When the Model Maker Enters the Room

The press release landed on a Tuesday morning. I read it the way I read most press releases about things that will take a while to understand — twice, quickly, on my phone, in the parking lot of the coffee shop on El Cajon Boulevard where I tend to end up when I’m thinking about something I don’t yet have words for. A model company, it said. A new enterprise AI services company. Three of the most consequential investment firms in private capital as co-founders. The kind of announcement where the words are plain and the implications take a minute to land.

A founder I’d been working with sent me a message that afternoon. Short, which is how she always texts when she actually wants an answer: Had you seen this. Were you worried.

No, I said. Not in the way you mean.

What it actually says

Most of the coverage I read that week described the announcement as Anthropic “raising money” or “partnering with enterprise.” That is not what it says. What it says is that the company that makes one of the two or three frontier models in current production use has created a joint venture, co-capitalized with institutional investors, to sell services. Not to sell API access. Not to sell a platform license. To sit down with enterprise clients, understand their problems, and deliver outcomes.

That is a different thing. I want to be precise about why — and to get there, it helps to read this announcement alongside a second one that landed the following day from the same company.

The day after the enterprise services announcement, Anthropic disclosed that it had signed for the entire compute capacity at SpaceX’s Colossus 1 data center: 300 megawatts, 220,000 GPUs. The announcement also noted, in a subordinate clause, an interest in developing orbital AI compute capacity. I mention this not to editorialize but to be precise about the scale of what is being assembled. A company that signs for entire data centers and floats the idea of orbital compute is not making a side-bet on a services business. It is building a vertical stack from infrastructure to engagement. The buyer on the receiving end of that sale should understand what structure they are standing inside.

When a model company sells you API access, the relationship is clean. You write the integration; you own the outcomes. If the system doesn’t work, the question is whether you built it correctly, not whether the company that makes the weights is accountable for the result. The API is a raw material. What you build with it is yours.

When the model maker also sells the services engagement — when they are both the supplier of the ingredient and the chef who cooks it — the accountability question gets more complicated. Not worse, necessarily. But more complicated. And that complexity is worth understanding before you sign a contract with either side of it.

What institutional capital means for the incentive structure

Every professional services business has an obligation structure. The structure shapes behavior — not because anyone is malicious, but because people optimize for what they’re measured on, and firms optimize for what their investors need.

A boutique practice with six engagements a year is measured on whether those six engagements produced outcomes that would make the operator who commissioned them want to come back, or tell a colleague about them. The entire business is referral-density, which means the incentive structure is perfectly aligned with the client’s interest in a working system and a candid answer.

A joint venture backed by three firms whose combined AUM exceeds the GDP of countries I could locate on a map is measured on something different. I don’t know exactly what — the terms aren’t public — but I know what institutional capital is for. It is for producing returns at scale. That is not a criticism. That is a description. Scale changes the incentive structure of a services business whether or not anyone intends it to.

The question this creates for a buyer is not is the large firm worse. The question is what am I optimizing for when I choose who sits across the table from me for the next eighteen months?

The question to ask before you sign

Here is the one question I would ask every AI services vendor in 2026, whether they are a boutique practice or a Goldman-backed joint venture:

If the system we build together fails to hit the evaluation metrics we define at the start of this engagement — not “fails to be perfect,” but fails to clear the bar we both agree on in week two — what happens to your fee?

The answer to that question tells you everything about incentive structure that a capabilities slide does not. Most large services firms will give you a polished answer that does not answer the question. Most boutique practices will give you a less polished answer that does. You are looking for the version where the person across the table is willing to be wrong, on the record, against a number, in a way that has a consequence.

That answer requires the person who built the system to also be the person who bet on whether it would work. It requires skin in the game that is not denominated in reputation — which is what every senior executive claims to have — but in deliverables, with a definition of done that was agreed to in writing before anyone opened a notebook.

This is not a critique of large firms or of institutional capital. Some of them will do excellent work. Some boutique practices will do terrible work. The size of the balance sheet is not the variable. The incentive structure is. And incentive structures are readable if you ask the right question and pay attention to whether the answer is specific or smooth.

The press release was neither a threat nor an invitation. It was a market signal. The services tier of AI is consolidating, the same way every professional services market consolidates when the underlying technology matures: the large firms get larger and more institutional, the boutique practices get more specialized and more personal, and the buyers who know the difference get better outcomes than the buyers who don’t.

I drove home the long way that Tuesday, past Balboa Park and back down Sixth, thinking about which of those two things I was building and whether the answer was still the right one. It is. But it’s worth asking the question.