After writing The Human Bet (opens in new tab), the question I kept returning to was: what has to get built for any of that to be real?
The demand side is about people: what they'll trust, what earns a lasting place in their lives, how intelligence expands human agency rather than just automates tasks. I believe that argument. But believing the demand side doesn't tell you where to invest on the supply side. For that you have to look at the other half of the equation separately, on its own terms.
We have been working through what we think has to exist, not as a single thesis, but as a collection of bets. The tooling that lets modern businesses operate at a new level. The infrastructure that keeps people in control as agents act on their behalf. And the specialist models that can push intelligence further in domains where focus or access to proprietary data creates an advantage.
These aren't extensions of The Human Bet's argument. They're what we believe has to be true on the other side of the demand, and where we see an enormous new landscape for company creation taking shape.
The Capability Layer
Application companies won't emerge alone. Every major platform shift creates an entirely new generation of tools built to serve them. We've seen this before: as modern consumer brands emerged, companies like Stripe, Shopify, Faire, and Klaviyo were built alongside them. They weren't the story themselves. They existed because a new generation of businesses suddenly needed capabilities that hadn't existed before.
Intelligence is creating the same dynamic. AI-native companies make decisions, build products, and serve customers in fundamentally different ways, increasingly running with far smaller teams than the businesses they're usurping. That creates demand for an entirely new software stack designed around how these companies work. We think this will become one of the most important investment categories of the next decade.
“The opportunity here isn't automation for its own sake, but giving smaller teams the capabilities of much larger organizations.”
The opportunity here isn't automation for its own sake, but giving smaller teams the capabilities of much larger organizations: better judgment, faster execution, richer context, entirely new workflows. The founders who win will have alpha from understanding their industries well enough to know where the real friction lives, then compound it by turning the intelligence generated through the work into better decisions and execution.
That's why this category feels compelling to us. For nearly two decades, we've invested in brands, retailers, healthcare companies, financial services businesses, marketplaces, and consumer platforms as their workflows evolved, watching their constraints accumulate. AI finally makes many of them solvable. We're seeing that opportunity show up in very different places already. Koah (opens in new tab) is building monetization infrastructure designed for conversational interfaces. Blue Labs (opens in new tab) is giving AI systems a deeper understanding of human context. Marklo (opens in new tab) is using intelligence to rethink how commerce brands plan and operate across channels. Depthfirst (opens in new tab) is applying AI to cybersecurity as the development stack itself changes. What connects them isn't a common product category. It's a founder who knows a problem deeply enough to recognize that intelligence changes how it can be solved. We think there are a lot more of those problems still hiding in plain sight.
The Control Layer
If AI is going to act for us, what makes those actions trustworthy?
The internet was built around a simple assumption: people initiate actions. People log in, approve payments, sign contracts, and make decisions, and identity, security, and payments were all designed for a world where a human sat at the center of every transaction. That assumption is changing. Agents are beginning to research, negotiate, purchase, schedule, and coordinate work with increasing autonomy, and as that happens, the question stops being whether an AI can do something and starts being whether it should be allowed to. We see this as one of the defining infrastructure questions of the AI era.
A new control layer needs to emerge, one that gives people confidence their AI is acting within the authority they intended: identity systems that carry intent, permission frameworks that determine what an agent can and cannot do, audit trails that make autonomous actions transparent, and payment infrastructure built for agents transacting with agents.
Making AI more capable is not really the point here. Making autonomous systems trustworthy is. Much of today's infrastructure was built for a world where humans performed every important action themselves. The next generation needs to work in a world where software acts constantly while people remain in control. We see that tension emerging across our own portfolio….
"where the hardest questions about increasingly capable agents are no longer about intelligence. They're about authority."
We're already investing against that problem. Natural (opens in new tab) is rebuilding payment infrastructure so agents can transact on behalf of businesses and users. ZeroClick (opens in new tab) is working on what happens when the buyer and seller are both agents, including the pricing, identity, and controls required to make those transactions work. These are two places where the problem is already visible to us. There will be many others as agents move from answering questions to actually doing things.
The Model Layer
The Human Bet argues that enduring value lives in the relationship an application builds with its users, not in the underlying model, and that remains true as general-purpose models continue to converge. But there is an important exception, and potentially a significant category of opportunity: specialist models built for a particular domain, paired with applications designed around that expertise.
The case for specialist models isn't that general-purpose models will somehow fall short. It's that some problems reward a level of focus they are unlikely to prioritize. Concentrating training, evaluation, and product development around a specific domain can produce meaningfully better performance, and sometimes enable a product experience that would not be possible with a more general model. In those cases, the model is not simply infrastructure underneath the application. It is part of why the product is better.
Sometimes that starts with proprietary data. A specialist may have access to information a general-purpose model simply cannot reach because it is regulated, relationship-driven, generated inside a closed ecosystem, or otherwise unavailable. That can create a durable advantage because the constraint is access, not effort.
In other cases, the advantage comes from focus itself. A company willing to concentrate all of its resources on one domain can often push performance further and faster than a generalist for whom that domain is one priority among many. That lead may not last forever. If a horizontal model decides the category matters, its scale and resources may eventually close the capability gap.
But that does not make the early lead unimportant. It gives the specialist time to build around it. Better performance can produce a better product. That product can generate more usage, more domain-specific data, and a much deeper understanding of what customers actually need. Over time, the company may become harder to catch for reasons that extend well beyond the original model advantage.
So the investment question isn't simply whether a specialist model is better today. It's why it is better, what that difference makes possible at the product layer, and what the company is able to build before the generalists catch up.
"The most interesting opportunities are likely to be the ones where specialized intelligence is the starting point, not the whole moat."
We can already see both the performance advantage and what can be built around it. Speechify (opens in new tab) has gone deep on speech, pairing its own models with a consumer product and an API that makes those capabilities available to others. Suno (opens in new tab) has taken a similar approach to music, where purpose-built models make possible a consumer experience that general-purpose models haven't matched. In both cases, specialized intelligence is the starting point. What gets built from that lead, and how much better these models can become through the products around them, is what makes this layer particularly interesting to us.
There are likely many domains where general-purpose models will ultimately be good enough. We're more interested in the ones where "good enough" leaves something important on the table.
The supply side of AI is not one opportunity. The Capability Layer, the Control Layer, and the Model Layer each have different reasons to exist and different ways value can accrue. That's why we think they are more useful when considered separately than collapsed into a single infrastructure thesis.
Our early investments here are evidence of what's starting, certainly not a map of where this ends. If anything, what we're seeing points to a much larger opportunity still taking shape.
