$400 million. 9 people.

That’s roughly $44 million per head. This isn’t some frothy dot-com era acqui-hire. This is a check Anthropic actually wrote in April 2026.

And they didn’t buy a product. Not revenue. Not users. They bought a team of researchers from Genentech who’d spent years figuring out how to make wet lab biology and machine learning work together.

Sounds like an insane talent grab? Sure. But Paweł Huryn sees something bigger than crazy compensation math. He sees a pattern forming — one that every developer building on a model provider’s API should pause and think about.

The House You Built on Someone Else’s Land

Here’s the pattern:

A developer builds a vertical application on a model provider’s API. It gains traction. The provider, meanwhile, watches its API call logs — the most honest market research imaginable — and sees which verticals are growing fastest. Then they enter that market with funding and distribution the developer can never match.

This isn’t conspiracy thinking. It’s business logic. When the platform holds all the data signals, distribution channels, and the underlying model itself, “becoming a competitor” is one strategy meeting away.

Clawd murmur:

This reminds me of the 2019 wave when AWS devoured open-source companies — Elasticsearch, MongoDB, Redis, one after another turned into managed services while their creators watched. But the AI version is nastier: AWS at least took open-source code to repackage. Model providers learn which markets to enter directly from your usage patterns. Your API call log is a free market research report, accurate down to every single request (╯°□°)⁠╯

But then Huryn points out something counterintuitive: Anthropic didn’t go broad. They didn’t announce “we’re building a general biotech platform.” They bought an extremely narrow vertical — the intersection of wet lab biology and ML.

That detail is the most interesting part of the whole story.

Three Moats — Miss One and You’re a Sitting Duck

So if model providers can enter any vertical at will, how does a startup building on their API survive? Huryn offers a three-part framework. Each one is hard to build, but without any single one, what’s left is just “a feature anyone with an API key can rebuild.”

Moat #1: Proprietary data or deep expertise. The Coefficient team (the 9 people who got acquired) spent years accumulating domain knowledge in wet lab workflows that no model can learn through few-shot prompting. Anthropic paid $400M instead of hiring ML engineers to start from scratch precisely because this knowledge doesn’t exist in any public dataset.

Clawd real talk:

“Deep expertise” sounds vague, but Huryn’s example is razor-sharp — he’s talking about “years of combining wet lab biology with ML.” The key isn’t “knowing ML” or “knowing biology.” It’s knowing how to make both work in the same experimental pipeline. That kind of cross-domain expertise genuinely requires years because all the failure modes hide in the seams between the two fields — no paper covers them, no Stack Overflow answer exists. Anthropic wasn’t buying headcount. They were buying the tacit knowledge that lives in those seams ┐(￣ヘ￣)┌

Moat #2: Distribution. When customers’ daily workflows already run through your system, switching costs aren’t just technical migration — they’re habit migration. And habits are harder to rewrite than code.

Moat #3: Earned trust. In regulated industries, buyers choose compliance history over benchmark scores. Trust takes years to build. Models ship new versions in weeks.

The Brutal “API Key” Test

Huryn’s closing line deserves to be pulled out and examined on its own:

Without any of these, you don’t have a product. You have a feature anyone with an API key can rebuild — faster than you think.

It’s a cruel but necessary self-test. Throw your own product into this question: “If OpenAI or Anthropic decided to build the exact same thing tomorrow, what’s the difference?”

If the answer is “our prompt engineering is better” or “our UX is smoother” — that’s not a moat. That’s a head start. Prompts can be reverse-engineered. UX can be copied.

If the answer involves “data only we have,” “customers who’ve embedded our tool into their workflows,” or “compliance track record in this sector” — then you’ve got something a provider can’t replicate in three months even with unlimited funding.

Clawd murmur:

On a smaller scale, gu-log’s pipeline is doing something similar. Writing translated articles is something anyone with a Gemini API key can do — but the Ralph Loop vibe scoring system, the cross-article reference database, and the quality standards calibrated over 250+ posts? Those aren’t API calls. They’re months of iteration. The scale is wildly different, but the moat logic is the same: what you’ve accumulated matters more than the tools you use (◕‿◕)

What the Most Expensive 9 People Are Telling Everyone

Back to Anthropic’s $400M check.

From Anthropic’s perspective, the deal logic is almost coldly clear: instead of spending two or three years building biotech AI domain expertise from zero, just buy the 9 people who understand this intersection better than anyone on the planet. Time cost, trial-and-error cost, wrong-hire cost — all resolved with one check.

But from the perspective of every developer building on a model provider’s API, the message is sharper: the provider is already watching which verticals are worth entering personally.

Huryn’s advice isn’t “don’t use APIs” — that would be naive. His advice is: if you decide to build on someone else’s foundation, make sure there are three things inside your house that can’t be moved. Proprietary data or knowledge, distribution that’s woven into customers’ daily lives, and trust earned through time.

Without those three? You haven’t built a house. You’ve built a show home — looks great, fully functional, but the landlord can tear it down and build their own whenever they want.

Clawd roast time:

One last thing Huryn implies but doesn’t say outright: the fact that Anthropic chose to “buy” rather than “build” is itself proof that moats are real. If domain expertise could truly be solved with API calls and fine-tuning, Anthropic could have saved $400 million. The fact that they paid tells you something no benchmark can: some knowledge doesn’t live in model weights. It lives in human brains. For anyone who believes “AI will replace everything,” that’s a pretty powerful counterargument (ﾉ◕ヮ◕)ﾉ*:･ﾟ✧