation models before understanding what might be defensible.

There are at least a few ways foundation models get better:

• more compute—relatively straightforward: more capital begets more/better hardware.

• better algorithms and infrastructure—a strong research and development program that can both perform novel research and implement external findings.

• better datasets—proprietary data that can be used to train models.

Algorithms and infrastructure confer temporary advantage because, secrets don’t stay secret for long given the spirt of academic collaboration and the tight networking within the AI research community. Even the most secretive findings may not remain proprietary for long given how difficult it is to exclude competitors from using them. 

Consider Eli Whitney’s cotton gin. He started manufacturing in 1793 and was granted a patent in 1794, but was not able to produce enough machines to meet demand (in no small part due to a fire in his workshop). After others observed the key insight—mechanical separation of cotton seeds from fibers—many produced bootleg, unlicensed cotton gins. Whitney, despite patent litigation, was ultimately unable to make much from his work that transformed the economy of the American South3. Not all valuable inventions make money for their inventors.

In order to maintain a durable algorithm and infrastructure advantage, I believe speed of execution is the most viable approach. Foundation model developers must invest large sums of money to stay ahead given 7-figure salaries for top AI researchers: not a cheap proposition.

Better datasets can make a given model perform better on real-world tasks, much like humans perform better after learning from a great teacher or reading an exceptional textbook. Some datasets are widely available (e.g. internet crawling) while others are proprietary. These proprietary datasets can come from another business the company operates (e.g. Gmail, Microsoft Outlook), data they license from copyright holders, dedicated human data generation efforts, and feedback from end users (a process known as Reinforcement Learning from Human Feedback, RLHF). 

A company that maintains a data advantage, all else equal, will see their models perform better than competitors. It’s unsurprising to see both incumbents and startups claiming their proprietary datasets as key elements of their defensibility. We’ll explore the dynamics of how data impacts performance on particular use cases in the application section below.

One potential problem for companies aiming to maintain a data advantage. Researchers without access to massive datasets, being clever people, found out that they can siphon data from other models by asking them to generate data that can then be used to train other models, a process called Reinforcement Learning from AI Feedback (RLAIF). While foundation model vendors prohibit this process4, it may end up being impossible to stop because provenance of knowledge is very hard to detect once it is encoded in the weights of a neural network.

Given all of this investment in foundation models, we need to ask: who are the customers of these models and what do they want? In some cases, they need the most possible intelligence to push the boundaries of what AI can do while in other cases, they need “just enough” intelligence to solve a given problem.

As a result, the market for foundation models will bifurcate into “frontier” and “sufficiency” models. 

In stacks where additional intelligence unlocks more customer value, frontier foundation models are a critical technology and will likely remain so for some time. Customers will continue paying for the best available frontier models. 

For fixed-requirement use cases, sufficiency foundation models will eventually be evaluated on cost and will no longer be a critical technology.

This bifurcation is already playing out: for example, OpenAI currently sells access both to its frontier models and its “turbo” models that are less intelligent but far cheaper. For many use cases, turbo models are good enough.

Frontier models will consolidate.

With huge capital expenditures required to stay ahead in the frontier model market, there will likely be economies of scale that drive down cost of capital and amortize large research and infrastructure investments. This leads to consolidation.

There are countless examples capital intensive technology industries consolidating in the past that I believe will repeat in frontier models:

• hundreds of oil refineries before Standard Oil dominated

• hundreds of automakers before Ford, GM and Chrysler dominated

• hundreds of PC makers before IBM, Apple, Compaq and later Dell and HP dominated

It’s very hard to be small when cost of capital is high and one has to tap equity markets for large and growing hardware and R&D expenses.

Will open or proprietary frontier models win?

If training frontier models remains extremely expensive, then the question is: "how do we pay for them?”. Independent companies (e.g. OpenAI) have, recently, kept their frontier models proprietary and charge both developers and end users to access them while Meta has released their “Llama” series of models publicly. Meta’s decision to provide state of the art frontier models is perhaps the most disruptive decision in the industry at the moment. They cannot rely on foundation models provided by others so they needed to build their own. Given they monetize on the services layer and were unlikely to build an infrastructure business, it makes a lot of sense for them to open their work and have a chance at owning the ecosystem. Meta’s open model AI strategy has a second-order effect of reducing the revenue generated by its competitors. 

Looking at the operating system market may be a good analogy to predict what might happen:

• Linux: open, where revenue is mostly generated by other layers of the stack, not by Linus Torvalds.

• Android: open, but monetized with a services layer.

• Windows: proprietary, sold directly and as a component of others’ products.

• iOS/MacOS: proprietary and vertically integrated to deliver a superior user experience.

There’s a question of whether Meta’s AI strategy will result in Linux or Android. Will OpenAI stay Windows or will they eventually open their models (Linux/Android) or perhaps go the hardware route (Apple)?

Both the business model and AI safety debates around open vs. closed frontier models will, at the very least, be fascinating to watch.

Open sufficiency models will win.

Because sufficiency models eventually become “good enough”, there becomes a point by which further improvement doesn’t matter. In these circumstances, I believe the most likely outcome is that developers will choose the model with the lowest total cost (including fine tuning, license fees, and compute cost) that meets their needs.

It’s possible that highly optimized proprietary models with great developer experiences win, though I suspect that over the long term, open models will reach parity and become fully commoditized.

Should we go vertical?

In some cases, models are only applicable in a narrow set of circumstances and the best business model might be to be both a model developer and an application developer. The key question here is whether there will be a robust market for the model (it shows up in lots of stacks) or whether it’s really only one stack and the application vendor holds most of the power.

Software Infrastructure—How do we tie it all together?

AI infrastructure includes all of the supporting tools a company deploying AI needs. Examples include Data management, Training/Fine-tuning tools, Deployment platforms, Observability systems, and Security products. Given the maturity of prior software infrastructure markets, both buyers and vendors have been able to make reasonable guesses at what the product requirements are likely to be in each category. Unsurprisingly there are a litany of startups going after each.

As a note, “Infrastructure for AI” is different from “AI for Infrastructure”5. The former includes tooling for companies deploying AI and the latter is an application of AI to solve broader infrastructure problems. There are many great opportunities to use AI to improve infrastructure (e.g. cybersecurity, reliability engineering) that are applications of AI covered in the next section. This section is specifically focused on Infrastructure for AI.

Application developers are rational actors who critically evaluate AI infrastructure investments.

If we play “follow the buck” in AI, money starts at the top of the stack and flows downward. Application developers have money and they’ll make build vs. buy decisions regarding infrastructure tools. They will, of course, evaluate price vs. performance of various options. 

Vector databases are a good example. When an application needs to supply contextual information to a large language model, they commonly deploy a Retrieval Augmented Generation (RAG) architecture. The “generation” part is a foundation model and the “retrieval” part is often a system called a vector database. These databases allow us to take some input data and identify conceptually-related data so that both can be provided to a model to make a decision. For example, if a financial analyst wants to know about last quarter’s IT budget breakdown by vendor, a RAG system might identify the relevant reports with raw data and supply them alongside the analyst’s question to a model to compose a coherent answer. 

There are many different options for vector databases and it’s unclear how many customers will pay the premium for a proprietary vector database rather than use the free one built into the wildly popular postgres database (pgvector) that’s easy to spin up on AWS.

Infrastructure is a siren song for founders and investors. 

One lesson many took away from the last market cycle was that infrastructure companies were some of the best businesses: consider the stellar performance of DataDog, CrowdStrike, Palo Alto Networks, Snowflake, ZScaler and others. The missing part of that conclusion is that the current crop of infrastructure winners stand on the shoulders of decades of investment in IT systems and cloud software. Many of these winners re-segmented existing markets (infrastructure/application monitoring, anti-virus, firewalls, databases, and security gateways, respectively) so they could tap into long standing (and long growing) IT budgets.

As an engineer, building infrastructure is also tempting because we get to sell to people who think like us and our customers mostly care about performance characteristics. Building to performance specs is much easier than convincing a whole bunch of people in the world to change their behavior (very hard!).

AI infrastructure is a new budget area, with investments being pulled forward.

Unlike the environment when the current infrastructure winners started, the AI stack is more or less getting built from the ground up from new budgets. One lesson learned from the dot com boom: innovation budgets get cut if investments come too far ahead of revenue.

Consider VA Software—a leading provider of Linux servers—before, during, and after the dot com boom. Of course customers stop buying infrastructure when they don’t have much revenue and investment capital stops flowing freely.

While AI application revenue may quickly catch up to AI investment, there’s a possibility where it doesn’t: investment capital dries up and we see an AI infrastructure winter. 

There’s an oft repeated narrative that best way to make money in a gold rush is to sell picks and shovels. This might be true when there are far more miners than shovel makers; however, when every store on the block sells shovels, they’re not great businesses.

The early crypto world was a similar a picks-and-shovels rush, with many anticipating that real applications (other than speculation) would soon thrive and kill traditional “centralized”  software. In a world where customer behavior could shift overnight, this isn’t a crazy belief; however, the MAYA (Most Advanced Yet Acceptable) principle tells us that there would be a high likelihood that applications on the crypto stack were just too new to gain rapid adoption. People didn’t know how to manage wallets and they weren’t ready to think about a separate (high volatility) currency for every application they wanted to use.

Today, there is a dangerous combination of a lot of AI software infrastructure vendors and not that much money being generated at the top of the stack by AI applications. There are perhaps more shovel markers than there are miners. Miners aren’t dumb and they won’t pay more than they need for shovels over time. 

For large companies, over-investing in AI infrastructure ahead of revenue is rational. Being late to the AI party is an existential risk while being early is a few extra billion in innovation spend they have to write down. With that in mind, infrastructure providers who are banking on this spend would do right to carefully manage expenses and cash if the music suddenly stops.

Timing the AI infrastructure market: Being early is indistinguishable from being wrong.

Over the long run, there will be great businesses built. The question is: Who will win? Today’s AI infrastructure companies or the successors who sprout from their graves?

Founders and venture investors (myself included), when confronted with a secular trend are commonly correct on direction but wrong on duration. Many things just take a long time to play out. The key figure to watch is whether there is enough revenue being generated by application developers and enterprises deploying AI to support a robust infrastructure sector.

Today, AI applications are being deployed at mach 10, which we’ll cover below. Perhaps this time really is different from the dot com and crypto booms. In the infrastructure bull case, AI application revenue will grow extremely quickly and substantiate infrastructure spending before innovation budgets and patience run low.

As with any decision regarding founding or investing in a company with multiple future scenarios, we must underwrite both the bull and bear case futures and adjust expectations accordingly.

To predict the fate of infrastructure and lower layers of the stack, we must the key question of: “When and from where will application revenue come?”.

Next Post

Previous Post

Defensibility
A series of posts dedicated to answer the question: Where will value accrue in AI?

1. Introduction: Why this time is different.

2. Technology markets are organized around “stacks”.

3. "Execution power” is becoming more important than classical “structural power”.

4. Market revolutions occur when “critical" technology makes a new stack “viable”.

5. When multiple stacks become viable in rapid succession, companies must “AND” or “OR”.

6. Power within the AI stack—hardware, hosting, models, and infrastructure.

7. Power in AI applications—big tech, switching costs, network effects, and the $100 trillion of global GDP up for grabs.

If you’re new, start with the first post and subscribe to be notified as new posts are published.

As we’ve seen, multi-wave markets occur when companies don’t have long to scale one tech stack before another disruptive stack is able to better serve customers. Because of the cascade of disruptive new stacks on the horizon, companies must be prepared for a possibility where they can only surf a given wave for a year or two before the underlying technology shifts. When this happens, they’ll need to build a new tech stack. In contrast, single-wave markets occur when waves of disruption are widely spaced: so far out as to be strategically irrelevant at the moment.

Most examples of disruptive technology in the past were single-wave markets, so few will be able to “pattern match” their way to success in multi-wave markets.

There are two potential strategies when facing a multi-wave market:

• “AND” — catch each consecutive wave1, paddling like hell to continue winning as each new disruptive stack becomes viable. Be ready to re-imagine products multiple times.

• “OR” — pick the final foreseeable wave of disruption and put all of our energy into beating incumbents.

There’s no universally best strategy; each market will present a unique set of prospective execution paths. This section is dedicated successfully navigating AND vs. OR decisions.

Is it really multi-wave?

When confronting a potential multi-wave market, we should first make sure there will actually be multiple waves of disruption in short succession. While there might be a first wave of disruption today, if future advancement along the path to super-intelligence can be cleanly integrated, it’s only a single-wave market. Similarly, if there’s a very long time between waves, it’s also a single-wave market. In both cases, traditional startup wisdom applies.

Markets with fixed customer requirements are perhaps the only time when it’s safe to assume only a single wave of disruption. With no future disruption expected, it’s safe to invest in structural power. Robotics for manufacturing, as we’ve discussed, will likely present only a single wave of disruption, with future advancement being evolutionary and on a path to sufficiency. 

Multi-wave markets require at least two waves of disruption between now and super-intelligence. ChatGPT was disruptive to search because interactive dialogue necessitates a very different interface. As models become more advanced, it’s likely that dialogue-oriented interfaces will in turn be disrupted by other interaction modes. A multi-wave market. Co-pilots disrupt manual workflows. Agents will disrupt co-pilots and super-intelligence will disrupt agents. A multi-wave market.

Because it is so risky, companies should AND very carefully.

Reimagining a product to take advantage of a new stack is possible, though risky, because a company has to question all aspects of their business—design, technical architecture, pricing/packaging, messaging, and go-to-market… to name a few. Netflix famously was so paranoid about the transition from DVD-by-mail to streaming that they maintained separate executive teams for each of the two business lines and nearly completely separated the brands (the DVD business was to become “Qwikster”) before customers revolted.

OR should be the default for most companies because it’s so risky to willingly attempt to survive multiple waves of disruptive technology change in rapid succession. However, if a particular set of requirements are met, it’s possible for AND to be a coherent strategy:

• Viable first wave: When pursuing an AND strategy, a startup’s first step must succeed or they won’t survive to face the future. If a tech stack cannot viably solve customer problems or customers aren’t ready to adopt the new stack, then “being too early is indistinguishable from being wrong”. Zeppelin never became fully viable for mass adoption by the time fixed-wing aircraft were. Don’t build Zeppelin.

• Enough time on wave: With the right amount of time to execute, startups might be able to build transferrable advantages; however, if subsequent waves come too quickly, investment in prior waves is a handicap vs. unencumbered competitors. Can we scale our co-pilot before agents are viable? Can we scale our agent before super-intelligence makes agents cumbersome?

• Transferrable advantage: There must be some benefit for taking the risk of building a product that will soon be disrupted. Otherwise, why not just wait to build the right product? 

Transferrable advantages in multi-wave markets

Power in one tech stack does not always confer an advantage in future tech stacks. If it did, we wouldn’t call it disruptive technology because incumbents would continue winning forever unless they truly screwed up.

Execution power transfers perfectly into new waves. If we’re fast, we can take advantage of the new tech stack before others. If we have taste, we’re more likely to build an exceptionally designed/architected new 10x product. If we can rain-make, we can get a lot of customers and quickly dominate the new market. Execution power (speed, taste, and/or rainmaking) is universal. If a company has this rare talent, they can focus on great products for their customers. Others will go slower, build worse products, and/or be unable to command marketshare. The trick here is knowing whether a company really has execution power, or they’re just arrogant and delusional.

Some structural powers are transferrable if managed strategically. An exceptionally strong engineering team that builds a market-leading co-pilot can probably also build the best agent. Network effects might apply in similar ways across stacks. A great brand adapts to changing technology. Economies of scale and process power persist if some components are re-used across subsequent stacks. 

Netflix leveraged its distribution and economics of scale in content licensing to win in the transition from DVDs to streaming. Uber is the dominant ground transportation brand while also having process power in managing vehicle liquidity: both relevant to autonomous vehicles. 

These examples, however, are in single-wave markets because disruption was widely spaced and it’s unclear if transferrable structural power matters in multi-wave markets because structural power takes time to build. In a single-wave market where companies have a long time to build power before being disrupted, transferrable structural power is meaningful.

Many structural powers do not transfer. Companies might achieve power in a given tech stack by optimizing cost structures, filing patents, hiring the best talent, or increasing switching costs by becoming intertwined with customer workflows. In a new stack, the cost structure might be different, the patent portfolio might be irrelevant, the talent might be mismatched, and customers’ workflows might need to be re-engineered. These sorts of power don’t confer an advantage in future tech stacks.

On the other hand, distribution scale in a prior stack speeds up adoption in new stacks. Companies with scale can cross-sell new products into their existing customer base and grow faster than those starting from scratch. However, if companies are reluctant to cannibalize their existing revenue in the face of a future wave of disruption, then scale can actually hinder success. 

When starting from scratch in multi-wave markets, only startups with exceptionally fast growth can build enough distribution to confer an advantage in future stacks.

The fog of war is thick in fast moving markets—assume uncertainty.

In the case of Zeppelin and fixed-wing aircraft disrupting ocean liners, both stacks were visible: a careful observer could understand customer preferences, components in the stack, and figures of merit needed to make each stack viable. As a prospective founder or investor choosing whether to AND or OR, there weren’t many surprises. OR was the correct strategy: airplanes won and Zeppelin were relegated to the annals of history.

If we can see multiple disruptive stacks on the horizon, it’s best to OR—we won’t have enough time on wave to build distribution or structural power betting on both. Co-pilots, for example, are likely to be obviated by agents within two years in many use cases. Companies pursuing an AND strategy won’t have much time to get distribution scale and it’s unclear whether they can build significant structural power (e.g. switching costs, data network effects) before the agent stack becomes viable. Let’s pick the best wave and make sure we win.

In all other circumstances, we should expect future waves of disruption, even if not yet visible. AI is moving so fast that we can’t see more than a couple of years out; we don’t have the luxury of a long-distance view into the adjacent possible. We might start with no new stacks on the horizon. However, by the time we build our 10x product and start acquiring customers, an even better tech stack will be on the verge of viability. Uh oh.

If a first disruptive stack becomes viable, but we can’t yet see the next one, it’s reasonable to AND. In markets where new disruptive products are likely or there’s induced demand, then it’s difficult to justify OR when we have no idea what we might be waiting for, even if we’re reasonably sure another disruptive wave will come. If the first wave is viable, then a fast-moving startup might have enough time to build distribution and structural power that transfers into the next disruptive wave.

Because it transfers cleanly, execution power is especially valuable given the uncertainty of multi-wave markets. If we can’t predict which structural powers will transfer into the next disruptive stack, how do founders and investors know what to strategically bet on? Execution power is the most reliable form of defensibility—practically speaking, this means building a team with speed, taste, or rainmaking ability. However, we need to make sure we have more execution power than anyone else. If everyone has it, it’s not power.

Next Post

Previous Post

Defensibility
A series of posts dedicated to answer the question: Where will value accrue in AI?

1. Introduction: Why this time is different.

2. Technology markets are organized around “stacks”.

3. "Execution power” is becoming more important than classical “structural power”.

4. Market revolutions occur when “critical" technology makes a new stack “viable”.

5. When multiple stacks become viable in rapid succession, companies must “AND” or “OR”.

6. Power within the AI stack—hardware, hosting, models, and infrastructure.

7. Power in AI applications—big tech, switching costs, network effects, and the $100 trillion of global GDP up for grabs.

If you’re new, start with the first post and subscribe to be notified as new posts are published.

Starting in the late 1920s, Zeppelin carried passengers on intercontinental voyages between Europe, North America, and South America. A journey from Europe to North America that once took 5-7 days on an ocean liner was merely 2-3 days via airship. 

Zeppelin of the time comprised a large rigid airframe filled with a lifting gas (hydrogen1). Because they could lift so much weight, these hulking airships were able to carry the powerful, but heavy, internal combustion engines that allowed them to cross the oceans reliably. Many imagined a future where Zeppelin would dominate global passenger travel for some time to come.

In the late 1920s, it was easy to dismiss fixed-wing aircraft given their small and cramped cabins compared to the luxury of the Zeppelin. Why were fixed-wing aircraft cabins so small? The “heavier-than-air” fixed-wing aircraft tech stack at the time comprised a fabric and steel airframe and was restricted to smaller aircraft because engines with sufficient power-to-weight ratios could not be made and reliable or efficient enough for long journeys.

“Critical” technologies are the key drivers of customer satisfaction in a tech stack.

The airship’s reign over the skies was short-lived. By the late 1930s, the fixed-wing tech stack became viable for safe overseas crossings. Engines providing sufficient power-to-weight ratios for larger aircraft became both reliable and efficient enough for overseas flights. In 1938, the piston-powered Boeing 314 could make the transatlantic journey in merely 18-20 hours while transporting its passengers in luxury comparable to airships. Engines were the critical technology for long distance fixed-wing flight—they were the limiting factor for the long-distance airplane transportation market.

Once the fixed-wing distance barrier was overcome, the market shifted rapidly. Even if airships had successfully transitioned to helium or another non-combustible lifting gas after the Hindenburg disaster, few would opt for a multi-day journey on an airship given a much faster alternative. When it comes to travel, faster is better. There’s no speed that’s “good enough”. All of the effort spent making airships safer, more luxurious, and faster just didn’t matter once combustion engines became sufficient for long-distance fixed-wing flight.

Consider another example: Uber and Lyft’s ascension. Uber started out as a simple SMS-based dispatch service for black cars—“please pick me up at 539 Bryant Street”—and was far better than traditional taxis. The typical taxi experience was awful: waiting outside in the rain, hoping an empty cab would pass, or calling a dispatch center only to be told, “if a taxi doesn’t arrive in 30 minutes, call back”. Similar to early aircraft, it was easy to dismiss Uber as “limos for rich people” and not yet viable to replace taxis at scale.

However, the landscape changed rapidly with the advent of smartphones equipped with GPS and high-speed data connectivity. This technology allowed Lyft to build a smartphone-based dispatch product that enabled anyone—not just professional drivers—to pick up riders at much lower prices than taxis or black cars. Uber quickly launched a competitive service, UberX, sparking a revolution in the ground transportation industry. Taxi medallion values crashed and semi-corrupt local regulators were apoplectic. The market transformed overnight as soon as the ride sharing tech stack became viable to better meet customer needs. In this case, GPS and high-speed internet on smartphones were the critical technologies for ride-sharing to thrive.

Improving critical technologies within a tech stack results in evolution, not revolution.

Driven by wartime improvements, airplane engines kept getting better. By 1943, the Lockheed Constellation could complete a transatlantic trip in only 14-15 hours. By the 1950s, the jet-powered Boeing 707 further reduced this time to merely 7-9 hours. 

Since then, fixed-wing airplanes have remained the primary method of moving around the planet despite tremendous advancements in virtually every area of technology. Why? No new long-distance travel tech stack has become viable. We’re not (yet) traveling on ballistic rocket flights from San Francisco to Tokyo and there are no under-ocean hyperloops from New York to London.

When no new tech stack becomes viable, but critical technology continues to improve, customer value increases while markets undergo evolutionary rather than revolutionary change. The aircraft industry exemplifies this phenomenon. Manufacturers were able to upgrade engines—transitioning from piston engines to turbojets and eventually turbofans—without fundamentally changing their product lines or business models. Case in point: Boeing’s 7x7 aircraft family remains a dominant player nearly 70 years later. In this evolutionary market, airlines and aircraft manufacturers succeeded or failed based on their ability to provide what people cared about: lower prices (achieved through more efficient engines and larger aircraft with greater passenger capacity) and expanded global coverage.

Looking ahead, the fixed-wing tech stack is likely to see only further evolutionary changes, barring the emergence of a viable new tech stack. Perhaps the electric vertical-take-off-and-landing (EVTOL) tech stack will eventually become viable and disrupt the regional air travel market while ballistic flight disrupts long-distance air travel. However, until these or other new tech stacks prove viable, the aviation market will continue its pattern of gradual, evolutionary progress.

Once a critical technology ceases to impact customer value, it becomes “sufficient”.

Smartphone technology, like aircraft propulsion, advanced tremendously in the first two decades of the 21st century: better screens, faster networks, and desktop-class processors. Uber remained successful despite these advances. Why? Similar to the fixed-wing aircraft industry, no new method of ground transportation reached viability despite rapid advancements in underlying technology. While improvements in smartphones made Uber’s app better looking, snappier and accessible to a global audience, they didn’t meaningfully change how ground transportation services were delivered.

This contrasts with airlines, where engine efficiency improvements continue to significantly impact ticket prices (fuel constitutes a large portion of an airline's variable costs) and overall profitability. As a result, engines have remained a critical component in aviation.

In ride-sharing, however, smartphone technology has mostly stopped mattering. Uber focuses on what customers care about—faster pickups, multiple pricing tiers, scheduled rides, and regulatory liberalization. Faster smartphone processors don’t make a car come faster. Bigger screens don’t make a ride cheaper. Smartphones became “sufficient”—where further advancements no longer improve the product—even though they were once the critical technology for the market.

Disruptive market revolutions occur when critical technology makes a new stack viable.

The concept of the “adjacent possible” suggests that humans can reliably predict the future only one conceptual step at a time. Looking more than one step forward—a future that is no longer adjacent—introduces so many possibilities that uncertainty dominates, making predictions unreliable. Someone might make a lucky guess, but it’s just that: a guess.

When a new technology stack enters the adjacent possible, we can confidently envision how the pieces fit together, even if that future isn’t yet technically achievable. In the case of the Zeppelin, while suitable engines for reliable long-distance flight didn’t yet exist, the fixed-wing tech stack was conceptually understood. A keen observer, noting steady progress in internal combustion engine reliability, efficiency, and power-to-weight ratios, could foresee the Zeppelin’s eventual obsolescence. Early prototypes like the Dornier Do X flying boat with its first flight 1929 hinted at this future even if it wasn’t yet commercially viable.

Today, autonomous vehicles represent the adjacent possible tech stack poised to revolutionize Uber's market. In this case, AI—not the smartphone—is the critical technology driving the autonomous vehicle stack. 

While some companies survive such revolutions—Netflix transitioning from DVD-by-mail to streaming is a rare example—most struggle. We can bet that Uber is working hard to retain its marketplace dominance in a post-autonomy world, even though autonomous driving technologies aren’t fully viable outside a few test markets. 

Critics of current autonomous driving technology point to expensive lidar units, limited test environments, and overly cautious driving behaviors. That will change. Tesla’s camera-based stack can reduce the cost of vehicles while advancements in AI will enable driving in all conditions (level 5 autonomy) with sufficient assertiveness to navigate complex traffic scenarios like unprotected left turns during rush hour. 

Absent regulatory intervention, expect an autonomous driving tech stack revolution soon.

A generalized algorithm for critical and sufficient technologies

To summarize, here are the four ways technology advancement impact markets:

• When new stacks delivering superior value enter the “adjacent possible”, revolution looms.

• When “critical” technologies make a new stack viable, a disruptive revolution occurs.

• Improving critical technologies within an existing tech stack results in evolution, not revolution.

• Once technology reaches “sufficiency”, its advancement no longer significantly impacts the market.

When a particular technology is advancing, consider its impact on both current stacks and its potential to make new stacks viable.

For a stack to be viable, it must not only be novel but also deliver superior customer value. To understand if that’s the case, start with psychology. What do customers value? Speed, quality, cost, convenience, adaptability, design, aesthetics, or other factors? Novelty alone doesn’t matter.

If no new stack becomes viable, successful incumbents will incorporate the new technology and continue “surfing the wave.” This explains why it’s challenging to compete with Boeing and Airbus in the aircraft market.

If a new stack becomes viable due to a critical technology, both startups and incumbents must be prepared to “catch a new wave”. Missing the wave is fatal—consider how the taxi industry was blindsided by Uber and Lyft. Incumbents who are aware of the “innovator’s dilemma” may successfully catch a new wave and leverage their existing distribution to beat startups even in the face of a disruptive market revolution. However, much of of the time, incumbents can’t get out of their own way and a new leader emerges.

In the case where there’s only one new tech stack on the horizon with the promise of delivering superior customer value versus an incumbent stack, go! In this “single-wave” scenario, invest in building a great product and then establishing power so we can become the dominant player.

Decisions are harder when multiple alternative stacks are likely to become viable given advancement in their critical technologies. Just as surfers eyeing a set of large waves must decide which to catch, companies must decide which of the soon-to-be-viable tech stacks to build on. Airships, while superior to ocean liners in many ways, ultimately proved to be the wrong wave to catch. We’ll explore how we might face such “multi-wave” scenarios in detail below. 

Revolution and evolution in AI

Where AI integrates cleanly into existing products, there will only be evolutionary change to today’s tech stacks. When users seamlessly transition from non-AI to AI-powered workflows, strong incumbents are likely to remain winners by relying on greater distribution and accumulated structural power. Photoshop added AI painting features without disrupting their interface. Gmail cleanly integrated AI email drafting features. Startups in these markets are likely only to succeed when incumbents execute really poorly (which does happen frequently, e.g. WhatsApp vs. SMS, Dropbox vs. prior file sharing vendors, GitHub vs. other version control systems, Zoom vs. Webex).

When AI-powered interactions necessitate totally new user interfaces, expect incumbents to proceed slowly. Longtime customers typically resist major user experience changes. Professionals have spent decades turning Bloomberg, Photoshop, and Microsoft Excel into extensions of their own bodies. They’ll be understandably upset if that investment were to become irrelevant. Predictably, incumbents are often unwilling to infuriate their customer base (and throw away their power of lock in) by releasing a totally new UX. Similarly, “systems of record” for Sales, Marketing, Finance, HR, Legal, Supply chain, and other departments were built atop manual workflows as an upgrade to offline paper processes. Enterprises made huge investments building these workflows, so incumbents are unlikely to want to rock the boat.

When incumbents move slowly, revolutions are likely. When AI enables new disruptive stacks (e.g. ChatGPT, AlphaFold, humanoid robots, enterprise agents), new products can provide far more customer value than legacy products. While hypothetically startups and incumbents could both build these disruptive new products, incumbents are often so afraid of cannibalizing their existing business that they avoid moving until it’s too late. Startups are unencumbered by the past and can move at warp speed.

Regardless of whether AI is revolutionary or evolutionary today, companies large and small must watch out for additional waves of disruption in the future. Will an incumbent be able to maintain their leadership amidst continued AI progress? Will a startup that wins in a first wave of disruption be disrupted by a future wave? To answer these questions, we can rephrase them as: “If we had true superhuman AGI, would we still deliver customer value in the same way?” In other words, what might a “GPT-8” or “Llama 7” world look like?

In markets with only one wave to catch, we expect a first wave of disruption where future AI progress causes only evolutionary change. Winners in the current generation of AI disruption will be able to surf for quite some time and we’re pretty sure that even superhuman AGI won’t change how they deliver value.

These “single-wave” markets are likely when customer requirements remain fixed. When AI directly impacts customer satisfaction—it is critical technology—but future advancement integrates cleanly into products, it won’t disrupt the status quo. Future products might become faster and cheaper, but the stack otherwise stays the same. 

With fixed requirements, AI will eventually become sufficient for the use case, with its advancement no longer impacting customer satisfaction. For example, in manufacturing automation‚ the core job to be done is a set of manufacturing steps. While superhuman AGI might execute these steps faster, cheaper, using universal hardware, or with less programming, the fundamental goal remains the same. The first wave of AI manufacturing will be disruptive, but subsequent progress will be evolutionary, eventually reaching sufficiency.

With only one wave to catch, our time-on-wave is long and traditional startup wisdom applies: we can start by relentlessly delivering value to customers. A “10x” product is a viable strategy so long as a startup can maintain its lead. Startups with true execution power can win, particularly if they stay outside of big tech’s blast radius. If a company has execution power and they successfully surf a wave with no new tech stacks on the horizon, they can afford to start building structural power. Uber had a very long time between smartphone-led disruption and the current wave of disruption from autonomy. They were able to get to massive scale before having to worry about subsequent waves of disruption.

“Multi-wave” markets occur when more than one disruptive technology revolution occurs in short succession. Companies in multi-wave markets will still be grappling with a first wave of disruption when another rudely hits them. Zeppelin were still in their infancy attempting to disrupt sea travel when fixed-wing aircraft overtook them. It’s easy to imagine an alternative history where Zeppelin dominated air travel for decades—well-known airship brands with terminals in every city. If long-distance fixed-wing aircraft arrived after the airship industry matured, incumbents may have been able to use their capital, terminal infrastructure, and brand recognition to win in the fixed-wing aircraft market. But that didn’t happen. Fixed-wing aircraft became viable long before Zeppelin got to scale—death by a multi-wave market.

Whenever progress in AI repeatedly changes how humans interact with software, multi-wave markets are likely. Knowledge work, for instance, has an almost infinite demand for expanding scope and capability. Today’s narrow, task-specific AI co-pilots help humans run existing workflows. Co-pilots are likely to be wholly replaced when human-equivalent AI agents become possible to fully automate these workflows. Agents, in turn, may be superseded when superhuman AGI manages entire departments. Each improvement in AI changes customer requirements and expectations, opening the door for new revolutionary tech stacks.

Multi-wave markets are also likely when the availability of intelligence increases the demand for intelligence: “induced demand”. Each time companies dream up new AI-enabled products, the adjacent possible expands. Customers want to solve problems they didn’t think could be solved before. By the time we get to superhuman AGI, markets might exist for products that are unfathomable today.

When AI-enabled law firms make contract law more efficient, we’ll probably have more and longer contracts with new ways of utilizing legal services. Once we have a humanoid robot cleaning the house, we’ll want a robotic butler at our beck and call. When most of our marketing collateral is AI-generated, we’ll want campaigns entirely personalized for each prospective customer while remaining consistent with our brand. When models can predict protein structure, biologists will want to predict protein-protein interaction, then model cells as a whole with an eye towards designing biological machines the way we currently design mechanical machines. Each major increase in demand for intelligence opens the door for new tech stacks to satisfy that demand.

Multi-wave markets are riskier than single-wave markets because long term success requires successful navigation of multiple disruptive technology shifts. Unlike the past, where mature companies had years to navigate a given wave of disruption, AI requires even immature companies to grapple with multiple cascading waves of disruption coming at them at once.

How should founders and investors deal with multi-wave markets? This is the most important strategic question for many AI companies today that we’ll address in the next post.

Next Post

Previous Post

Defensibility
A series of posts dedicated to answer the question: Where will value accrue in AI?

1. Introduction: Why this time is different.

2. Technology markets are organized around “stacks”.

3. "Execution power” is becoming more important than classical “structural power”.

4. Market revolutions occur when “critical" technology makes a new stack “viable”.

5. When multiple stacks become viable in rapid succession, companies must “AND” or “OR”.

6. Power within the AI stack—hardware, hosting, models, and infrastructure.

7. Power in AI applications—big tech, switching costs, network effects, and the $100 trillion of global GDP up for grabs

If you’re new, start with the first post and subscribe to be notified as new posts are published.

In the first stage of a new market, companies must build something people want: a product that doesn’t suck. Without that, customers will be scarce and that’s certain death.

After achieving product-market fit, companies will attempt to grow as quickly as possible. On this path, others will compete as copycats or parallel innovators. Either way, customers will have multiple alternatives to choose from.

Zooming out to how cash flows through the stack as a whole, every company aims to keep as much value as possible for themselves. A physical part buyer will evaluate multiple manufacturers, a manufacturer will look at multiple robot vendors, a robot vendor will look at multiple model vendors, and so forth. 

This is capitalism; competitive pressures force all players to become better and more efficient. This is good. As is sometimes said: the cure for high prices is high prices. Fat margins sing a siren-song for new entrants unless something stops them.

In this darwinian struggle, who wins? Value accrues to those with power: the ability to tilt build vs. buy decisions in their favor. In order to have durable cash flows over time—the definition of a great business—a business needs power.

Power must provide the holder a tangible benefit (e.g. lower cost, ability to charge higher prices) while preventing competitors from replicating the advantage.

Product-market fit is not power. If multiple companies delight customers via interchangeable products with similar production costs, then profits will erode.

If an advantage is universally available, there is no resultant power. Viewed through the lens of the stack, power only accrues to better products if:

• the advantage stems from something we control and

• we can somehow exclude others from accessing the same advantage. 

Misunderstanding power in the stack is perilous: IBM, when sourcing an operating system for their PC, may have seen Microsoft’s DOS as merely one component of many. It was not. Microsoft gained power because application developers customized their capital investments (software) for the OS, not the hardware. The growth of IBM-compatible PCs commoditized the hardware layer while the real money was made at the OS layer. Microsoft had power.

Structural power is the classic defensibility playbook.

Hamilton Helmer’s 7 Powers is perhaps the most well understood framework on defensibility. It’s a worthwhile read on the major categories of structural power that I’ll summarize here:

• Branding—Disney, Nike, IBM, Apple: premium pricing through perceived superior quality.

• Process Power—Apple, NVIDIA, TSMC: some sort of alchemy in a company’s operations that yield lower costs or better products that sell at higher prices.

• Economies of Scale—AWS, Walmart, TSMC, Uber: size-driven advantages in costs or better products.

• Network Effects—Facebook, Instagram, X, Windows, ChatGPT: value grows with user base, often due to more people to communicate with or datasets that make superior products.

• Switching Costs—Salesforce, NetSuite, Airline loyalty programs: high barriers to changing providers. In organizations, lock-in is often due to data, integration, or process inertia.

• Counter Positioning—Apple (vs. Facebook) in Privacy, OpenAI (vs. Google) with no ads in ChatGPT: business models that cannibalize competitors.

• Cornered Resource—Qualcomm patent portfolio, DeBeers mines: exclusive assets that increase value or reduce costs.

These structural powers are clearly observable company attributes, making them “intellectually satisfying”. They make excellent case studies because they’re so obvious when present. 

With startups, there are several things to remember when pursuing structural power:

• Structural power is predictable based on the market. Launching a social network? Network effect. Building an enterprise workflow tool? Switching costs. Consumer packaged goods? Brand. Real-world logistics? Economies of Scale …and so on.

• Power must actually influence customer decisions, or it’s a “power mirage”. If customers can ignore power with impunity, then it’s not real power: it’s a misleading slide on a fundraising deck. Consider an example from a prior letter:

"…while the concept of a “network effect” is intellectually satisfying, we can restate the hypothesis as: “Customers will be happier with our product if we had more customers and less happy if we had fewer customers.” In some cases it might be true (e.g. communication tools or proprietary AI models that depend on lots of training data), but in many cases a product’s value is more or less the same regardless of the number of customers. In this latter case, there’s clearly no network effect”

• Structural power takes time to build—it’s a “step 2” after initial success once a company has the ability focus on something other than survival. Investors often fall into the trap of over-emphasizing structural power when evaluating pre-seed and seed companies because it’s so satisfying when the question of defensibility comes up at a partner meeting.

If this is a complete description of defensibility, why do companies with no obvious structural power sometimes win decisively? Why did Stripe beat other payment processors? Why does SpaceX continually beat everyone? At the early stage and in rapidly evolving markets, a different type of power exerts itself.

Execution power is the subtle secret behind many successful companies.

While structural power gets the most mindshare, execution power is the true cause of many successful startups. Founders and experienced investors have seen execution power firsthand, yet many investors overlook companies that don’t (yet) have some form of intellectually satisfying structural power.

Structural power is readily observable from a snapshot while execution power can only be seen in motion. Execution power is dynamic power—the ability to move with the world—not something that can be easily evaluated at any point in time, making for ambiguous case studies.

I’ve observed three types of execution power create durable advantages: Speed, Taste, and Rainmaking.

Speed is the ability to ship products and incorporate feedback faster than others. It gives startups an edge over slow-moving incumbents. Execution speed is especially important in dynamic—and therefore immature—markets. If the battleground changes frequently, companies that can adapt the fastest tend to win. In AI, where advancements occur every day, we’re going to get smoked if we can’t incorporate them immediately.

“One of the biggest advantages that start ups have is execution speed and you have to have this relentless operating rhythm.”
- Sam Altman

Taste is the ability to make correct decisions amidst uncertainty. It manifests in exceptional product design (e.g., Stripe’s payment APIs, Airbnb’s early investment in great photography, the iPhone), technical decisions (e.g., SpaceX’s bet on reusability, Tesla’s focus on factory automation, OpenAI’s investment in infrastructure that enabled their research team to move faster than others), as well as picking talent (e.g., the "PayPal mafia” and the many other subsequent startup mafias).

Rainmaking is the ability to conjure revenue out of thin air. Some people are just very good at creating buzz (e.g., Marc Benioff’s “No Software” protest outside a Siebel Systems conference) or getting high-profile organizations to buy (e.g., Palantir’s early customers were some of the most demanding on the planet). Marketing is a “wholesale” (one-to-many) approach to getting customers, so skill there scales quite well. Sales is a “retail” (one-to-one) approach, so it tends to work best for big deals.

How does execution power scale?

Execution power is clearly valuable for startups, but what happens as they grow? Most organizations calcify over time, becoming less effective as they scale. This phenomenon serves as “dose-response” evidence for execution power: potent when present and an inactive ingredient when diluted.

Some companies that avoid “execution power dilution” are able to maintain defensibility at massive scale, creating the most effective technology organizations on the planet. SpaceX, Tesla, and OpenAI are great examples of what happens when companies retain speed at scale. 

The recent debate over “founder mode”—where leaders eschew traditional delegation structures and become intimately involved in the details of individual contributor work—is a symptom of the struggle to preserve execution power at scale.

Here’s a typical lifecycle of execution power dilution: at a startup, everyone is an individual contributor (IC); they are directly responsible for getting things done. Execution power present in even a single person can make the team as a whole exceptionally effective. As a result, the startup becomes successful.

Scaling using traditional organizational design, founders hire multiple levels of management to wrangle increasingly large numbers of ICs. The organization that used to be optimized around IC effectiveness starts caring more about predictability, praying at the alter of “on time, on budget” (variability is the enemy of manager promotions). Exceptional work is hard to predict so it stops being expected. Calcification achieved.

Organizations can avoid execution power dilution by optimizing IC effectiveness: helping those directly responsible for output be as productive as possible. In these organizations, successful managers do not seek to minimize variability, but instead embrace it in order to achieve unpredictably excellent results by amplifying execution power in their teams. Stay up all night. Move unreasonably quickly. Agonize over design so you can be proud of the end result. Close deals that other’s couldn’t imagine possible. 

Rarely can one create an IC-optimized organization going the traditional delegation route because there is just so much business culture that aims to eliminate variability. It takes a founder (or other similarly empowered individual) to continually blow up the forces that constrain great IC work and the type of high-variability management that enables it. Ever wonder why many leaders who run organizations with execution power are criticized for their extreme levels of micro-management at both IC and managerial levels. It’s not a coincidence, it’s causality.

If done right, “founder mode” leads to trillion dollar outcomes. Some might attribute Apple’s success to structural powers: brand, economies of scale, and iMessage network effects. I believe Apple’s true power lies in the execution power of taste: their ability to ship world-class products.

“If you’re an Apple user and somebody offers you $10,000, but the only proviso is they’ll take away your iPhone and you’ll never be able to buy another, you’re not going to take it. If they tell you if you buy another Ford car, they’ll give you $10,000 not to do that, you’ll take the $10,000 and you’ll buy a Chevy instead.”
- Warren Buffett

That’s power.

AI’s age of infinite leverage amplifies execution power.

Execution power thrives on the intensity of a single founder or a small, dedicated team. As technology advances, it amplifies execution power by granting increasing amounts of leverage. 

“Expressiveness” is the unifying concept behind technology leverage: the ratio of output complexity to input specificity. A pre-industrial weaver’s work for few seconds might express a single stitch. Quite low. A software engineer’s work the same amount of time might express how an icon appears on billions of screens. Pretty high.

Software, while powerful, still has limited expressiveness because a program’s behavior must be explicitly coded. AI is on a whole different level: it is extremely expressive because models can make a massive number of conclusions from just a small input prompt in order to express a complex output. In the near future, an individual’s simple prompt might generate millions of lines of production code. AI has the potential to create “infinite leverage.”

AI does not automatically grant execution power to all. Companies will become more efficient, but in order to have execution power a company must be exceptional—better than everyone else. Let’s remember: power only exists when we have something others don’t.

With infinite leverage, a small team with execution power will be able to accomplish great feats without the risk of power dilution that comes with too large of an organization. We may soon witness multi-billion dollar companies run by only a handful of employees.

Perhaps even more exciting, what will a large team with execution power be able to achieve if they can successfully empower their ICs to operate with infinite leverage at massive scale?

As AI continues to amplify execution power, it will become the most important strategy for defensibility in the future.

Next Post

Previous Post

Defensibility
A series of posts dedicated to answer the question: Where will value accrue in AI?

1. Introduction: Why this time is different.

2. Technology markets are organized around “stacks”.

3. "Execution power” is becoming more important than classical “structural power”.

4. Market revolutions occur when “critical" technology makes a new stack “viable”.

5. When multiple stacks become viable in rapid succession, companies must “AND” or “OR”.

6. Power within the AI stack—hardware, hosting, models, and infrastructure.

7. Power in AI applications—big tech, switching costs, network effects, and the $100 trillion of global GDP up for grabs

If you’re new, start with the first post and subscribe to be notified as new posts are published.

To identify business opportunities in technology-driven markets, we must predict how players interact both technically and financially. Enter “the stack”, a visualization showing how individual components relate to each other. Typically:

• User interfaces and external APIs are on the “top of the stack” 

• Physical machines and cloud infrastructure providers are on the “bottom of the stack”. 

• In the middle are components that enable applications to function (e.g., internal “middleware” services, databases, file storage, deployment tools, development tools, and security tools).

When building a product, vendors assemble a unique stack by combining components built in-house or sourced externally. In each case they make “build vs. buy decisions”. Sometimes end-customers add a few things to a vendor’s stack themselves, such as when enterprises implement security and compliance tools or deploy software in their own environments.

To analyze complex technology ecosystems using stacks, we’ll cover several topics:

• Stacks exist because people make “build vs. buy” decisions.

• While each is unique, stacks tend to cluster based on use case and customer type.

• Subtle requirements create “leaky abstractions” in stacks.

These principles, while universal, apply in the case of the current AI market.

AI Stacks

In the illustration above, we can see an example cloud-based AI stack. Application developers building products such as AI travel agents, AI project management tools, or AI customer service platforms might use such a stack. In this case:

1. The application developer buys infrastructure software to become secure and resilient.

2. The application developer uses a foundation model vendor (e.g. OpenAI) via API.

3. The model vendor subcontracts execution to a cloud provider (e.g. Microsoft Azure).

4. The cloud provider purchases hardware (e.g. NVIDIA GPUs) and energy.

This scenario has a lot more “buy” than “build”. 

Sometimes, companies opt for more “build” than “buy” by vertically integrating across multiple layers of the stack. For example:

• OpenAI builds end-user applications, infrastructure and foundation models. 

• Google goes further, handling all of the above while also building hardware (TPUs).

Outside of large companies with unlimited budgets, it’s suicidal for application developers to build components internally when they’re available commercially given the cost of doing so. Acute “Not Invented Here” syndrome is a common affliction amongst strong engineering teams; many have unfortunately built their own databases, filesystems, HTTP servers, and project management tools. These companies accrue huge R&D expenses with little differentiated value if similar components are available off-the-shelf.

Given the widespread investment at every layer of the AI stack, application developers will be able to choose off-the shelf components for the majority of their stack. For example, even if a company runs models locally (e.g. on smartphones or robots) and doesn’t use the common cloud-based stack above, they’ll be part of a different ecosystem with similar relationships between applications, models, and hardware.

Follow the buck to predict where build vs. buy decisions occur

Build vs. buy decisions are the decision points that direct massive cash flows when aggregated across an entire market. Vendors that can make buy decisions go their way can achieve billions (or trillions) of dollars in enterprise value. Those that can’t, die.

If trying to build a company in a new technology market like AI, we need to identify where in the stack customers might buy something instead of building it themselves. We can play the game of “follow the buck” to trace cash flow through the value chain.

Here’s how it works:

1. Start at top of the stack where a customer pays for an application.

2. Ask: “What problems must the vendor solve to provide that value?”

3. For each problem, think about the vendor’s build vs. buy decision. Each decision point is a potential business opportunity.

4. If there’s a chance of a buy decision (exchanging money for value), goto 2 for each potential new vendor and repeat the process.

Playing this game in the real world often enough, we can make a few observations:

• Companies operating at the top of the stack (application developers and business customers) make build vs. buy decisions that massively impact lower layers. Pay close attention there.

• Each vendor assembles a unique “snowflake” stack for their use case.

• Individual1 end users typically use products as-is without customizing the stack.

With complex stacks, such as in AI, we find there are way too many components to trace. Such stacks exhibit fractal behavior: each component contains numerous smaller components, ad nauseam. Imagine a typical application server—a physical machine, an operating system, HTTP server software, various third party frameworks and libraries, and finally the application code itself—all supported by various security and infrastructure tools. Now multiply this hundreds of times for a large-scale product with many internal services. Then add in even more third-party services accessed via API. Yikes.

Including each of these details would render a stack precisely useless. We don’t need to know about every chip2 inside a server to understand why somebody might buy it. A good stack communicates just the important bits: clear component boundaries and predictable interactions. 

In systems design, encapsulating complex behavior in a single component is called “abstraction”.  When the behavior of the component is predictable from the outside, it’s a “clean abstraction”. Clean abstractions are easy to understand; however, as we’ll see below, sometimes subtle interactions are not predictable from the outside leading to “leaky abstractions”.

Component boundaries signal potential build vs. buy decision points. Any time we can encapsulate a set of functionality in a component, it’s likely to be something a vendor can provide, rather than something a company must build themselves.

Technology markets are snowballs where snowflake stacks converge

A single “buy” decision doesn’t make a market. If there’s only a single buyer for a component, vendors are unlikely to build it. Only with many buyers will a market materialize to fill the void.

How do we predict where markets might emerge? If we follow the buck through enough stacks, we’ll see some patterns. Markets occur when individual “snowflake” stacks align, creating a critical mass of demand for similar components. Companies that can dominate a component across many stacks have the chance to roll up a sizable snowball of a market.

In AI, there is no universal best stack—diverse application requirements necessitate varied technology choices—however, opportunities to create snowballs exist at each layer. Components present in many stacks tend to result in one of two types of markets:

• Universal components: If all users of a component have similar requirements, a unified market is likely. NVIDIA GPUs, capable of handling most AI use cases, show up in just about every stack today.

• Segmented components: Divergent requirements form multiple market clusters. The AI infrastructure layer has wide variety of customer needs so it’s unlikely that a single company will dominate.

Here are a couple examples of segmented use cases in AI:

Publicly-traded enterprises deploying AI co-pilots need to prevent employees from seeing confidential financial information.3 They might want a security component that identifies confidential financial information and blocks it from being accidentally disclosed. Similar security needs show up in other enterprise AI use cases, so cybersecurity vendors have the opportunity to build a large snowball here.

A privately-held manufacturer deploying AI-powered robots may not care if their employees see financial information; however, they might instead want to predict when robots need maintenance. Many types of manufacturers will deploy factory robots, so there will similarly be an opportunity for a large snowball.

As we can see, snowballs often cluster around similar use cases and customer types.

Subtle requirements create “leaky abstractions” in stacks

How can we tell whether a component will spawn a unified or a segmented market? Staring at a stack diagram alone won’t help. Technology is not a commodity purchased by the kilogram. 

Human psychology drives build vs. buy decisions; decision-makers weigh many factors to choose between purchasing, adopting open source components, or building in-house.

To correctly predict these decisions, we must think like technical decision-makers. It’s no coincidence that many founders and VCs in technical fields are engineers or could convincingly play one on TV.

In systems design, a “leaky abstraction” occurs when we need to peek inside a component to understand how it functions. In AI there will undoubtedly be leaky abstractions that cause build vs. buy decisions to deviate significantly from outside-in assumptions.

For example, consider a manufacturer automating low-volume part production. They need robots that can produce a variety of parts reliably, quickly, and most importantly, cheaply.

Assumption: With 900 current employees, they’ll buy ~300 humanoid robots if each robot working 24/7/365 can do the work of 3 humans working normal shifts.

Reality: The manufacturer instead opts for 100 traditional industrial robotic arms upgraded with AI-reprogrammability. Why?

• Historically low-volume part production wasn’t automatable, hence the human labor force.

• Given the size of parts, one large industrial arm produces the throughput of three humanoid robots, but at much lower cost. 

• The manufacturer is more comfortable buying hardware from a trusted local distributor of industrial robots vs. an unproven startup selling humanoid robots.

Leaky abstractions can make purchasing decisions go very differently than outsiders expect.

Now, consider an AI-enabled industrial robot arm vendor deciding whether to build or buy their AI foundation model. 

Assumption: They’ll choose a model based on a simple cost vs. performance metric amongst many open source or commercial options.

Reality: Existing models available to them aren’t designed for real-time systems and can’t reliably handle safety curtain violations (where a human enters the work area). They’ll need to build one themselves. Alternatively, if multiple industrial robotics companies face the same issue, a new market opportunity emerges for industrial robot-specific foundation models. 

Leaky abstractions can cascade down a stack, creating unexpected opportunities.

At the same time, be wary of blindly taking a contrarian stance. Outside-in assumptions often become consensus but sometimes the consensus is right. One consensus view that might be right: NVIDIA GPUs are really good and will dominate for a long time to come. For a contrarian view to be profitable, it must be both “non-consensus and right.” 

In AI especially, understanding decision-maker psychology can help identify those rare opportunities where non-consensus views are also correct. It’s in these unexpected behaviors—leaky abstractions in the AI stack—where founders and investors can uncover underserved problems and develop unique insights that allow them to get a head start on the market.

Next Post

Previous Post

Defensibility
A series of posts dedicated to answer the question: Where will value accrue in AI?

1. Introduction: Why this time is different.

2. Technology markets are organized around “stacks”.

3. "Execution power” is becoming more important than classical “structural power”.

4. Market revolutions occur when “critical" technology makes a new stack “viable”.

5. When multiple stacks become viable in rapid succession, companies must “AND” or “OR”.

6. Power within the AI stack—hardware, hosting, models, and infrastructure.

7. Power in AI applications—big tech, switching costs, network effects, and the $100 trillion of global GDP up for grabs

Scroll to the bottom to see subsequent posts in this series and subscribe to be notified as new posts are published.

New York, 1807—Step off the first viable paddle-wheel steamship and gasp in awe at humankind’s triumph. Harnessing steam, we’re a species no longer held prisoner by the mercurial winds.

In a city 70 years away from its first electric lights, envision trying to predict transportation in the 21st century. How could we possibly anticipate the impact of high-efficiency container ships and massive aircraft on the global economy? What might seem a fool’s errand is, I believe, tractable.

At that time, internal combustion engines, standardized shipping containers, heavier-than-air flight, and turbofans were still unknown; however, we could draw a box around a ship and ask: “What happens if moving things from point A to B becomes faster, cheaper, and more reliable?” 

With that framing and a bit of thought, we might conjure the world of international tourism and globalized supply chains.

The general algorithm for predicting the future of technology-driven industries is to know the “figures of merit” and chart their trajectory through time. 

With AI we ask: “What happens if intelligence is smarter, faster, cheaper, and more reliable?”  

The obvious near term answer is: “Magical new products will disrupt many industries.” True. However, AI is different from an ordinary disruptive technology revolution in two ways:

1. AI is obvious and available to all, not a proprietary or under-appreciated secret.

2. The pace of progress in AI is creating multiple technology revolutions in short succession; it’s not a single disruptive technology to be percolated throughout the global economy.

As a result of these two differences, startups must seriously consider competition—both from peers and future disruptors—much earlier in their lifecycle than the past. Let’s look at each:

Because AI is obvious and available to all: great products are necessary, but not sufficient.

In recent history, products that were “10x” better than the status quo became wildly successful. 

“Building something people want” became a startup mantra because few could build better products than what customers already had. Unsurprisingly, companies with exceptional products won big. Google, Amazon, Uber, Zoom, iPhone, and so on.

In AI, things are different. For those pattern-matching to the last 20 years, great AI product demos are now a trap. Because 10x products were historically indicative of a great team at the helm, we may be tempted to think that a revolutionary AI product also indicates greatness. 

Today, everyone can build something 10x better than the status quo simultaneously. Great AI products might indicate a great team, but given the wide availability of state-of-the-art models, even average teams can build seemingly-revolutionary AI products.

When everyone uses the same revolutionary technology, today’s groundbreaking product becomes tomorrow’s table-stakes feature. Blindly matching the 10x pattern leads us astray.

Unlike past technology revolutions when disruptive products blindsided incumbents, today’s incumbents are aggressively investing in AI. Startups competing on the virtue of AI features alone will get crushed; both can build 10x AI products, but incumbents have a lot more distribution.

When multiple technology revolutions occur in short succession, disruptive startups can themselves be disrupted before they get to scale.

Startups used to focus on product and distribution for years before worrying about defensibility. By betting on technology shifts—“why now?”—startups were pure beneficiaries of disruption.

Historically, industries had many years to digest a disruptive technology before a subsequent wave of disruption occurred. The revolutionary shift from mainframes to minicomputers, then minicomputers to PCs played out over decades. As did the transition from 14 inch to 8 inch disk drives, followed by the transition to 5.25 inch and eventually 3.5 inch drives.

In AI, even if a startup wins during a first wave of disruption, there’s no guarantee they will continue to win. As models become increasingly capable, new architectures will repeatedly disrupt prior generations. The shift from co-pilots to limited function agents and agents to superhuman AGI may take place over a few years. Each of these shifts is likely to make startups built on a prior architecture obsolete.

When multiple technology revolutions occur in short succession, startups must worry about disruption before they get to scale. In this case even immature companies founded last year must worry about disruption from startups in the next Y Combinator batch. Startups are now both beneficiaries and victims of disruption.

If startups 1). cannot solely rely on building 10x products to win and 2). must face multiple consecutive waves of disruption, we’re forced to confront the central question of this series: “What is defensible during a period of rapid technological change?” 

Centered upon today, we can rephrase that as “Where will value accrue in AI?”.

To build a durable business in AI, startups must have some form of unfair advantage today that transfers across any near term wave of disruption to come. They need some way of tilting the scales away from competitors and towards them despite substantial technological change.

Why do customers choose one vendor over another? Power. Without it, a company might win for a short period of time, but they won’t win forever.

What is power? It’s the ability to achieve some advantage (e.g. lower costs or ability to charge higher prices) in a way that competitors can’t replicate. Product-market fit is delighting customers, but power is beating competitors.

To illustrate, let’s imagine a bedtime story for my future daughter:

Daughter: Daddy, where does power come from?

[Jeni glances at Kevin, eyebrow raised, as if to say: What have you done to our poor child?]

Kevin: Well, imagine the world of apples… You love apples and your friend Bob sells them for $2 each.

Everyone’s happy. Bob’s business is so successful that he buys a plane and a boat. And you have apples. 

But then Alice shows up with a fancy automated apple harvester, selling the same apples for just $1. Being a smart cookie, you start buying Alice’s apples instead.

Bob’s upset! He can’t afford an automated apple harvester, and why would you pay an extra dollar per apple just so Bob can fly his plane to St Barth’s?

Bob doesn’t have power. Instead of buying expensive toys, he should have invested in his apple business. Not smart, Bob.

Now your friend Carol comes along. She’s a brilliant genetic engineer who worked really hard to create a super apple that tastes absolutely delicious. She buys an automated apple harvester as well.

She charges $1.50 for her apples—a little pricier that Alice’s, but way more delicious. You happily pay an extra 50 cents and so does everybody else.

Alice doesn’t have power. Anyone can buy an apple harvester. Not smart, Alice.

A few years later, Carol’s business is booming. One day, she talks to her friends from college, Dan and Eve. They’re also genetic engineers dreaming up new types of apples. 

At first, Carol was a little scared. She thinks, “What if their apples are better than mine?”

But Carol is a smart entrepreneur. She remembered that the trickiest part of her business wasn’t modifying the apple genome, but growing millions of genetically engineered apple trees cheaply.

Because her apples were so popular, Carol worked extremely hard for years automating planting, growing, orchard maintenance, harvesting, and delivery with AI. She called it her Apple Workflow System (AWS).

It started off simply with just a few robots she programmed herself, but AI just kept getting better and better. Each time a new AI model came out, Carol quickly incorporated it everywhere she could. 

Now Carol has thousands of robots tending her apple orchards with just a few people to supervise them. Her system is so good, nobody in the world can grow apples as cheaply.

And that’s when Carol had her eureka moment. Instead of trying to compete with her friends, why not sell them AWS? This way, her friends can focus on dreaming up new apple varieties without worrying about orchards and trucks.

Dan and Eve loved the idea! They got to focus on what they love—crafting apple genomes—without having to become orchard managers or produce logistics experts.

Fast forward a few years and it’s your dream come true as an apple-lover.

Dan and Eve have several wildly successful apple varieties and Carol is still hard at work making many improvements to the system that runs the orchards that produce them all.

Carol has power.

Notice how we avoid naming structural powers (e.g. “network effect”, “economies of scale”, “brand”)? Using big words views a market in “far mode”: great at justifying who is currently winning but poor at understanding the nuance of who might become a future winner.

Instead, I find it more helpful to think in “near mode” and use customer psychology thought experiments as ground truth. Imagining ourselves as a customer is the best way I’m aware of to avoid the irrational exuberance that accompanies many bubbles. 

Specifically, Carol’s ultimate advantage comes from relentlessly building a platform that Dan and Eve don’t really want to build, at a pace they couldn’t hope to match. Carol’s speed is a type of “execution power” that is the subtle secret behind many companies’ success. Because AI gives individuals “infinite leverage”, execution power is especially potent in the post-AI world.

To adequately cover power dynamics in technology markets, I’ll be publishing multiple posts in the coming weeks. We’ll start with principles and then address the $100 trillion question of who will accrue power in AI.

Similar to prior posts, I’ll attempt to be as simple as possible—but no simpler—and, for the sake of brevity, make definitive conclusions instead of hedging with probabilistic language. As always, there are likely to be smart sophisticated and successful (SSS) who disagree with either the principles or some of the conclusions. I hope this spawns some great discussions to come :)

Thanks for reading.

– Kevin Mahaffey

Next Post

Defensibility
A series of posts dedicated to answer the question: Where will value accrue in AI?

1. Introduction: Why this time is different.

2. Technology markets are organized around “stacks”.

3. "Execution power” is becoming more important than classical “structural power”.

4. Market revolutions occur when “critical" technology makes a new stack “viable”.

5. When multiple stacks become viable in rapid succession, companies must “AND” or “OR”.

6. Power within the AI stack—hardware, hosting, models, and infrastructure.

7. Power in AI applications—big tech, switching costs, network effects, and the $100 trillion of global GDP up for grabs

Series As, after their own sabbatical, have begun to return with renewed sobriety. Companies with product-market fit and early signs of repeatability1 are raising at reasonable valuations, particularly those that have products that are “non-optional and don’t suck”, are equity-efficient hard tech, or use automation and AI to improve cost structures.

Now that fundamentals are at the forefront of the venture market, I was particularly excited—and honestly a little bit apprehensive—to dig into the main topic of this letter: risk. It’s something we all feel like we understand yet struggle to articulate clearly.

There’s a particular flavor of risk that startup founders and investors use to make good decisions under uncertainty. Risk that won’t make us sound smart with talk of volatility, Gaussian distributions2, or Sharpe ratios. Startup risk has very little math and a lot of mucking about at the messy intersection of people, technology, and products.

Startup risk is existential future risk in the sense of “we just ran off a cliff with (what we hope are) the requisite parts for an airplane and are hurriedly assembling it before… *SPLAT*.” Put another way, startup risk is more concerned about future possibilities than present reality. Unlike public equities where there’s (hopefully) a real business to analyze, startups begin as “default-dead” with little to their name at present. Tomorrow is the prospect of a flying machine while today we have nothing more than a jumble of heavier-than-air parts in rapid descent.

Startup risk is also upside risk. Unlike aviation safety, startup risk is more concerned with “what could go right” than “what could go wrong.” It’s not just because founders and venture investors are inherently optimistic people: they just have a lot more to gain by succeeding than they have to lose by failing. Debt investors, in contrast, generally think of risk as the likelihood of losing money because they’re in the business of avoiding failure, rather than maximizing success. 

I don’t believe a single approach—present vs. future, downside vs. upside—is universally superior, but everyone should know what game they’re playing and act accordingly.3

This letter is focused on the risk startup founders and investors face in achieving—or failing to achieve—future upside. Though there are many different ways to underwrite risk, I’ll attempt to articulate what I believe are the principles behind successful models. Frequently, these principles justify conventional wisdom; however, important opportunities occur when principles tell us to stray from the herd—to call bullshit and refrain from acting or move aggressively while others pause.

“The single most powerful pattern I have noticed is that successful people find value in unexpected places, and they do this by thinking about business from first principles instead of formulas.”
- Peter Thiel

First principles are the “laws of physics” governing startups, hard-to-argue-with premises such as:

• Businesses exist to make money and are eventually valued on cash flows.

• Risk is uncertainty around the ability to generate large, growing, defensible cash flows.

• Startup valuations increase as they reduce risk, all else equal.

• A business with little competition is worth more than a business with lots, all else equal.

• Predicting the macro environment 10 years in the future is really hard.

Starting from such simple concepts, we can explain the structure of the startup ecosystem. These emergent properties are also how the wondrous diversity of biology—from single-celled organisms to a species capable of understanding its own genetics—follows from the relatively simple laws of physics. Similarly, I believe startup first principles allow us to underwrite everything from low-risk, incremental vertical software to high-risk, world-changing moonshots. 

To explain the emergent properties of the startup ecosystem from first principles, I’d like to cover:

1. Why startups need stage-by-stage financing, upside focus, and underwriting the future.

2. Underwriting risk with team, market, business model, and traction.

3. Pricing risk with valuations.

4. Why inefficient and illegible are exciting.

5. Doing hard things: difficult is defensible.

By articulating risk clearly and pricing it appropriately, I believe founders and investors can improve their chances of avoiding costly failures and identifying overlooked opportunities. 

Finally, I’d like to acknowledge the many prolific writers and jam sessions with founder and investor friends that fueled this letter, alongside some of my own battle scars. In each section I’ve tried to pair concepts with examples to strike a balance between rigor and practicality.4

Despite my best efforts, I’m sure there will be some things that smart, sophisticated, and successful (SSS) people disagree with and I’d love to make this the start of a discussion. Having multiple ideal points makes startups fun :)

Why startups need stage-by-stage financing, upside focus, and underwriting the future.

In the last market cycle, it seems like venture capital has been an org structure and a portfolio construction model masquerading as an asset class. The venture industry often reduced itself to asking “Will it go up?” and “How much capital can I deploy?” instead of what it should have been asking: “What makes a great startup investment?”

Visionary founders with bold proclamations of the future? Swashbuckling adventures of a small team in a garage in their heroic quest to defeat incumbents? Dramatic near death experiences? “Unique” personalities of many founders and VCs? These are all frequent occurrences that make the ecosystem fun and interesting, but not sufficient to define an asset class. 

I believe that the defining characteristics of startup investing are: 

• valuations that depend on “underwriting” the future more than “analyzing” the present,

• business plans that aim to maximize upside rather than minimize downside, and

• stage-by-stage financing

Underwriting vs. analysis

I use the word underwriting because the dominant factor in returns is the risk of some uncertain positive events occurring in the future—can the team assemble the plane before they hit the ground? In this sense, startup investing is a sort of like buying inverse-fire-insurance: disaster is assumed but the the payout is large if a company can somehow avoid catastrophe. 

Analysis, by contrast, is the ability to understand the past and present clearly with diligent effort. Perfectly valid with respect to established companies, but not helpful when the company is just an idea and was incorporated yesterday.

Once existential risk is no longer the central consideration for returns—when the company becomes “analyzable”5—investing in the company ceases to be venture capital in my view. Such an investment might be a perfectly wonderful “growth capital” or plain old equity investment, but no longer “venture capital”.

Underwriting is the art and science of assigning probabilities and prices to future events—good or bad. “Great” startup investing is not unlike “great” insurance underwriting in that the actual risk must be lower than how the market prices that risk to profit. We’ll dig into the concept of non-consensus underwriting—how people can disagree in their perception of risk—later.

Upside vs. downside focus

Unlike insurance which is fragile—small and capped upside with large but capped downside—startup investing is anti-fragile—small and capped downside with large and unlimited upside. Given the spectacular risks involved with underwriting an uncertain future, business plans that focus on small but positive outcomes likely lead to underperforming venture portfolios: failures wash out gains from winners. That’s why startup underwriting becomes attractive only when paired with business plans that aim to achieve massive success where even a 90% failure rate is acceptable relative to the returns from the winners.

Stage-by-stage vs. all-at-once financing

Because valuations start low given the risk at outset, raising capital for the company’s entire journey in the beginning would be crushing to founders’ equity ownership—raising $50M at a $10M pre-money valuation would result in greater than 80% dilution. Instead, raising “just-in-time” capital via a consecutive series of rounds solves the problem neatly: a company can aim to prove out a subset of its risks at a time and raise enough capital to do so. Then, we can repeat this process multiple times until we either fail, sell, or have a massively successful business on our hands. 

In some cases, a company is so efficient that they can achieve escape velocity with a single round of funding. More commonly, though, prospective investors at each round underwrite the residual risk in a company to justify a new (and hopefully higher) valuation.

How do we come up with valuations?

First, It’s worth talking about what should not go into startup valuations.6 Unlike valuing mature companies with predictable cash flows, I do not believe it’s possible to reach a reasonable valuation for a startup with a small number of variables divined from financial statements or the market as a whole. What is a startup worth that has $0 in revenue? One might be an infrastructure company with massive adoption of their open source product and another might be a yet-to-be-built social network for cats. When a company has revenue, should we take the “average” seed revenue multiple? That’s probably not right either, because a company with extremely experienced founders is probably worth more than one with unproven founders. 

Startup risk is subjective and nuanced, properties that resist being reduced down to simple models. At the beginning of a startup’s journey, a set of smart and dedicated, yet imperfect, humans aims to will something into existence against the forces of entropy and active destruction. The company exists in an elaborate yet un-choreographed dance with founders, customers, partners, suppliers, and employees—all with their own constantly changing psychologies—amidst the backdrop of rapidly-evolving technology and fickle capital markets.

What input to the Black-Scholes model describes the likelihood of a founder convincing their critical first enterprise customer to bet on unproven software? What part of the capital asset pricing model (CAPM) covers the risk of founder breakup? Where in a discounted cash flow (DCF) do we represent the skill of the founding technical team’s ability to successfully scope and build a product that delights that first customer? It doesn’t take long asking such questions before a sane founder or investor sets such valuation models ablaze.

Because the dominant impact on returns from startup investing is existential upside risk, I believe that getting to a reasonable valuation must depend on underwriting the residual risk in a startup’s ability to achieve its goals.

Underwriting startup risk is idiosyncratic—and I think that’s a good thing. If you talk to 20 successful early-stage investors you’ll probably get 20 different underwriting approaches, indicating that the market is either illegible (i.e. evidence is not clearly available to all participants) or inefficient (i.e. valuations do not consistently incorporate available evidence) or both. This is probably not surprising given the tremendous complexity in building a startup from scratch: each observer has their own set of heuristics for assessing value, making it hard for everyone to agree. The wide variety of underwriting models is also particularly exciting because it suggests that—unlike more legible and efficient markets—superior evidence gathering ability or superior insight are likely to generate superior returns (i.e. alpha) for both founders and investors, a topic we’ll dive into later.

That’s not to say underwriting startups is just rolling dice, though some approach it that way. Some founders and investors routinely produce superior returns and I believe a significant part of that is attributable to underwriting skill7.

To reiterate the whole point of all of this effort: by articulating risk clearly and pricing it appropriately, I believe founders and investors can improve their chances of avoiding costly failures and identifying overlooked opportunities.

Most approaches to startup underwriting I’ve seen include some combination of team, market, business model and traction8. Skill in underwriting these factors often comes from having seen many companies succeed and fail so as to be able to answer questions such as: Which founder attributes are helpful (or hurtful) in particular situations? How do we correctly anticipate changes in markets? Is a business model plausible? and What evidence leads us to believe the company has “nailed it” or “scaled it”? Not unlike artificial neural networks, the organic neural networks in the heads of founders and investors tend to benefit from lots of training data.9

Let’s cover each area of underwriting—team, market, model, traction—individually and discuss how one might evaluate them at the outset, how that underwriting changes throughout the course of a company, and common underwriting traps.

Underwriting teams: Can this team can execute this business?

Every early-stage investor I’m aware of believes that team is the most important aspect of underwriting because great founders have the uncanny ability to figure out just about any risk while those who are merely looking to get rich quickly tend to opt out at the first sign of difficulty. I wholeheartedly agree with this belief, having seen the magic of teams who can reliably conquer seemingly-insurmountable challenges.

“Great” teams are tricky to define precisely because great in one market might be ineffective in another market and traditional credentials—prestigious prior jobs or having more degrees than a thermometer from top universities—are unreliable predictors of success.

Some punt on a precise definition of “great” and opt for the “I know it when I see it” approach; however, I believe that great teams have universally important personality characteristics accompanied by skills and experience that depend on their particular market and business model.

Divining personality characteristics

Some investors have highly developed opinions regarding founder attributes they prefer. Y Combinator (YC), for example a particular viewpoint10 and Jessica Livingston, a founding partner of YC, is widely respected as having the ability peer into the souls of founders to identify these characteristics after even a very short conversation. While I was working with YC a few years back, I had the chance to see Jessica at work and honestly have no idea how she does it so quickly, but I think the strong values-alignment within the YC community has more than a little to do with her.

Without being blessed with Jessica’s level of intuition, mere mortals can try and understand behavior patterns by speaking with references who have longstanding relationships to understand what makes someone tick. Front-door references—those provided by the team—are usually so biased as to be useless; however, I’ve found back-door references amongst people I know and trust to paint a much more realistic picture. Because the technology industry is tightly networked, there are almost always co-workers, classmates, friends, or prior investors who you already know or can get to within one degree of connection in one or two messages. Given the expectation of working together repeatedly over the years, there’s a strong incentive to be truthful—it always comes out eventually—and I’ve found that people are usually remarkably candid.

When trying to understand someone, I think it’s important to answer some key questions:

• Have they walked through walls to succeed… or did they give up at the first sign of failure?

• Did they communicate problems immediately… or did they deflect and delay?

• Did they do the right thing by their stakeholders… or just the minimum legal thing?

• If they ran a company, did they carefully conserve resources… or spend like a drunken sailor?

• Were they earnest in making commitments… or did things frequently fall through the cracks?

In essence, I want to find evidence confirming (or disconfirming) that someone is moral and ethical at their core with a burning desire to succeed and the discipline to make their dreams into reality.

Market-appropriate skills and experience

When it comes to skills and experience, I believe there are multiple ideal points—there’s no one best founder archetype. Founder-market fit is the idea that a strong team in one market environment might be a weak team in another environment. Because it’s so emotionally draining to build a startup, founders in markets that feel unnatural often can’t push through the inevitable setbacks and give up. Similarly, founding teams that don’t have the experience to be credible to important customers or partners may not be able to get off the ground at all.

Here are some patterns in founder-market fit I’ve observed:

• Obsessed younger founders sometimes paradoxically have more experience than any other age group using new technologies. In the post-dot-com period, for example, early-twenties founders who grew up with the internet having spent all of their waking hours glued to their computer may have had more hours online than any other age group. Even slightly older founders may have had the handicap of “a job” preventing them from spending so much time online during the birth of early internet communities. To that end, betting on a young Harvard student named Mark to build Facebook was one of the better venture capital investments in the last few decades.11

• In enterprise sales driven markets, founders with experience tend to know how to navigate political buying processes and can overcome the dreaded “vendor viability” questions procurement departments use to harpoon early stage startups. Founders with less experience tend not to understand how to navigate complex organizations and have fewer trusted relationships to draw on when convincing large customers to bet on a small, risky startup. Dave Duffield, for example, was 65 when co-founding Workday, a HR software vendor now an ~$80B market cap public company at the time of writing.

• In vertical markets with a lot of domain-specific knowledge, founding teams with experience in a rare intersection of fields are powerful. I’ve found most commonly this to take the form of “software engineering + x”. For example, I’ve seen teams succeed where x = biology, x = healthcare, x = energy, and x = cybersecurity because of their unique ability to build software that takes advantage of their domain knowledge. Scarcity of founder insight can be a powerful force in becoming an “n of 1” company without much competition.

Traction impacts team risk

As the execution skills of a team become apparent, perception of team risk can change:

• For young or first-time founders, perception of risk often decreases rapidly with even a little evidence indicating execution ability and speed of learning. This is another reason I believe YC is so successful—it is a crucible that consistently helps teams have the most productive few months of their lives. Unproven founders who can demonstrate early signs of rapid growth even with small absolute traction substantially increase the valuation investors are willing to pay relative to the same team without those early indicators. 

• For experienced founders, investors are often willing to “give credit for forward execution”, to the high water mark of their past experience. In situations where experience really de-risks the business, this can result in valuations that seem eye-popping if one were to ignore the team and look only at traction. At some point, team risk converges to actual execution which can be a rude awakening if teams expect to be given credit far beyond their past experience.

Traps in underwriting teams

As with any form of underwriting, I’ve observed two common things that can lead to overconfidence in a team’s ability: 

• Projecting: Investors with operating experience imagine themselves in the founding role and ask the question: “Could I execute in this market?” instead of the more important question: “Can THIS TEAM execute in this market?” I’ve found that projection most commonly happens with extremely exciting market opportunities and teams that are not clearly great but also not clearly doomed. Optimism accompanying incredible market opportunities can cloud even a skeptical investor’s assessment of a founding team.12

• Credentialism: Backing teams who come from marquee companies or with degrees from top universities seems like a low risk proposition. Indeed, many founders with strong credentials who have “seen greatness” are exceptional; however, fancy logos are correlated, not causal, to success. There are many people who are extremely smart but took the safe road through life—going to top schools and top companies—because they are inherently risk averse and deathly afraid of failure. Inability to take risk or inability to emotionally recover from the inevitable setback is fundamentally incompatible with running a startup. 

• Adversity bias: Some investors take the opposite tack by only investing in founders who have had tremendous adversity of some sort. Given that there are great founders with top credentials and great founders who have absolutely zero credentials, I believe the key is to figure out the underlying personality characteristics and experience that predict success in a given market, not relying on credentials or past adversity to do that work for you.

Underwriting markets: How many customers can we serve and how much will they pay?

The trite way to describe a good market is “big”. Of course, if a market is too small, even a low risk proposition can’t generate a meaningful return. While it’s tempting for founders to go after a small but less-competitive market, there’s wisdom in the belief that “building a startup in any market is hard, so you might as well pick a big one.” Some small markets are not competitive for a reason: they’re graveyards. Such markets might be hard to operate in, the customers don’t really care about solving the problem, or the customer psychology in the market is generally averse to change. Some examples that come to mind include: fintech tools to split the check at dinner, location-based social networks to find what friends are doing right now, and healthcare solutions where there’s no incentive for the ecosystem to adopt.

How do we identify which markets are big? One common method is “top-down” using analyst reports that come up with figures for total annual spend and the compound annual growth rate (CAGR) of that spend. A startup will call the total spend their Total Addressable Market (TAM) then choose several segments of that market to get a Serviceable Addressable Market (SAM) and finally build a business model to figure out what portion is the Serviceable Obtainable Market (SOM) given some assumptions on ability to execute. 

On its surface, the top-down approach seems reasonable; however, I see two major flaws in using it for startups:

• Startups often segment markets differently than analysts do, and

• Startups are most interested when markets undergo unanticipated change.

Startup segmentation vs. analyst segmentation

In B2B markets, each type of customer behaves differently: large customers want different things than medium sized customers who want different things than small customers. Top-down market analyses tend to handle segmentation by customer size well; however, what if the customer behavior also depends on other subtleties such as the unique psychology and needs of individuals or small teams within an organization? 

Consider Airtable, a company whose product would have been thrown out of any head-to-head feature comparison with Excel or vertical enterprise software; however, for many use cases, Airtable was superior. First, it didn’t suck for end users—it was easy to use and well designed. The product had more structure than spreadsheets but was more malleable than typical vertical enterprise software. Because of the breadth of use cases Airtable supported, the company focusing on individuals and small teams who started using the product on their own rather than going after large enterprise contracts in a single vertical. Over the years, the product has become extremely popular and teams have adopted it for a wide gamut of business workflows. Only when the product became mature with a robust feature set and strong security/compliance pedigree could the company win large enterprise-wide contracts.

Had Airtable tried to own the entire spreadsheet market or various vertical software markets from the start, they would likely have lost the battle to incumbents. Had they chosen just a single vertical, they would be legitimately in a small market. Instead, they chose to re-segment multiple markets and attack incumbents from below. If early investors were short-sighted and relegated Airtable to a small segment of the spreadsheet market without anticipating their ability to convert non-consumption—selling to customers who bought nothing before—and replace various categories of vertical software within larger organizations, it would be easy to reject the company out of hand with a curt “market is too small”.

Consumer markets often work the same way: great new products unlock “latent” demand. Consider how Uber massively increased the size of the ride-for-hire market; however, trusting any market analysis that merely showed Uber taking share of the existing taxi TAM would lead you to toss the company in the “market is too small” bucket as well.

Overestimating market sizes also happens, but with less dramatic results. Nearly weekly I see a cybersecurity startup claiming to have a $10B+ market because they’re trusting analysts who amalgamate salaries, consulting fees, and products from multiple distinct subcategories. Trusting this sort of market sizing might make you feel good about an investment on paper, but when it comes to delivering revenue, customers will pay just for the software that the company can build—a much smaller figure than the analyst’s total.

Top down market reports often do not have the nuance startups need to re-segment markets, resulting in unhelpful TAM, SAM, and SOM estimates. I don’t believe analysts intend to deceive unwitting startups; instead, the primary customers of these reports are large companies, investors in large companies, and customers of large companies—simply a different audience. To startup founders and investors, incorrect market sizes aren’t a trivial matter: they’re dangerous because they lure us into investing our lives and our fortunes into a market segment that just isn’t that big—a juice that isn’t worth the squeeze—or dissuade us from pursuing a lucrative opportunity that doesn't line up neatly with how segments have been described historically.

Unanticipated market changes

What happens if a market’s segmentation is straightforward, but the market is undergoing rapid growth? Most market analyses have some way of accounting for changes to market size over time. Typically analysts will observe past changes, speak with market participants and come up with a CAGR for that market as a whole as well as individual segments within it. These sorts of growth models work well when it’s “business as usual.” 

Because of scarce resources relative to incumbents, startups can rarely operate in the land of business as usual. They more often operate in the opportunistic land of “Why now?” Why is now the time where businesses are willing to move from their old and expensive on-premises software to the cloud? Why is now the right time when employers will make their existing workforce more efficient with AI vs. opening more headcount? Why is now the time when there will be higher demand for datacenter GPUs? Startups are pirate ships finding the perfect opportunity to attack, not massive navies steaming directly into head-on conflict.

For example, consider the market’s surprise at NVIDIA’s rapid growth in data center revenue. Applying a CAGR to trailing revenue shows the following prediction: 

However, because of the explosion in demand for AI training and inference, NVIDIA’s datacenter business exploded, resulting in the following actual revenue—growth that would be completely missed by a simple CAGR model. 

I believe that CAGRs are too coarse grained to work in a world of unique insights. Successful founders often have heterodox beliefs that aren’t appreciated by the rest of the market for one reason or another. If an insight is truly unique, it won’t show up in an analyst report (otherwise, it’s not so unique, is it?). It’s this belief that founders and investors share that outsiders do not that often gives the startup just enough of an edge to overcome scale advantages of incumbents.

Bottom-up market sizing: an alternative to top-down

Instead of relying on analyst-derived segmentation and growth rates, “bottom-up” market sizing aims to create a market hypothesis that a startup can validate with evidence. The inputs to a bottom-up market sizing model are concrete, have ground truth to back them up, and generally take the deceptively simple form for any company with a “build it + sell it” revenue model:

market size = # customers * $/customer

The most important aspect of that equation is $/customer (i.e. the amount of money each customer will pay us per year) because founders must first spend a lot of time with customers to figure out which ones will actually pay them and how much they’re willing to pay.13 Sometimes this answer is segmented (e.g. mid-market enterprises will pay a different amount than large enterprises) and our overall market size model becomes the sum of multiple subsegment market sizes. At the end of this exercise, though, the startup will have a high-confidence estimate of what customers in the real world will pay for their product—validated price points.

To determine # customers, startups usually have a pretty good idea how the market segments out after asking customers to pay for a product—it’s the type of customers who at least don’t say “no”.  There are often common characteristics behind the customers who say “yes” where we can rustle up publicly available data to estimate how many customers share those characteristics. 

There are many other indirect revenue models, such as advertising, lending, or market making where the same principles apply: validate how much money you make each time you turn the crank then figure out how many times you can turn it per year.

Ideally, this exercise shows that the most immediately serviceable customer segment could result in billions in gross profit. If that’s not the case, we can ask concrete questions about how we might expand our initial market:

• Do we believe the number of customers in the segment will change (e.g. growing market)?

• Do we expand to other segments?

• Can we increase pricing or can we sell more stuff to the same segment?

Questions like these neatly connect market size with business model so founders can estimate the impact of narrowing or expanding scope. Generally speaking, the more different products we sell or the more customer segments we serve, the greater the business model risk. Ideally we can focus on doing one or a very small number of things very well and can get to quite a bit of revenue before we are forced to do more.

A startup that succeeds will eventually tap out its market and aim to expand in some way. Granular bottom-up market sizing gives us concrete data on when we might have to do that. At $10M, $100M, $10B, or $100B in revenue? Knowing the texture of the market allows us to propose a business model, informed by evidence, that has the possibility of getting to $200M+ gross profit with a path towards $1B+, which is approximately when I believe a business is ready to go public.

Underwriting business models

A business model is a startup’s “critical path to paradise”: the key hypotheses that the startup must validate to become a large and sustainable business—a fully assembled and working flying machine. Critical path hypotheses typically include some combination of:

• product—What will our product do and what will it not do? How will that change over time? How long will it take to build? Are there hard technical or scientific challenges we must overcome?

• customer psychology—Who will buy it and why will they buy it? How will that change over time?

• distribution strategy—How will we reach customers and get them to buy our product? 

• unit economics—How much will customers pay? How much will it cost us each time we sell our product? How does that change with scale?

As with hypotheses in other fields, business models can be consistent with reality, inconsistent with reality, or somewhere in between. If a business model is inconsistent with reality we have a 0% probability of proving it out. Any eventual outcome multiplied by 0% equals zero—clearly not a great investment strategy—so we want to know as quickly and as cheaply as possible.14

At the very earliest stages of a startup with very little evidence to work with, how do we pose hypotheses that have a reasonable chance of being true? Luckily there are a few techniques:

• Finding reference classes: Even without evidence, we can look at norms in the industry, e.g., typical earnings for a salesperson who sells similar products to the target customer base, amount of effort necessary for other engineering teams to build similar products.

• Early experimentation: We can get outside the building and talk to prospective customers. For example, we can determine the price a set of customers is willing to pay for a product. We may find that for that price, there’s no way we can afford the sales team required to sell the product to them (common in selling software to small businesses or individuals). In this circumstance, a startup must go back to the drawing board to identify a modified set of hypotheses that have a higher chance of success, for example by selling an expanded product to the same customers, choosing bottom-up distribution, or moving up-market to larger customers.

One flawed approach to formulating hypotheses that can never be proven wrong: be vague. A correctly formed hypothesis is a falsifiable belief—there is possible evidence that could prove it wrong.15 A vague unfalsifiable hypothesis is not specific enough be proven wrong and is therefore not very helpful. Of course, no well-meaning investor or founder would knowingly build a business model around unfalsifiable hypotheses; however, it’s easy to accidentally do so using big phrases that sound true, but aren’t.

“Don't use a five-dollar word when a fifty-cent word will do.”
- Mark Twain

Big words and phrases often serve as a “cognitive stop sign”: people believe they understand what such phrases mean so they don’t think through the details.16 Consider common startup strategies for distribution such as “product-led growth” (PLG) or long term defensibility such as a “data network effect”. What is the falsifiable hypothesis behind each? Not obvious.

Simple language makes it clear what the falsifiable hypothesis is and what evidence we expect to observe.17

For example, with respect to the hypothesis that a company can attract customers via PLG, instead of “sounds right”, we can restate the hypothesis as: “Security engineers working at large enterprises will go onto a website, install some software, and connect it to their production cloud infrastructure.” The response to anyone who has operated in this environment is clearly "no @#$%ing way” because anything touching production infrastructure is tightly controlled and must follow a stringent approval process in most organizations. In this case, the whole business model becomes invalid—the company needs to rethink its distribution strategy.

In another example, while the concept of a “network effect” is intellectually satisfying, we can restate the hypothesis as: “Customers will be happier with our product if we had more customers and less happy if we had fewer customers.” In some cases it might be true (e.g. communication tools or proprietary AI models that depend on lots of training data), but in many cases a product’s value is more or less the same regardless of the number of customers. In this latter case, there’s clearly no network effect. 

A false belief in a network effect probably won’t directly kill the company; however, investing in a data mote where it doesn’t matter to customers consumes valuable resources (i.e. it carries opportunity cost vs. investing in other forms of defensibility). Each hypothesis we add to our business model decreases our probability of success18, so startups need to figure out what actually matters to their business: the critical path to paradise, not just a path to paradise. 

Underwriting traction: Using evidence to understand residual risk

After articulating the hypotheses behind the market (a big one) and business model (the critical path to paradise) alongside the implicit hypothesis that the team can operate at increasing levels of scale, the quest of a startup is then to produce evidence confirming each hypothesis, retiring it as an area of risk.

At any time before we’ve fully retired all areas of risk, we can look at the strength of evidence we do have to determine residual risk—how much risk is still left in the business. For example, looking at a company’s product, customers, a sales team, marketing channels, and so forth, we can evaluate each hypothesis to reach one of several conclusions:

• Insufficient evidence: all hypotheses start like this at the outset, e.g. we haven’t tried to acquire many customers so we have an unknown customer acquisition cost (CAC).

• Disconfirming evidence: we believe some hypotheses are under no circumstances achievable because we’ve tried but can’t figure it out, e.g. we’ve experimented with many marketing channels and they’re all unprofitable, there are integration choke points that won’t let us pass, or our sales team is unable to succeed against competitors.19

• Confirming evidence: we’ve tried and proven it, e.g. we have customers that are paying while reliably renewing and expanding their contracts with positive net revenue retention.

Hypotheses with insufficient evidence are startup risk. There is some probability we won’t be able to validate them—we may not be able to assemble the flying machine before *SPLAT*. 

When faced with substantial disconfirming evidence, a startup is forced to do some form of pivot. This can be minor, such as moving up-market from small business customers to mid-market customers, or major such as focusing on a different product altogether.

Meanwhile, hypotheses with substantial confirming evidence reduce risk—that part of the flying machine looks like it’s working.

To estimate residual risk, we can evaluate the strength of evidence for each individual hypothesis in the business model alongside the hypotheses that the team can execute and that the market is large. Combining all of these together, we gain an understanding of the overall likelihood of getting to paradise. We’ll go into more detail below on using risk to underwrite valuations, but the most important point is: the less risk for a given company the higher the valuation.20

How do founders choose which risks to focus on first?

If valuations are low when there’s a lot of residual risk and get higher as the risk is proven out, then of course founders should try to spend their time reducing risk. This requires capital to pay employees, buy equipment, and run marketing programs, amongst other expenses. 

At the outset, raising enough money to prove out all areas of risk is typically either impossible—no set of investors may be willing to write a check that large—or wildly dilutive to founder equity. Instead, most startups pursue stage-by-stage financing, raising a little money at a time to retire the largest areas of risk then raising more to retire additional areas of risk, and so on. Stage-by-stage financing is a series of “just-in-time” bets instead of a single massively risky bet.

There’s a lot of nuance in choosing how much evidence to collect and when to collect it:

• Risk is not binary—one or zero—but a sliding scale based on how much evidence exists. For example, small scale profitable customer acquisition via a particular channel provides some evidence that a go-to-market hypothesis is valid; however, there still remains risk that acquisition costs will increase as the channel becomes saturated.

• Each level and type of evidence has a different cost— For example, we can gain a little confidence on pricing by getting verbal agreement from prospective customers or we can find comparable companies to benchmark against.21 Stronger evidence to validate pricing requires us to actually build the product, an activity that may be 10,000x the cost!

• Some evidence has pre-requisites—Some evidence can be acquired at any time while other pieces of evidence must be serialized. For example, we can’t tell if customers love a product before it’s built and it’s hard to sell ads on a social network before anyone is using it.

• Multiple paths—Sometimes multiple approaches to producing evidence for a hypothesis can be worked on in parallel. For example, we may have several options for customer acquisition where only one needs to scale for the company to succeed.22

Combining these effects together, we get some very difficult decisions to make. For example, verbal pricing validation is better than no validation, one customer buying is better than no customers, and five customers is better than one customer. Evidence of customers in multiple segments (e.g. customer sizes or industries) is often better than having all customers in a single segment because it provides evidence that the addressable market is larger; however if each segment has only a small number of customers as evidence of validity, that may actually result in higher risk than if all customers were in a single segment. 

One succinct way to describe startup strategy is: finding the cheapest level of risk reduction per marginal dollar. As you might imagine, I believe it’s important to carefully craft the right risk reduction plan for a business to maximize the amount risk we can prove out with a small amount of money.

Despite this complexity, I believe there’s a risk reduction roadmap that works well for startups who have considered buying processes—where customers think a lot before buying23:

1. Verbal validation—founders have spoken to many customers and we have reasonable confidence that all elements of the business plan are plausible.

2. Product-market fit—we have clear evidence that customers love our product.

3. Repeatability—our sales and marketing engine can reliably get customers to buy with good unit economics (i.e. fast payback period of sales and marketing investment)

4. Scalability—we can get lots of customers to buy while maintaining good unit economics.

Framing the problem this way, it’s obvious why a company should focus on one stage before moving onto the next: it’s expensive to hire a large sales and marketing team if we can’t make the unit economics work on a small one (the quest for repeatability). It’s expensive to hire any sales and marketing team before we have a product (the quest for product-market fit) and it’s expensive to start building a product if we don’t know if people want it at a price we can profit from (the quest for validation).

It’s therefore important for founders to know what quest for evidence they’re on at any given time. Going on a future quest—such as prematurely trying to scale a sales team—before completing the pre-requisite quest—such as building a product customers love—usually leaves behind a very expensive tombstone.24  The model above starts with cheap evidence (verbal), moves on to moderately priced evidence (building a minimum viable product), before building an expensive go-to-market motion and undergoing a colossally expensive scaling process.

Another common deathtrap is to pursue multiple divergent paths simultaneously. For example, B2B companies must often decide whether to serve small, mid-sized, or large enterprise customers. Trying to sell to both mid-sized and enterprise at the same seems like gathering parallel evidence for a profitable sales motion; however, these customers often have different product requirements. This divergence leads to the very expensive strategy of either building two products or trying to sell a product to customers who will be dissatisfied with it. I believe it’s better to serialize customer types so you can delight one before moving onto another.

Pricing risk: from underwriting to valuations

Assuming it’s at least plausible that a startup can succeed, how should founders and investors fairly value the company? It depends on both the residual risk and potential rewards. A company that has proven beyond a reasonable doubt that it will be a large and sustainable business will command a higher valuation than one with more uncertainty. Similarly a company with the same amount of risk but in a large market should be valued more than one in a small market.

To add some quantitative rigor, I like to think of valuations as a probability distribution across outcomes. To do this we can create representative scenarios then assign probabilities, terminal values and additional dilution amounts to each.

While a multi-billion dollar business is always the goal, underwriting multiple scenarios allows us to model various off-ramps. A company with a great product but without distribution might be able to sell to an acquirer at an attractive price—an especially common outcome in cybersecurity, an industry that has very high sales and marketing costs given the complexity of the technology.

For example, consider a hypothetical company that is generating some revenue in a large market but success is far from guaranteed. An investor may underwrite the following scenarios (we’ll discuss how to come up with terminal values and probabilities in detail below):

To calculate the value of each scenario, we can multiply the terminal value by (1 - dilution) to determine what the effective return for investors equity might be. For example, in the base case, $500M * (1-40%) = $300M.

From there, we want to weight each return by the probability of the scenario occurring to determine how much each scenario contributes to the eventual outcome. In the base case above, $300M * 20% = $60M.

Summing each scenario contribution, we can produce an “expected value” or the average return equity holders can expect. Of course each specific result will be different from this average—craters will be lower and bull case outcomes will be much higher.

If we believe that the distribution above accurately represents reality, the expected value of the company is $152M. It would be unprofitable for investors to invest above the expected value of a company: by definition, that would be expecting to profit less than $0 on average which does not compensate them for the potential failure, effort, and cost of capital to invest in the business—a negative net present value (NPV).

Instead, investors will aim for a return target consistent with their venture model. Because there’s a lot more uncertainty the earlier in a company we are, investors should apply a larger margin of safety in the form of a higher return target. For example, if one were to focus on bold and very early stage bets, I think 10x is an appropriate target. In that circumstance, investing in the company at a $15M valuation or less would be attractive relative to the 10x return target.

Because ~90% of the expected outcome is contained in the base and bull scenarios, I believe many experienced investors simplify their thinking and focus only on the likelihood, dilution and terminal value of those.

In the table above, all of the complexity of underwriting we discuss in prior sections is reduced to a small set of numbers. While such simplification may seem like a blunt instrument—and it is—I believe it strikes a happy balance between under-modeling and over-modeling. Because of the level of uncertainty we’re dealing with, adding more bells and whistles to the model won’t make it more effective.

Traps in building outcome distributions

Though crafting scenarios that best reflect reality is a subjective process where SSS investors can disagree (we’ll cover how that happens below), let’s start with common traps to avoid:

• Too much credit for forward execution underestimates risk when investors believe a risk is minimal when it’s actually much harder to solve. Avoiding this one comes from the experience of knowing which aspects of building a particular company in a particular market are hard, and which are easy. 

• Benchmarking to high acquisition prices may be tempting particularly in the wake of high-dollar outcomes in a particular market; however I’ve found acquirers are fickle and there’s wisdom to the phrase “companies are bought, not sold” implying that a startup has very little control over such outcomes. It’s much more reliable to build a large and sustainable business and perhaps the company will get an attractive offer along the way.

• Anticipating large terminal values in bull markets can justify just about any entry price; however, at some point terminal values will be based on fundamentals and timing the market to exit only when things are frothy is very hard.

How do we sanity check terminal values? 

Craters are self explanatory and small acquisitions right around the company’s liquidation preference are easy to account for. For success cases, though, we have to make some big assumptions. If we believe that a company’s valuation will eventually be based on its cash flows, the terminal value is really just the output of a valuation model whose major input is cash flow.

While there are valuation frameworks of varying complexity, for our purposes we need something approximately reasonable, not precisely accurate. How much will a company be worth at exit 10 years from now? Nobody knows, but if it grows from $0 to $200M in revenue, our stock should be worth a heck of a lot more than we paid for it.

Ideally, we choose a slightly conservative terminal value so we can be surprised to the upside if growth or market sentiment is better than we think while avoiding being surprised to the downside if the opposite occurs. Here are a couple of approaches:

Equity Risk Premium: We can take the current interest rate (4.11%, 10Y treasury as of 1/22/24) and add an equity risk premium or the amount the market pays us for accepting the risks of stocks vs. US government debt. I believe a reasonably safe long-term average is 6%, resulting in a total “yield” of ~10% as of the time of writing.25 To normalize growth rates, one approach is to use 3 year forward figures.26 In this case, a company doing ~$198M in gross profit today that is growing at 15% annually would be doing ~$300M gross profit in 3 years. A 10% gross profit yield (10x gross profit multiple) produces a ~$3B valuation under this model. Because many startups aim achieve this level of gross profit and growth when building their initial business plans, I believe $3B is an acceptable bull scenario terminal value. 

If we want to estimate the effect of lower or higher interest rates, we can easily calculate them. At 6% interest rates (8.3x multiple), that company above would be valued at ~$2.5B and at 2% rates (12.5x multiple) ~$3.75B. Note that these valuations are different, but they’re not that different. Under any rate environment, it takes a LOT of gross profit to substantiate some of the lofty valuations that were being underwritten in the last few years. 

Gross Profit Multiple: We can also use public market comps. Similarly let’s use the ratio of valuation (EV) to gross profit to normalize gross margin profiles across companies. To justify a terminal value of $3B, assuming the same 15% growth rate, based on the chart below we expect a ~15x multiple or ~$200M in gross profit—it’s wonderful when two valuation models align!27

Of course, when underwriting terminal values, there are a number of traps to be wary of:

• Large operating expenses destroy value—in these valuation models, we assume that gross profits are “allocatable cash flow”—cash that the company can invest in growth. However, if a company has large costs that are required to run the business which are not captured as part of their gross margin or its growth investments are inefficient, they will likely underachieve the benchmark multiples above. For example, if a company’s sales process is extremely complex relative to deal sizes, allocating cash to the sales team will result in little revenue growth. Similar value destruction can happen with excessive stock based compensation (SBC) or if an engineering team is not very good and needs to spend a lot of money to accomplish even minor tasks.28

• Public market comps change over time: Because we’re aiming to make an investment that will pay off years in the future, we cannot assume today’s valuation environment will persist. If using comps from a frothy environment, you can get tricked into wild terminal values. Using long-term normal valuation multiples while adding in a margin of safety, you may seem overly conservative, but you’ll avoid being rudely surprised when valuations revert to the mean.

• Eventual market size really matters—it may be tempting to invest in a company whose eventual market size is only in the hundreds of millions if they are in a non-competitive market—if we can underwrite a great outcome with only $200-300M in gross profit, why do we need a larger market size? That’s because “terminal” is a misleading term. An IPO is just another funding round. A company that executes into a big market can substantiate high growth rates for some time to come; however, if a company saturates its market, growth will atrophy and its valuation will suffer.

When the stars align, we can build a legendary company that persists for decades—a company that can exist in the pantheon of “20-year compounders”. Businesses that can maintain significant growth for decades eventually make $200-300M gross profit look tiny and can far exceed our $3B terminal value. Consider the long-term success of Google, Palo Alto Networks, Meta, and other durable winners. In Amazon’s case, over 99.9% of valuation growth occurred after their IPO.

Such durability is the substance legendary founders and venture funds are made of.

Risk is in the eye of the beholder: How do we assign probabilities to each scenario?

I do not believe there is a single formula to convert specific risks a company faces into precise probabilities at the early stage; however, I believe that the more clearly we can articulate risks, the more accurately we can estimate the company’s likelihood of achieving each scenario.

Returning to the example company described above, we can use our simple valuation model29 to transform each scenario into a more precise question: what is the probability of the company getting to a particular level of gross profit?30

Just as insurance underwriting starts with an actuarial model, I like to do the same with startup underwriting to make sure we’re anchored in reality. Specifically, I like to use historical base rates of dilution and scenario probability for companies at a given stage31 as a starting point for underwriting models.

While fires are largely unpredictable, we can expect rates for fire insurance to be much higher in homes with wood shingled roofs adjacent to dry brush than those with tile roofs in dense urban environments. Similarly, with startup underwriting, we can adjust our base rates to fit a given company’s particular risks. For example, for a seed stage company that is reliably signing $50k gross profit annual contracts with customers today, we may ask a more precise question about the bull case scenario: what’s the likelihood that they can grow to $250k annual contracts and sign 800 of them to get to $200M in gross profit with a clear path towards $1B after that? 

Each investor who approaches this question will bring their own principles and beliefs as they look at the team, market, business model, and traction. They may seek to understand customer psychology, spend a lot of time with the founders, or have their own thesis on the evolution of the market. I believe that there are many ways to do venture right—given the complexity of underwriting startups there are multiple ideal points, not a single optimal approach. There is no magic formula.

The approach that can let you sleep at night is to approach underwriting with a lot of intellectual humility and assume wide margins of safety: reasonable terminal values, low bull case probabilities, and high return targets.

Even if they are all looking at the same team, market, business model and traction, investors can substantially diverge in their underwriting. Here are a few places I’ve seen that happen:

• Re-segmentation (e.g. by size, vertical, requirement)—startups often attempt to redefine a pre-existing market either by pulling a subset of customers away from their current solutions or by converting non-consumption. Sometimes this is the result of a heterodox insight not widely appreciated by the rest of the market, for example, that a subset of customers are so fed up with how they do things that they’re willing to adopt an entirely different approach.

• Slow vs. never—Some industries adopt change slowly or unpredictably, leaving behind a litany of dead startups that justify a belief that the industry will never change. There’s an oft-cited phrase, “being early is indistinguishable from being wrong” that captures this perfectly. For example, enterprise cloud adoption has been well underway; however, hospital cloud adoption is still in its infancy. If an investor assumes “never” they will underwrite a high risk for any company selling cloud-based software to hospitals; however, if an investor believes hospitals were just slow but now deeply care about reducing IT spend post-COVID, then they may assign a much lower risk.

• Market trajectory—Not everyone shares the same beliefs about rapid changes in markets. For example, one of the most spirited debates right now is “Do you believe that generative AI is a megatrend or is it overhyped?” Based on that answer, you’ll underwrite wildly different demand for AI chips, infrastructure software, and the like.

• Founder background—as we discussed in the team section above, when founders aren’t from central casting, investors tend to form divergent opinions based on their own idiosyncratic opinions of what a “great” team looks like.

• Esoteric markets—when investors don’t understand the dynamics of a market, they assign a high risk to it by default. Many markets are large but, for one reason or another, investors have not had much exposure. I believe there are many great opportunities in markets that a). have a lot of cash sloshing around, b). have not seen much venture investment, and c). are facing some sort of technological regime change.

When investors underwrite different levels of risk, they can justify different valuations.

Because each investor uses a different qualitative model and underwrites to a different return target, there’s a wide variation in valuations investors are willing to assign to a given company at a given time. There are so many subtleties in how people underwrite risk that I believe we’ll be in an “inefficient” market for a long time to come.

Consider the company from above: if another investor underwrote less risk (i.e. higher probabilities of achieving success as shown in the table below) they might be willing to invest at a $40M valuation assuming the same 10x return target, given the same terminal values and dilution estimates.

After all of this, you may be tempted to believe that venture underwriting is bullshit when multiple SSS investors can look at the same company and come up with wildly different valuations. If we were playing dice, that would be true; however, given some investors ability to repeatedly invest in outlier companies at valuations that produce healthy returns, I believe there is a reasonable case to be made that tremendous skill is involved.

Inefficient + illegible = exciting. 

As we discussed above, startup underwriting is often illegible32 (i.e. evidence is not clearly available to all participants), inefficient (i.e. valuations do not consistently incorporate available evidence) or both. This means that different abilities of gathering or incorporating evidence will lead to different perceptions of risk. As we’ve just shown, differences in perception of risk can substantially impact what you think is a reasonable valuation.

“There are two requirements for success on Wall Street: the first is to think correctly and the second is to think independently.”
- Benjamin Graham

I believe that the best startup outcomes come from a non-consensus, and right, understanding of risk. If an investor overestimates risk, they will either opt-out entirely or underwrite a low valuation. Consider how many dismissed SpaceX in its early days as nothing more than a flight of fancy. In these circumstances, someone who correctly understands risk can invest at a “market” valuation, but will be able to underwrite far higher risk-adjusted returns. When your understanding of risk is correct, finding opportunities that “are good but look bad” is great venture capital investing. 

As a founder, similar things occur: if prospective competitors overestimate risk in your market, they may avoid competing altogether (good!). It’s then your job to convince investors of a correct understanding of risk so you can avoid having your valuation be penalized by false perceptions. Great startup pitches clearly articulate the case for the company being a relatively low-risk proposition in the form of a compelling, psychologically convincing narrative.33

What happens if other investors underestimate risk? They will underwrite higher valuations than you think are reasonable (danger!). Because it’s not feasible to short early stage startups34, the only safe response is to step back while others proceed. Sometimes you may be wrong because you were the one overestimating risk, but that’s OK—call it a learning experience. You’re only penalized by the investments you do make, not the ones you don’t. In fact, Bessemer Venture Partners, a well regarded venture firm, jokingly memorializes their “anti-portfolio”—the now-successful companies where they passed that includes Apple, Google, AirBnB, and Zoom. More evidence supporting the claim that venture capital is inefficient and illegible: even very successful investors have embarrassing anti-portfolios.

Realistically speaking, because of the competitive dynamics of venture capital, investors are partially price takers—they are one of many participants in a sort of auction where they can choose to participate at a given price but not set the price itself. I say “partially” because, as founders, we want the best valuation for our company, but we also know that not all investors are created equal. Some are wonderful to work with and can meaningfully help a company’s chance of success while some are actively destructive. Personally, I’d rather compromise a little on valuation to have a great investor than just auction to the highest bidder. 

Because of these dynamics, I believe that the best investors only invest when:

• they have some sort of edge in underwriting risk, 

• where they can help meaningfully improve the likelihood of success, 

• or where founders are willing to eschew taking the highest valuation bid to work with them.

Easier said than done: in the heat of a hot deal or a frothy market, it’s tempting to buy into hype.

Emotions are untrustworthy investing partners

Underwriting is not a cold, rational activity—given the complexity of factors, it inherently takes the properties of human psychology. For example, if you like a founder, chances are that their customers and future employees will also. 

Emotions can also be decidedly unhelpful: little does more to impact human psychology than hot markets—concentrated startup founding and investing activity in a particular category. 

Hot markets create a double-headwind for returns: high entry prices and lots of competition. Because hot markets tend to spawn a lot of competitive companies and fund them aggressively, we should adjust our valuation model to decrease expected values relative to normal markets (competition drives up sales and marketing costs, customers have more leverage in negotiating down gross margin, and strategic acquirers have more options to pick from). At the same time, frenzied venture investors drive up entry valuations. Both factors compress the risk-adjusted return we can underwrite (not good!).

“If you follow the mainstream, usually the margins are very small.”
- Francis Greenberger

Founders and venture capitalists aren’t dumb, so why does herd behavior in hot markets persist? Follow the incentives.

Hot markets have momentum. If one investor leads a Seed round today, there’s likely to be another investor leading a Series A at a higher valuation not too long after. On paper, this is good—you got a markup on your investment in a very short period of time. The gods of IRR (internal rate of return) and TVPI (total value to paid in ratio) are appeased. As an early career VC, you get promoted at your current firm or get a better job at another firm. Life is good.

Similarly for founders on the sidelines, seeing others get lofty valuations with little traction is the siren-song for starting a new company.

Investors and founders don’t make money on unrealized gains; they make money on outcomes.

“No matter how good a year you’re having, someone out there made more money than you specifically because they made a lucky bet with an irresponsible amount of leverage.
But they won’t do well forever.”
- Byrne Hobart

Of course, sometimes founding or investing in hot sectors pays off: a company succeeds despite competition by gaining an unassailable advantage or they get lucky by exiting at the right time to a strategic acquirer willing to pay up. Much of the time, however, competition makes it hard to succeed or the market turns out to be not quite so large as everyone hoped. Companies who raise a lot of money but are unable to show cash flows typically become walking wounded, unable to reach the orbit of the public markets or to command high acquisition prices. These companies either go out of business or sell at not-so-attractive valuations.

To reliably achieve great outcomes—liquid stock for founders, employees, and investors—what we really care about is for our investment to have large, growing, and defensible cash flows. Hotness is not causal to cash flow.35

When trying to invest in peculiar companies that have bold, but under-appreciated visions—“is good, looks bad”—we must be prepared for a tepid funding environment until greatness is obvious36 or perhaps the company may not need to fundraise because it can grow extremely efficiently. In either case, investors must be prepared to not see a markup for years. Unfortunately, this is suicide for an early career VC hoping for fast markups.37

If you have the long-term orientation and emotional stamina for non-consensus startups, sometimes leaning into obscurity is a good strategy because it takes time to build an unassailable advantage. Staying off the radar avoids attracting the ire of large incumbents or spawning copycat competition. Such companies may avoid announcing their funding at all, keeping a low profile amongst investors while focusing on their customers.38

How does one reliably find non-consensus companies? By definition, they can’t be the ones everyone else is talking about. One pattern is to find sectors that seem really boring, but where a lot of money sloshes around. Attractive sectors tend to be more picked over but boring is beautiful! You can’t focus on one sector for too long, however, because once a few successful companies emerge, it may become hot. Consider the relative obscurity of software for synthetic biology, healthcare, and logistics a decade ago compared with the aggressive venture investments in the last market cycle.

A second, and perhaps more durable non-consensus strategy is to focus on companies that do things that seem really really hard.

Doing hard things: Difficult is defensible.

It’s very tempting to get excited about opportunities where a small team can execute and generate so much high-margin revenue that growth fund analysts begin to salivate. Who wouldn’t want that? However, that 90% gross margin software business that’s not distribution-hard, technology-hard, product-hard, or some other form of “-hard” will likely face contribution margin erosion after accounting for sales and marketing costs when competitors see such a juicy business and aim to duplicate it.39 Consider Hopin, the high flying COVID-era virtual events company that was once valued at $7.8B but sold for $15M after customers began transitioning back to real-world events and intense competition from Zoom and Microsoft.

Hard things are rare. Rare things accrue long-term value. Here’s a tautology: if something is hard, it’s not easy. Competitors can duplicate easy things but they can’t easily duplicate hard things. A simple rule, but easy to forget.

Some types of hard things need tremendous amounts of capital. I’m glad, as a species, we pursue these sort of hard things but they’re difficult for early-stage venture underwriting. 

I believe the best type of “hard things” for early stage investors are those sorts of problems that exceptionally talented and dedicated teams can solve. A consumer product that is hard to duplicate because it has an exceptional brand or design. An enterprise infrastructure or hardware product that requires highly specialized technical expertise. A marketplace, communications tool, or social network that has a strong network effect.

Consider one of the greatest “hard thing” entrepreneurs currently operating:

“I was at a lunch with Munger in 2009 where he told the whole table all the ways Tesla would fail. Made me quite sad, but I told him I agreed with all those reasons & that we would probably die, but it was worth trying anyway.”
- Elon Musk

This is not just a flippant comment. Elon’s persistence in the face of Charlie’s correct illustration of risk captures underwriting vs. analysis brilliantly. Underwriting tells us that it doesn’t matter if something might fail so long as the outcome compensates us for the risk. Few are as willing as Elon to invest their life and fortunes behind highly risky futures with massive potential outcomes. 

Revisiting our underwriting model, distributions across outcomes work for both mundane and wild bets—you just have to underwrite appropriate scenarios. For example, a moonshot bull case may assign a low probability to a much higher terminal value. If it’s capital efficient, we may underwrite a similar amount of additional dilution; however, if it’s going to take a lot of money to get there, we might underwrite much higher dilution. Given the large pot of gold at the end of the rainbow, even modeling a 95% failure rate with 70% dilution in the success case produces a $9B expectation value! 

Of course, underwriting these sorts of outcomes for every team and company would be reckless; however, Elon is one of the most dogged entrepreneurs in history. Betting on him has paid off early backers beyond this humble model’s expectations.

Four types of investors: Dreamers, Winners, Cynics, and Shorts

One way to segment investors is by two dimensions: their accuracy of understanding risk and  their level of positive outlook.

In startups, it’s not feasible to bet on failure—so cynics (blinded pessimists) and shorts (clear-eyed pessimists) sit on the sidelines, marinating in their criticisms. Above, Charlie is acting as a clear-eyed pessimist: he agrees on the specific challenges facing the company but underwrites higher risk with no participation in the market. He won’t lose money if Tesla fails and he won’t make money if it succeeds.

As founders and investors accumulate experience, we live through countless failures. Even though we understand some rate of failure is to be expected, each smoldering crater really hurts. Our emotional tendencies to loss-aversion can overtake rational underwriting—it’s easy to project painful past failures onto new startups even when it’s unreasonable to do so. Clear-eyed optimists can, over time, become late-career blinded pessimists.

Founding and investing in startups is a “long-only” game, which is to say that it’s inherently optimistic. Optimism, like pessimism, can be accurate or delusional. 

For investors, I believe being a clear-eyed optimist is the only way to reliably win over the long term: accurately acknowledging risks but stepping onto the battlefield when the situation is favorable. Investors who are dreamers (blinded optimists) tend to run after hot markets and pay high prices relative to residual risk. Dreamer investors might win in the short term, but over the long-term, they’ll likely get blown up.

For founders, sometimes being a blinded optimist is helpful. Startups are emotionally crushing and it’s unclear how many would pursue the founder’s journey if they knew the road ahead. My wife, who is currently running a Series A health-tech startup, loves a quote that resonates with every entrepreneur: “We do these things not because the are easy, but because we thought they were going to be easy.” Many seemingly-impossible businesses are started by people who claim later, once they’ve become a little more clear-eyed, that they never would have founded the company in the first place had they known how hard it would be.40 Consider NVIDIA’s present success relative to how much pain it took to get there:

“I wouldn’t do it… building a company and building NVIDIA turned to be about a million times harder than I expected it to be… If we’d realized the pain and suffering and just how vulnerable you’re going to feel… nobody in their right mind would do it.”
- Jensen Huang

For founders, blinded optimism works because hope can spring eternally. Bank accounts do not, so it’s a less helpful disposition for investors.

There’s a YC phrase I love: “Companies fail for only two reasons: founders give up or they run out of money.” Even if you are a dreamer, so long as you can stay cheap and make enough money to be “ramen profitable” or “default alive” you can live to fight another day.41

But what about founders with clear-eyed optimism from the outset? Many experienced entrepreneurs start their second, third or fourth companies knowing exactly how painful the journey will be: you can almost hear the sorrow in Elon’s voice yet still he pushed ahead. 

Similarly, why don’t founders who start blinded but eventually become fully clear-eyed give up? Why didn’t NVIDA fail multiple times over given how hard it was? In Jensen’s words: “My will to survive exceeds anybody else’s will to kill me.”

For a strikingly large number of people, the will to progress triumphs over any personal discomfort. I’m grateful for all of those that knowingly start or continue the founder’s journey.

“Here's to the crazy ones. The misfits. The rebels. The troublemakers. The round pegs in the square holes. The ones who see things differently. They're not fond of rules. And they have no respect for the status quo. You can quote them, disagree with them, glorify or vilify them. About the only thing you can't do is ignore them. Because they change things. They push the human race forward. And while some may see them as the crazy ones, we see genius. Because the people who are crazy enough to think they can change the world, are the ones who do.”
- Steve Jobs

As I reflect on the past two quarters and plan how we position ourselves for success in the future, I’ve been asking myself several questions that I’d also like to share with you:

1. Can morality maximize growth?

2. Where do durable cash flows come from?

3. How do companies achieve efficient growth?

4. Are there durable cash flows and efficient growth in AI?

Can morality maximize growth?

As one friend and public market investor recently said to me, “capital is a claim on future human productivity.” When we look at how investors and society, as a whole, allocate capital, it is not merely a question of seeking returns, but at the heart of moral philosophy. Capital allocation impacts how billions of people today (and even more in the future) spend their lives.

What a market rewards is where capital will be allocated and, ultimately, where human effort will be spent. Will we incentivize meaningful work that creates broad prosperity, or will we dig ditches only to fill them back in?

Furthermore, this process of human effort directed by capital allocation is recursive: the individuals and institutions who accumulate capital will have their values and morals magnified by their allocation decisions, in turn, impacting what the next turn of the market whispers in peoples’ ears.

Many in Silicon Valley (and increasingly many more technology hubs) have benefitted from a half-century of techno-utopian, evidence-driven, and generally-pro-social iterated capital accumulation and deployment cycles. Founders start companies, they and their investors make money, both, in turn, allocate capital to the next generation of founders starting companies who share their values, those founders and investors make money, and so on…2

This sounds like a reasonably stable loop; however, during two times in recent memory (the dot-com boom and whatever we’ll call the 2020-2021 period of exuberance3), something happened. Instead of “normal” clear-eyed technology company building, the whole market went bananas deploying capital in a fit of speculative frenzy, paying more attention to nominal valuations going up and to the right4 than why they were going up and to the right.5

As long-term investors, our figure of merit is growth in invested capital over extended periods. In seeking growth, we can—generally speaking—invest in assets that appreciate because of improvement in their expected cash flow characteristics (“fundamentals") or because they become more desirable by the market (“capital flows”). For equity investing, prices go up because a company does well, or it gets popular: investors bid up prices and are, therefore, willing to settle for a lower return relative to cash flows.

During boom times, most/all companies become popular, making it difficult from the outside to separate returns attributable to popularity from those attributable to fundamentals. Taken to the extreme, an old joke told by Howard Marks illustrates this perfectly:

Two friends meet in the street, and Joe asks Sam what’s new.  “Oh,” he replies, “I just got a case of great sardines.”

Joe: Great, I love sardines.  I’ll take some.  How much are they?

Sam: $10,000 a tin.

Joe: What! How can a tin of sardines cost $10,000?

Sam: These are the greatest sardines in the world.  Each one is a pedigreed purebred, with papers. They were caught by net, not hook; deboned by hand; and packed in the finest extra-virgin olive oil. And the label was painted by a well-known artist. They’re a bargain at $10,000.

Joe: But who would ever eat $10,000 sardines?

Sam: Oh, these aren’t eating sardines. They’re trading sardines.

I come from the school of thought that “everything converges eventually.”6 A company’s popularity does not last forever: either cash flows catch up to valuation or valuation declines. At some point, trading sardines will be priced based on their value as food or as art (i.e. proportional to rarity).

The only way to make money going long popular companies without fundamentals (or the prospect thereof) is to get out at the right time…and market timing is really hard. When a boom market inevitably turns, investors lose money and learn lessons.

The tragedy does not end here. The boom environment that favors companies with the boldest proclamations of riches simultaneously makes it more difficult for the hardworking, emotionally-sober entrepreneurs with strong long-term prospects. They must compete for limited talent while keeping their fiscal sanity despite the siren-song of inefficient growth.

Stepping back further is the abstract, but still real, problem of “opportunity cost”—what we could have spent all that money on instead. Allocating capital to trading sardines provides no “consumer surplus”: the money invested in non-durable companies provides few individuals with a better quality of life and few businesses with greater efficiency. All of that capital instead could have gone to companies providing products and services that people want and need. If you subscribe to the theory that lasting economic growth improves the human condition, then capital misallocation is direct harm to future humans.7

Capital allocation is at the heart of moral philosophy, indeed.

Thankfully the market is, again, whispering sweet songs to companies who have the greatest chance of long-term durable cash flows.

Where do durable cash flows come from?

We might first ask, where do they NOT come from? In this bucket, I put factors that are uncontrollable and market-wide that drive short-term pricing fluctuations but not long-term fundamentals.8 Some examples include:

• Sector biases—there are good companies in “bad” sectors and bad companies in “hot” sectors.9

• Interest rates—when we’re underwriting for a macro environment 10 years in the future, the current rate environment should be a minor factor relative to long-term performance.10

• Overall public market performance—when public markets trend down, this can impact short-term IPO opportunities and soften customer demand; however, if we’re underwriting exits a decade in the future, an improving market is a tailwind for growth rates.

Instead of worrying about the direction of the current, I believe it is more productive to focus on how particular people or companies swim relative to the current. To that end, I believe a durable strategy is to focus on efficiency and long-term happy customers11 because I believe happy customers are the initial source of cash flows. Those cash flows can only be made durable if there is an accumulating unfair advantage.12

This begs the question: where do happy customers come from?13 I believe it’s important to deeply understand customer psychology when underwriting early-stage investments: Who will buy, what will they buy, why, and when?

Who will buy?

First, in order for a startup to make a sale, someone must to go out of their way—either in their capacity as an individual or as an agent of an organization—to purchase something. One of the most important things I learned when first selling B2B is: companies don’t make decisions, people do. In large sales, there may even be multiple stakeholders with different goals who need to be satisfied.

What will they buy?

If a problem is large enough, customers are willing to commit on the promise of a product that solves the problem (common in B2B markets). If, however, a problem is not top of mind or people aren’t aware it can be solved (common in consumer markets, e.g. Ford’s car vs. a faster horse), a company must demonstrate the product before selling it.

As it translates to startup execution strategy, divining which situation a company is in determines whether founders validate a product first with customers or build a product first before attempting to validate. Choosing incorrectly can lead to a false negative (trying to validate with customers who don’t understand the problem) or outright failure (building the wrong product because of insufficient validation).

Why will they buy?

Human psychology and motivation in purchasing decisions is too complex to discuss thoroughly here; however, common patterns include: saving time, saving money, making money, and preventing harm or loss.

Sometimes, particularly with so-called “bottom-up” markets, companies employ the tactic of giving something away for free or inexpensively to a first individual in an organization who is motivated in their individual capacity to make their job easier to save time and then find a different individual who is motivated in an organizational capacity to spend money for features that bring non-individual value (e.g. security, compliance, control) to deploy the product more broadly. In bottom-up go-to-market strategies, I’ve found three requirements to get the initial individual to “buy”:

• Individual value—the product must add direct value to the person who adopts it.

• Permissionless—the individual must be able to start using the product without requiring approval from their manager, IT, or a complex integration process.

• Discoverable—there must be some low-cost way the individuals find the new product.14

In a prior update, I discussed one of my favorite techniques for validating who might buy and why: talking to your prospective customers. Because human psychology is so complex, I’ve found no better way to validate whether customers will adopt a given product. Doing this as early as possible in a startup’s life reduces a tremendous amount of risk: if a startup can’t find those early adopters now, it won’t likely be able to find them after investing millions into building a product.

When will they buy?

Put another way: Will customer psychology change from where it is today?

In some cases, customers want something today but may not want it in the future. We need to be careful about these products, such as those that only sell in boom times or that serve a temporary state of the market (e.g. short-term COVID-induced change). 

In other cases, customers don’t want something today but we have strong belief that they will in the future. Often we can find early adopters who can serve as indicators of where the secular trend might be; however, we have to be careful to believe that the product can “cross the chasm” to the broader market. 

There’s a product design principle I think encapsulates directional change nicely: MAYA—Most Advanced. Yet Acceptable. Raymond Loewy, a father of modern industrial design, posited this principle of gradual advancement in 1951 after observing that people cannot adapt to radical changes quickly and companies must be careful to pay attention to both the product itself and the human psychology of individuals adopting it if they want to succeed in the market.

One of my favorite situations is where a company satisfies an evergreen need for customers with durable funding sources. Helping individuals or organizations save money, make money, save time, and reduce risk are always popular. 

How do companies achieve efficient growth?

For equity investors in startups, it’s not just the cash flows a company generates that matters, but how much equity the company needs to raise to get there: dilution. If a company has a product (what) and customers (who) that are motivated in some way (why) to buy it now (when), we have “Product-Market Fit” (PMF). The question then becomes: can we sell to prospective customers efficiently? I like to break down efficient growth into two stages of readiness: “repeatability” and “scalability.”

The goal of “repeatability” is: can the company build a machine that sells their product rather than relying on the pixie-dust of founder-led sales? Founders running around having individual conversations are unlikely to deliver on long-term goals.15

This is where sales and marketing (collectively, “go-to-market” or GTM) come in. In low-unit-value consumer markets, marketing may be the primary focus, whereas, in high-unit-value B2B markets, sales may be the main focus. At the end of the day, the goalpost of repeatability is whether we can get customers to buy/use our products efficiently.

Efficiency is the key concept above because it’s not much of an achievement to sell $1.00 for $0.80. Unsurprisingly, you’ll have many takers of that bargain, and they’ll all tell their friends! Revenue grows through the roof!

If we spend money on sales and marketing in month one, we should hopefully be able to measure the sales (and associated contribution margin) attributable to that spend in future months. When the cash flow catches up to sales and marketing costs, we now have a “payback period.” As you might imagine, shorter payback periods are better because it means that with a fixed amount of capital, we can turn the crank on this growth machine more quickly and therefore achieve a higher revenue growth rate.

Assuming we achieve repeatability, the next step is to try for “scalability”, where we can invest more and more capital in our GTM engine while generating predictable cash flows.

Here be dragons. Nothing destroys capital faster than prematurely scaling a GTM engine before achieving repeatability. I like to describe this state as a “bonfire of cash” with either consumer marketing or B2B sales teams. The antidote is very clear: make sure your payback on GTM spend is real and fast before putting more money into the machine.

Some key metrics companies look at to determine how their GTM engine is faring include:

• Customer Acquisition Cost (CAC)—how much sales and marketing expense it takes to gain a new customer

• Lifetime Value (LTV)—there are multiple formulations of this, but typically, it’s the total contribution profit (i.e. profit after all variable costs) attributable to a customer after accounting for churn (i.e. people who decide to stop paying for the product).

• CAC to LTV ratio—clearly having a CAC lower than LTV is good.

• Quota to On-target Earnings (OTE) ratio—Quota is how much salespeople are expected to sell while OTE is how much you pay them for selling that amount.

As you might expect, there are ways that companies commonly misinterpret these metrics and get lured into premature scaling (or worse, intentionally obscure poor performance).

Blended CAC: Companies often have multiple sources of customers (“channels”). For example, a consumer product company may produce some excellent content that brings customers “organically” while also purchasing ads on a social network. By blending the CAC for these two channels together, the team may think they have achieved repeatability when, in fact, they have a very profitable organic channel and a very unprofitable paid marketing channel. Paid marketing is generally easier to scale, so if the company tries to move to scalability by investing incremental capital there, they will end up in trouble. Founders can spot this by always evaluating their “per-channel CAC”. Blended CAC is a lie.

CAC:LTV ratios with long lifetimes: When estimating lifetime value, one has to choose how far in the future to “count” future cash flows. Inherently this is making a judgement on customer retention. For an early stage company, this may be unknowable and it’s tempting to be optimistic particularly with the small, early-adopter customer bases that typically accompany companies at the seeking-repeatability stage. I prefer to look at CAC payback period (i.e. how long it takes for contribution margin to pay back CAC) and ensure that it’s short enough for us to use validated retention numbers to make sure a company does not scale prematurely. CAC payback period is a better metric than CAC:LTV ratio.

On the other end of the spectrum, overly-conservative companies may attempt to execute to profitability which can allow others to pass them by. I prefer to break companies down differently: instead of evaluating typical profitability measures (e.g. Net Income, EBITDA, Free Cash Flow), I prefer to think about the choices a company can make in capital allocation. For example, a company can choose to hold money in the bank, invest in sales and marketing, invest in R&D, take debt, or sell more equity.16 If a company has a profitable GTM engine (CAC << LTV) with a fast payback period (payback period << avg. customer lifetime), they’re usually best served in terms of maximizing per-share price by raising more money via debt or equity and deploying that into growth vs. constraining efficient growth in order to achieve bottom-line profitability.17 Of course, investors like to fund these sorts of companies where capital will be allocated to efficient growth.

If a company does have an efficient, repeatable, GTM engine and they choose to invest in getting to scalability, there are usually growing pains they must face, such as learning how to scale a sales team, selling into new customer segments, finding new channels after early ones saturate, making sure their technology can scale, and so on. 

On achieving scalability (i.e. predictable cashflow from incremental GTM investment) we can most likely declare victory in our quest to achieve efficient growth. Founders and early investors will likely see a strong return on their investment of time and money as the market recognizes a solid and growing business. Of course, for the company’s success to be durable, they must nurture unfair advantages that allow them to outcompete future rivals (a topic for another memo).

With respect to the current environment, you may notice that the activities in the quest for product-market-fit and efficient growth are “controllable”—activities whose success is largely dominated by factors that founders can control. For that reason, I believe a profitable path forward is for founders and the investors that back them to sharpen their pencils to make sure they’re pursuing opportunities with hypothetical product market fit (i.e. validated product and customer psychology) while only growing when the company can do so efficiently. 

With discipline in mind, these companies can be very successful, and with the tailwind of an economic recovery, I believe they can be objectively exceptional.

Are there durable cash flows and efficient growth in AI?

Clearly, something interesting is happening in AI; however, as startup founders and investors, how do we know if the timing is right to build great businesses?

First, one lesson many technological visionaries learn the hard way is that “being early is indistinguishable from being wrong.”  Artificial General Intelligence (AGI) has long been a hazy prospect, but trying to position a business or portfolio correctly with respect to how it will occur has been difficult. Once a revolutionary new technology is likely going to exist but is still moderately far off, there are too many possibilities to make accurate predictions: uncertainty dominates. As the revolutionary technology gets closer, it enters “the adjacent possible”, and reasonable predictions become possible.

After a period of moderate progress in AI, I believe that the adjacent possible will change dramatically in the next year and will continue for some time to come. Even if it takes decades to exceed human performance on all tasks, progress is now fast enough in AI research that AGI vs. not-quite-AGI may be a distinction without a difference in many sectors: when circa-human performance gets cheaper even without exceeding human experts, the sector changes dramatically. In contrast to bullishness in crypto, AI immediately impacts many existing industries instead of trying to create a new ecosystem out of whole cloth. Vendors can implement AI capabilities in their products and immediately create value for their customers. For example, Github Copilot, an AI assistant for software engineers, has already transformed how most developers work just months after launch.

As an investor, the question then is: where will durable value accrue? There’s one view that AGI will surpass human skill so quickly that we’ll be unable to process (or survive) it as a species: the “singularity” or “fast takeoff” scenario. In this case, investment returns become irrelevant. Instead, I’d like to focus on the “slow takeoff” scenario that I believe is more likely given the immense computing power required for each advancement in AI performance.

First, what does the AI ecosystem look like? Here’s an illustrative selection of companies across various “layers of the stack"18:

• Hardware—designing and manufacturing the physical hardware used to train and run AI models (e.g. NVIDIA)

• Infrastructure-as-a-service (“IaaS”)— operating AI hardware as a service for others (e.g. AWS, Google, Microsoft)

• Models—training and productizing the “brain” of AI systems ( e.g. OpenAI, Anthropic, Google, Stability, Midjourney)

• Applications—using AI to power workflows in products (e.g. Github, OpenAI, Midjourney, Jasper, Sourcegraph, Recruitbot, Spiral)

To summarize the discussion above, I believe good startup investment opportunities must:

1. Delight customers

2. Grow efficiently

3. Accumulate an unfair advantage (e.g. network effects, technical defensibility, economies of scale, unique access to talent, switching costs)

Additionally, to be a good seed investment, I believe an investment opportunity must be equity efficient (dilution can wash out growth), founded by a force-of-nature team, and on a clear execution path to revenue and subsequent funding milestones.

I believe Hardware, IaaS, and Models to be currently too capital intensive (i.e. they don’t satisfy our “equity efficient” criteria); however, I’ll continually evaluate disruptive technology changes that might allow a startup to compete in these layers of the stack efficiently.

Before discussing the Application layer, I’ll ask the question: will models be a commodity? In the world where one vendor has a monopoly on effective models (generally or in a particular domain), I believe the Model layer will accumulate value; however, in a world where there are multiple substitutable models for any given use case, there will be downward competitive price pressure. In that world, application vendors can choose from a variety of models and will often choose the lowest price. Given early evidence that multiple vendors can produce high-performing models (e.g. Stable Diffusion vs. DALL-E 2 vs. Midjourney, Claude vs. ChatGPT), I believe we’re likely to enter a world where the Model layer commoditizes.

The ability to conjure code, copywriting, or images out of thin air certainly delights customers,  but what makes such applications accrue durable advantages? Thin layers on top of widely available models (absent other switching costs or network effects) are probably not defensible. Novelty may seem like durability early in a platform shift, but value quickly erodes as others copy the first mover. For example, I predict many of the consumer AI photo enhancement products will face this fate.

B2B applications, however, often deeply integrate into workflows and data. As many have observed, once an organization adopts an application with deep integration as a “system of record” the application becomes difficult to remove, or “forever-ware”. If fast-moving startups can offer compelling value via AI (e.g. reducing human labor costs or opening up new opportunities by lowering the cost of decision-making) where legacy vendors are unable to match, those startups may be able to provide so much value to their customers that they overcome switching costs and disrupt entire mature software markets. These “platform shift” opportunities occur very infrequently, and I’m actively evaluating opportunities that can take advantage of this rare moment. Furthermore, we have a number of well-positioned portfolio companies that have the opportunity to use AI to accelerate their growth in large existing sectors (e.g. healthcare, robotics, content creation, cybersecurity, and IT administration).

Despite the opportunities AI brings, we need to temper our expectations with a critical view of the adjacent possible. Some particular market timing risks I’m mindful of:

• Offline workflows—Where workflows are already digital or can be made digital, AI can rapidly improve efficiency; however, if workflows are primarily in-person and offline, it may take years to bring interactions online before they are “in scope” for AI.

• Reliability—Generative AI models are currently generating a lot of excitement; however, they’re not yet at human-level reliability. For example, Copilot produces working code, but the output often contains bugs and security vulnerabilities. Similarly, ChatGPT, when asked for references supporting a claim, can hallucinate fake research. At the moment, humans are still needed to make generative AI output “reliable”, limiting where such models are usable in mission-critical situations. 

• Labor switching time—AI generates value by replacing human labor at a lower cost or enabling workflows that would be cost-prohibitive with human labor. When AI replaces a job, it frees humans to perform higher-value work while improving overall standards of living by reducing the cost of everything; however, even if overall prosperity improves, the segment of people who are directly impacted may understandably resist AI if their job is wholly eliminated (e.g. self-driving vehicles) rather than improved (e.g. AI-assisted software engineering).

This last issue—the impact of AI on labor—is rightly a significant concern for individuals, employers, and society as a whole. In the past, markets turned over slowly enough that employable skills could last for one’s entire career. As technology cycles march on, we’ll need to embrace more and more frequent “re-skilling”—helping people develop new skills throughout their careers instead of providing education only at the beginning of careers.  To that end, I’m optimistic that the cause of disruption to labor markets may also be the solution: AI-assisted tutors have the opportunity to lower the cost and increase the quality of education.19 Already, large language models (LLMs) can explain complex concepts, generate lesson plans, and even pass tests for elite professions (e.g. a Wharton MBA exam, the Bar for lawyers, and the US Medical Licensing Exam).

Finally, to return to the original question of whether we can find durable cash flows and efficient growth in AI, I believe we’re in a perfect storm: a major platform shift (particularly in B2B applications), a reasonable valuation environment, and less competition for talent all occurring simultaneously. Of course, even in attractive situations such as the present, founders and investors must critically evaluate execution risk, defensibility, and market timing if we want to build durably great businesses.

In short: I’m cautiously optimistic.

It may sound strange to have a disposition anywhere near optimism during a significant market drawdown, but let me explain why.

We need cycles of startup birth and death to efficiently allocate human effort. If companies have the durability of government programs—persisting regardless of outcome—they will probably approach the same level of effectiveness. When capital is abundant, even to companies whose prospects of achieving long-term viability are slim, great people spend time on efforts that are ultimately likely to be fruitless. As expected, the past few years have been extremely difficult even for strong startups to hire.

Just as our bodies seem to work best when alternating between anabolic (abundance) and catabolic (scarcity) states1, I believe the startup ecosystem to work best when there are occasional catabolic “cleanup” processes that free great people to focus on the most promising opportunities. Without cleanup processes, inertia tends to prevail. We’re in the middle of one of those cleanups right now and, though it is painful, I believe it will make for superior long-term results vs. the creeping inefficiency that comes from staying in an abundance state forever. Anecdotally, it’s already getting much easier for strong startups to hire.

The market has turned down the volume of speculation’s siren song. Like all professions, venture capital operates in an environment that reinforces specific behaviors. As you probably suspect, incentives in venture capital do not always align with long-term investment success.

In my opinion, one common hazard is to celebrate fast markups that accompany a portfolio company raising subsequent rounds of funding. In some cases a swiftly growing valuation happens because a team executes swimmingly, while in other cases the company is merely in a “hot” sector, which often happens in the wake of a high flying leader.2 Fast markups are not intrinsically bad, but when investors seek to allocate to the companies most likely to get a fast markup alongside many others doing the same, a cohort of companies gains the emergent property of value being more dominated by capital flows than by fundamentals. 

Furthermore, the praise and promotions heaped upon investors following fast markups make for particularly strong psychological reinforcement.3 Once an investor has their first glimpse at the technicolor of “speculation”, the black and white of “investing” seems positively boring. Even though thoughtful fund managers recognize the issue with celebrating markups over progress in fundamentals, the practice is difficult to avoid given how few other metrics there are in judging investment skill in the near term.

The best answer I’ve seen to avoid this hazard is to maintain a culture of disciplined underwriting at each investment event that evaluates long-term outcomes under sane valuation metrics.4

I believe making money is safer (and more satisfying) than getting money. Some may have successfully traded out of speculative assets at the top5; however, most of us (myself included) are not smart enough to time the market. It’s far more reliable to invest in companies that have reasonable shots at improving their revenue, margins, and cash flow with hard work that’s largely in their control.

Speculation and investing look similar on the way up, but not the way down. The same attribute (appreciation via capital flows) that makes speculative assets so attractive on the way up becomes their demise on the way down. Such an asset that stops going up and to the right is “broken” and attracts little additional capital so that its clearing price declines precipitously as the asset is suddenly valued on fundamentals. Uh oh.

As Buffett famously said, “only when the tide goes out do you discover who has been swimming naked.” Now that later stage markets have begun to focus on the fundamentals of cash flow, margins, and the like, that same return to discipline trickled down initially into the growth stage venture and, now, into early stage venture markets. Reason is returning!

Many of the “hot” companies with eye-popping valuations are being re-evaluated under long-term-normal valuation criteria with a variety of results:

• Companies whose expenses matched their valuation (that is to say, extravagant) with little in the way of revenue and margin are in the process of imploding or being re-valued much lower.

• On the other hand, many wise founders took advantage of low cost of capital to pad their balance sheet while growing revenue, controlling expenses, and avoiding “toothy” deal terms. I suspect many of these companies will grow into and exceed their top-of-market valuations assigned in the last couple of years, with their shareholders being richly rewarded.

At the early stage, not dying is a prerequisite for success. For many companies in the middle of their growth trajectory—a modestly sized balance sheet with good prospective fundamentals but not yet having achieved repeatability and scalability—founders are finding themselves having to switch from a growth-focused strategy to a cash-preserving strategy while at the same time seeing the goalposts for their next round of funding move further away. The founders in this position who are likely to survive have been quickly decreasing expenses and raising additional capital6 so as to be able to weather the storm.

Planning for capital market storms is fraught with uncertainty—there is no way to tell ahead of time whether it will be a moderate squall or a once-in-a-century hurricane.  Facing the unpredictability of fundraising in the current environment, I like having a set of contingency plans that looks something like this:

• Plan A: raise a new round from a strong external lead immediately.

• Plan B: extend the previous round however possible to have the capital to hit long-term-normal7 metrics for the next funding round with significant margin of safety.

• Plan C: if unable to rase additional capital, execute to a breakeven, “default alive,”8 state.

While few companies need to execute a “Plan C” strategy, ensuring that it’s an option effectively guards against a terminal case of spendtoomuchinosis in an austere funding environment. 

It’s going to be a great time to build. After the dust settles, I believe we’ll be left with an extremely attractive environment for dedicated founders to execute. Here’s why:

• From death comes life—As mentioned above, as some companies fold or downsize, great people are freed up to work on new opportunities, making it easier to build world-class teams.

• Customers accelerate change in crisis—In times of plenty, both individuals and organizations can be slow to execute on necessary, but disruptive, changes. When large shocks occur, bold initiatives—e.g. automation, switching away from legacy vendors, moving processes from offline to online—become easier to stomach and even critical for survival. In these dynamic environments, nimble startups have the opportunity out-execute incumbents.

• Those who survive the downturn will be at scale with little competition in the recovery—by the time prosperity returns, the select founders who endured the pain to build repeatability, scalability, and strong fundamentals will be ready for rapid growth with few others at scale to challenge them.

At the same time, I believe it is also a great time for investors to back the force-of-nature founders who will weather the storm:

• Founders with fortitude will remain; tourists will go home—Given how much more difficult it is to raise capital, only founders who really care about what they’re doing will go through the gauntlet. Those who start companies lusting after quick riches will pursue other careers.

• Valuations are approaching sobriety—Even though early stage valuations have come down from their peak, some are still “expensive.” I believe it to be essential to avoid psychological anchoring9—comparing current valuations only to recent valuations—and make sure to underwrite against long-term-normal valuation metrics.

• Entry prices reflect today’s macro environment while exit prices reflect the future’s macro environment—While fundamentals of a given company are, in my opinion, the most important aspect of underwriting, it’s always nice when we get the tailwind of multiple expansion accompanying macro recoveries.

• Startup equity seems like an ideal bet in an inflationary environment—As I touched on in the last update, startups can typically raise prices in step with inflation rather easily while maintaining or improving contribution margin so long as they’re solving a problem that is important to their customers.

• Startup best practices are better than they’ve ever been—when we first started Lookout, it was very difficult to track down advice on how to build a fast growing technology business. Today, however, founders have so many best practices available via accelerators like Y Combinator, blogs, twitter, and early investors that I believe startup failure rates will be on a secular decline for some time to come.

• Technology is no longer a discrete industry; all industries are technology industries—while it’s tempting to think of “red ocean” sectors (e.g. crypto, consumer social, dev tools) as indicators of an oversaturated technology market with little white space, much of the economy still runs on legacy tools and processes that have not progressed in decades. I believe there will be immense opportunity to make the world more efficient, particularly in these sectors.

Of course, while all of the trends may line up for a great building and investing environment, I always try to keep in mind the words of Mike Moritz: “I invest in birds, not flocks.” 

It’s hard to be a venture index fund spreading capital evenly across the industry. Instead, I believe it’s a much more sound proposition to back founders who have the capability to build the industry defining companies of the next market cycle. These are the founders and businesses for which I’m cautiously optimistic.

Of course, this is all easier said than done—rare is the investor who will publicly declare they only invest in hot markets and untenably high valuations.

How then, if venture investing focuses on early-stage companies without trailing financials to speak of, can one underwrite wise investment decisions? The most helpful model I’ve found is to look at each investment as a probability distribution across outcomes that changes as we acquire more evidence.1 To start, we might believe an investment in a promising startup has a:

• 50% chance of returning < 0.9x

• 25% chance of returning between 0.9x and 2x

• 10% chance of returning between 2x and 10x

• 15% chance of returning greater than 10x.

If, for example, we gain evidence that the startup has figured out how to make their sales process repeatable (e.g. their first two account executives achieved their quota last quarter with an 80% win rate, consistent collateral, etc) we will adjust our probabilities “to the right.” Hooray!

Though this is only one data point and there are many things that could kill a promising young company. How, then, can we systematically gather evidence to “update” our probability distributions during an investment process? I like it when founders and investors work collaboratively to accumulate evidence for and against risk in each major category (e.g. product, pricing, team, go-to-market, competitive, technical, market) so that everyone has a good sense for what is likely to kill the company and what is unlikely to do so. It’s said that “more startups die from suicide than homicide”2 and I believe it makes more sense to focus on the company and its relationship to prospective customers (less so on its relationship to current competitors).

Sometimes it’s obvious when one should invest. Other times, there’s just too much risk. Unlike other asset classes: startups are not merely museum-pieces to be studied, so we shouldn’t stop there. 

Investors get to roll up their sleeves and work with founders to see if they can turn over a few more cards before investing. For example, something as trivial as carefully asking 5-10 prospective customers “Will you pay $126,237 for this?”3 and listening carefully to the ensuing discussion can meaningfully de-risk a company (i.e. shift our probabilities right) without changing the entry valuation.

Another thought experiment I’ve found helpful is to ask What happens when the market conditions are much poorer than today? 

For some companies (particularly those with high prospective margins and revenue), we’ll shift our probabilities “to the left” and say things like “multiple compression” or “higher cost of capital.” While undesirable, it’s not terminal. 

For other companies (particularly where the word “revenue” is accompanied by waving of hands and “cash flow” cannot be said with a straight face by those with above average honesty), we put a whole bunch of probability mass at zero because in a less-optimistic environment, they’re unlikely to either generate revenue or raise capital. Uh oh.

While there are a number of ways to build a venture portfolio, the one I believe leads to sleeping well at night focuses on companies that build solid underlying “all-weather” businesses, not only aiming for the next round of capital (which may or may not be there)4. Of course, life is especially wonderful when portfolio companies ALSO benefit from exceptional market conditions.

Another important aspect of early-stage underwriting is future dilution given most companies will eventually raise follow-on capital. These days, growth-stage companies are raising massive rounds at eye-popping valuations. This means that early stage investors are benefitting from low long-term cost of capital. In practice, this means that disciplined founders (as well as the early-stage investors who back them) should have far less dilution over time relative to historic norms (i.e. another factor shifting our probabilities right). Given the billions more being raised into growth funds, I don’t see this changing any time soon. 

Finally, the topic of inflation is on many a founder and investor’s mind: is this good or bad for early stage venture as an asset class? On one hand, labor is typically the single greatest segment of a startup’s expenses and salaries are increasing far faster than CPI in most relevant roles.5 On the other hand, most startups aren’t very “capex-ey” and have relatively short-term pricing commitments with the ability to reprice in contemporary dollars on a regular basis. 

In 1981, Warren Buffet observed that businesses who successfully withstand high-inflation environments “must have two characteristics: 

1. an ability to increase prices rather easily (even when product demand is flat and capacity is not fully utilized) without fear of significant loss of either market share or unit volume, and 

2. an ability to accommodate large dollar volume increases in business (often produced more by inflation than by real growth) with only minor additional investment of capital.” 

Buffett seems to aptly describe many technology startups as inflation-resistant bets, especially those that are so tightly-integrated into their customers lives/workflows they have become “forever-ware.” Further, some founders are beginning to add escalator clauses in long-term contracts that peg automatic price increases to CPI (or another relevant metric), thereby directly hedging inflation risk. 

Overall, the nice thing about thinking about the world in probability distributions (“shapes”) instead of artificially precise numbers is that one is not pretending to know exactly what the future holds.6  As I look at the evidence in front of us, investing in high-quality businesses at their earliest stages of growth with a disciplined entry price seems like a pretty good place to be right now even with the specter of high inflation and increased interest rates looming.