Why the Model Labs Became IT Services

Anthropic announced a joint venture Monday with Blackstone, Hellman and Friedman, and Goldman Sachs to deploy Claude inside mid-sized enterprises. OpenAI announced a parallel vehicle hours later, The Deployment Company, ten billion dollar valuation, with TPG, Brookfield, Bain Capital, SoftBank, and 15 other investors. Both will embed engineers inside customer teams. Both will target the portfolio companies of their private equity backers. Both copied the model Palantir built fifteen years ago. (TechCrunch)

A market analyst put it plainly. These are consultancies. (SiliconANGLE)

The model labs just walked into the services business. They did not partner with Accenture or Infosys or TCS. They built their own. And the structure they built, embedded engineers selling into mid-market enterprises with private equity providing distribution, is exactly the structure the existing IT services industry has spent thirty years optimizing.

Anthropic is closing a round at nine hundred billion. OpenAI sits at eight hundred and fifty-two. Neither needs another revenue line. So what are they buying?

The counterparty position.

The IT services industry has a property nobody describes correctly. The product is not software. The product is someone to blame when the software breaks. Code is what gets delivered. Accountability is what gets paid for. You can see this in the pricing. Median EBITDA margins have held at thirteen to fifteen percent for a decade. Public multiples sit around ten to thirteen times EBITDA. Private deals cluster in the same band. The numbers barely move. This is what a regulated utility looks like. And it is regulated by the procurement department.

A CIO can defend “three senior engineers from Accenture for six months” to her board. She cannot defend “we paid someone for their judgment.” Headcount is auditable. Judgment is not. So the unit of sale becomes the person-hour, labor becomes seventy to eighty percent of revenue, margins compress to whatever bench utilization allows. The buyer cannot evaluate the output, so she evaluates the input.

This worked for thirty years because execution was scarce. Writing software took time. Configuring systems took time. The services firm sold time, marked it up, and the buyer accepted the markup because the alternative was harder than it looked.

AI moves the bottleneck. The hard part of building software used to be the typing. The hard part is now the specification. Knowing what to build, in what order, with what tradeoffs.

This is a billing story. The time-and-materials contract has no SKU for specification. There is no PO line item that reads “thirty minutes of clarity that replaced three hundred hours of work.” Procurement systems were built downstream of execution. They cannot price upstream of it. When the work compresses, the revenue compresses faster than the value does.

Anthropic and OpenAI are not selling API access. They are selling embedded engineers who walk into a hospital, sit down with the clinicians, and build the workflow. From Anthropic’s announcement: “An engagement might begin with the company’s engineering team sitting down with clinicians and IT staff to build tools that fit into the workflows that staff already use.” (Anthropic)

That is a consulting engagement. Word for word.

The bottleneck for enterprise AI adoption is not model quality. It is integration. Goldman’s head of asset management said it directly this week. There is a “big shortage” of people who know how to integrate AI with existing business processes. (SiliconANGLE)

The traditional services firms assumed they would fill it. They have the relationships, the headcount, the procurement contracts, the methodologies. They were the obvious answer. The model labs reached a different conclusion. The integration work is too important to outsource to firms whose business model depends on dragging it out. So they built it themselves, with private equity money, distributed through portfolio companies they could not reach otherwise.

The traditional services firms have to win deals. The joint ventures walk into deals already won by their owners.

Blackstone owns hundreds of mid-market companies. Hellman and Friedman owns hundreds more. TPG, Bain, Brookfield, the same. OpenAI’s PE partners alone have access to more than two thousand portfolio companies. (Business Standard) Combined revenue in the trillions. The joint ventures get preferential access to all of it.

There is an optimistic read for incumbents. The total addressable market is large enough that even with the model labs taking a slice, the remaining slice keeps Accenture, Wipro, Infosys, and the rest fully employed for a decade. The pie grows faster than any one player can capture.

The non-optimistic read is that the model labs took the most valuable slice. The strategic transformation work, the boardroom-visible AI initiatives, the projects that become reference cases. They left the rest. Maintenance, integration tax, ticket queues. The firms that survive will look more like managed services providers than consultancies, and the multiple compresses accordingly.

The non-optimistic read is mostly right, with one modification. The most valuable slice is not the work itself. It is the relationship. Once a model lab’s embedded team has been inside a hospital network for two years, learning the workflows, building the custom tools, training the staff, that hospital does not switch. The same switching cost that has kept Epic and SAP entrenched for decades. The model labs are buying the lifetime relationship, using AI as the entry point.

Which is exactly what the IT services firms thought they were doing. The model labs got there with a better product, more capital, and distribution baked in.

The pressure on the traditional model does not come from slow erosion. It comes from a renewal cycle that does not renew, two or three years from now, when the embedded engineers from Anthropic and OpenAI have already built what the next contract was supposed to build.

Public market valuations are still pricing the old model. Median EBITDA margins stuck at thirteen to fifteen percent. Everyone reads that as stability. It is what happens when an industry has optimized every other variable and run into a structural ceiling. The next move in margins is down, and it comes from the revenue side.

The IT services firm got rich because the buyer did not speak the language. The firm was a translator. AI now speaks both languages. As of Monday, AI showed up with embedded engineers, private equity distribution, and a balance sheet that does not need to bill hours to stay alive.

The translators are still in the room. The buyer left.