The Number That Should Keep Founders Up at Night
OpenAI is valued at $157 billion. Anthropic recently closed at $61 billion. xAI hit $50 billion in under two years. These are not typos.
If you’re a founder or CEO building a company in 2026, you’ve probably had this conversation: “Should we build our own AI capabilities, license from one of these giants, or find a partner who’s already figured it out?” It sounds like a technology decision. It isn’t. It’s a capital allocation decision — and most founders are getting it wrong because they’re looking at the wrong signals.
This is a grounded memo for how to think through it.
Why Valuations Are a Misleading Input
Model vendor valuations are driven by narrative, not infrastructure returns. The market is pricing in a winner-take-most scenario where one or two foundation model providers become the operating system layer of the global economy. That may happen. It also may not.
What you can know right now: the returns from using AI infrastructure are compounding faster than the returns from owning AI infrastructure for most companies outside the top tier of hyperscalers and research labs.
Think about it this way. When AWS launched, most smart founders didn’t try to build their own data centers to compete. They used AWS and redirected that capital into product differentiation. The founders who tried to build infrastructure-first ended up with expensive distraction and slower iteration cycles.
The same logic applies today — with one important caveat we’ll get to.
The Build Case: When It Actually Makes Sense
Building proprietary AI capabilities — fine-tuned models, custom inference pipelines, internal training infrastructure — makes sense in a narrow set of conditions:
Your data is the moat. If you’re sitting on a dataset that no model vendor can replicate (patient records, proprietary transaction data, specialized domain annotations at scale), building gives you a defensible edge that licensing cannot. The value isn’t the model — it’s the feedback loop between your data and your model getting smarter.
Your use case is latency-sensitive and high-volume. If you’re making millions of decisions per day at sub-50ms latency requirements, the economics of API calls break down fast. On-premise or self-hosted infrastructure starts making financial sense at significant scale.
Regulatory constraints require it. Financial services, defense, and healthcare companies often can’t send sensitive data to a third-party API. For them, build isn’t a preference — it’s compliance.
Outside these three conditions, building is usually ego, not strategy.
The Buy Case: Why Most Founders Should Be Here
Licensing access to frontier models through APIs is the right default for the majority of venture-backed companies. Here’s why:
The capability curve is moving faster than your build cycle. GPT-4 to GPT-4o to o3 happened in less than 18 months. If you started building a model to match GPT-4 when it launched, you’d finish building something that’s already been lapped. The model vendors are running a race you can’t win by joining.
Your investors funded a product, not an AI lab. Every dollar you spend training models is a dollar not spent on distribution, customer success, or the product features that actually drive retention. AI capability is increasingly a commodity input — your differentiation is how you apply it.
Time to market is your real competition. Your competitor isn’t building their own model either. They’re shipping products. If you’re in a build cycle for infrastructure while they’re iterating on customer feedback, you’re losing even if your model is theoretically better.
The buy case doesn’t mean passive adoption. It means strategic integration — choosing the right models for the right tasks, building proprietary workflows on top of reliable infrastructure, and treating AI capability as a lever, not a project.
The Partner Case: The Underrated Option
Partnering — whether with a systems integrator, an AI platform company, or a specialized vendor in your vertical — gets dismissed too quickly by founders who conflate it with outsourcing or lack of ambition.
Done well, the partner path looks like this: you identify a company that has already solved the infrastructure problem in your domain, you build a commercial relationship that gives you capability access plus integration support, and you focus your internal resources on the customer-facing layer that only you can build.
The honest risk: partner dependency. If your partner gets acquired, pivots, or raises prices, your roadmap is hostage to their decisions. The mitigation is contract structure (favorable terms, portability clauses, source code escrow where relevant) and building an abstraction layer in your own codebase so you’re not hardwired to a single vendor’s API surface.
The Framework: Three Questions Before You Decide
Before your next board conversation about AI strategy, answer these three questions honestly:
1. What’s our actual data advantage?
Not theoretical data you could collect — data you have today, at volume, that a foundation model vendor cannot access. If the answer is “not much,” the build case weakens significantly.
2. What does our customer value that AI enables — and who needs to own that layer?
If the answer is insight delivery, workflow automation, or decision support, you probably need to own the product layer, not the model layer.
3. What’s our real cost of delay?
Every quarter spent on infrastructure is a quarter not spent on customers. If your market is moving fast, speed compounds — and the founder who shipped six months earlier has a retention and referral advantage that’s hard to overcome.
The Honest Bottom Line
The AI vendors are expensive because the market believes they’re building the next cloud. They might be right. But that doesn’t mean you need to own a piece of the cloud to win in your market.
Most founders should buy access, build product, and partner selectively — then revisit the build decision when they have the data asset, the volume, or the regulatory constraint that actually justifies it.
The founders who will look smart in five years aren’t the ones who bet on model valuations. They’re the ones who used the available infrastructure to get to customers faster, learned more, and built something people couldn’t live without.
That’s still the game. The tools just got better.
