In May 2026 OpenAI launched a forward-deployed services arm called the OpenAI Deployment Company, which it majority-owns and controls, with more than $4 billion of initial investment from 19 investors led by the private-equity firm TPG, at what Axios reported was a $10 billion valuation before the new money. In the same month, TechCrunch reported that Anthropic announced its own services joint venture backed by Blackstone, Hellman and Friedman, and Goldman Sachs. A buyer evaluating whether to hire an outside firm to ship AI into production is right to ask why the labs that build the models would not simply do this work themselves. They are trying, and the shape of their attempts is the argument.
What the labs actually announced
Before the joint ventures, OpenAI had already built a consulting channel. In February 2026 TechCrunch reported that the company announced Frontier Alliances with Accenture, BCG, McKinsey, and Capgemini to deploy its enterprise platform, Frontier. Denise Dresser, then OpenAI's chief revenue officer, said in May that the enterprise business was on track to make up half of the company's revenue by the end of 2026, per the broadcaster CNA. The Deployment Company was built with distribution attached. OpenAI's announcement lists McKinsey and Capgemini among its investors, says it will work alongside the Frontier Alliance partners, and says its investment and consulting partners sponsor more than 2,000 businesses. To bring the private-equity firms in, OpenAI offered them preferred equity stakes with a guaranteed minimum return of 17.5 percent, Reuters reported in March, and The Next Web's reading is that the firms, in return, agree to make their portfolio companies available as "a captive enterprise customer base." TechCrunch sees the same logic in both ventures: each takes money from alternative asset managers to give its lab a new sales channel, and each will "presumably get preferred sales access" to the companies its investors own.
Read together, the announcements describe a forward-deployed services model: engineers placed inside a customer's operation to ship the vendor's platform into production. That is the same delivery shape a buyer would want from any firm hired to fix a workflow. So the question is not whether the labs can field forward-deployed engineers, but what incentive sits behind the engineer when a workflow could be served by a different model than the one their employer sells.
A model company's services arm is a channel for its own model
The first thing that follows is structural. A services arm owned by a model company exists to deploy that company's model. The Next Web sums up OpenAI's strategy as making private-equity portfolios "a captive distribution channel," and Anthropic says its venture will work with mid-sized companies "to bring Claude into their most important operations." That is model-captive by design, which is a description of why the vehicle was funded rather than a criticism of either lab. A venture raised to grow a model maker's enterprise revenue is measured on consumption of that maker's model.
This is a different incentive from a buyer choosing the best tool for a given workflow. Some steps in a real operation suit a frontier reasoning model, some suit a small cheap model, and some suit no model at all and want deterministic code or a rules table. A firm that earns on one vendor's consumption has a reason to route work toward that vendor even where a different choice would serve the workflow better, while a firm with no model of its own to sell has no such reason. The distinction matters most on the unglamorous middle of a process, where the cost-correct answer is often the smallest model that clears the bar rather than the most capable one. Access is the moat, not the algorithm makes the related point that the durable advantage in this work sits in the integration and the data, not in which model answers a prompt.
The economics point at scale
The second thing that follows is about contract size. OpenAI's venture has more than $4 billion of raised capital to put to work, and it will do that alongside the tier-one consultancies of OpenAI's Frontier Alliances, two of which are among its investors. That combination is built for large-enterprise contracts. Those firms' delivery economics, their bench rates, and their account thresholds are calibrated for the Fortune 500. The arithmetic of deploying a multi-billion-dollar services vehicle alongside tier-one consultancies does not bend toward the regional manufacturer with two hundred employees or the freight carrier with a three-person back office.
That leaves a large part of the economy underserved by the labs' own channel. The mid-market operator does not have an in-house AI team, does not write seven-figure consulting checks, and does not appear on a tier-one firm's target account list. Why services-as-software firms scale sets out why the delivery firms built for this segment compound differently from per-seat software vendors, and a map of the third layer places the various forward-deployed firms relative to the labs and the legacy consultancies.
The honest complication
The clean version of this argument breaks on one fact, and it should be conceded plainly. Anthropic's own announcement of its services firm says the venture will work with mid-sized companies across sectors, and it names community banks, mid-sized manufacturers, and regional health systems as the companies that stand to gain. That is the same segment described above as underserved. So at least one lab has named the mid-market as its intended customer, and a buyer should not be told that the labs have no interest in it. They do.
What survives the concession is the category distinction rather than the customer-list overlap. Even a lab firm that sells to a regional manufacturer is selling that lab's model into that manufacturer. The captivity is the same regardless of contract size, because it is a property of who owns the services arm rather than of how big the customer is. A buyer who wants the work routed to whatever combination of frontier model, small model, and plain code serves the workflow at the lowest defensible cost is buying against an incentive when the firm doing the routing earns on one vendor's consumption.
Where the independent model is the right answer
There are two further properties a lab-owned firm cannot structurally offer. The first is a horizontal pattern library built across many small operators rather than a few large ones. A firm that has redesigned the cash-application step in a wholesale distributor, the three-way match in a manufacturer, and the carrier-settlement step in a freight broker carries reusable structure across those instances, because they are instances of the same horizontal pattern. A channel optimized for a handful of very large enterprise contracts accumulates depth on a few accounts, not breadth across the long tail of mid-market operations. The third layer of consulting describes how that cross-customer pattern library forms and why it is the asset.
The second is independence from any single model vendor's consumption incentive, which is the same point stated as a commitment. The MIT NANDA initiative's State of AI in Business 2025, in preliminary findings built on interviews and self-reported surveys, estimated that about ninety-five percent of enterprise generative-AI pilots were delivering little or no measurable impact on the profit-and-loss statement, and traced the failures not to model quality but to a learning gap, in tools that don't retain feedback, adapt to context, or fit the workflows around them. If the constraint is integration, the right firm is the one that optimizes for the integration and treats the model as an interchangeable component. A buyer reading this with the labs' announcements in hand has the strongest possible evidence that the labs are serious about services, and the reason to hire an independent firm anyway is not that the labs are unserious. It is that a firm with no model to sell is the only one whose interests align with picking the right tool for each step of the work.
Common questions
- If the AI labs are building their own forward-deployed services arms, why hire an independent firm instead?
- A services arm owned by a model company exists to deploy that company's model, so it is measured on consumption of that model. An independent firm has no model of its own to sell, which means its interests align with routing each step of the work to whatever combination of frontier model, small model, or plain code serves the workflow at the lowest defensible cost. The reason to hire an independent firm is not that the labs are unserious about services, but that only a firm with no model to sell has no incentive to favor one vendor's consumption.
- Will a lab's services arm work with a mid-market company, or only large enterprises?
- It varies by lab. Per Anthropic's own announcement, cited in the post, its services venture will work with mid-sized companies across sectors, naming community banks, mid-sized manufacturers, and regional health systems, so at least one lab has named the mid-market as its intended customer. The post notes, though, that a vehicle capitalized with billions and working alongside OpenAI's Frontier Alliance consultancies, Accenture, BCG, McKinsey, and Capgemini, is generally calibrated for large-enterprise contracts, which tends to leave smaller operators underserved by that channel.
- Why does model neutrality actually matter for the results of an AI deployment?
- Different steps in a real operation call for different tools, since some suit a frontier reasoning model, some suit a small cheap model, and some are best served by deterministic code or a rules table rather than any model. The post cites the MIT NANDA initiative's State of AI in Business 2025, a preliminary report built on interviews and self-reported surveys, which estimated that about ninety-five percent of enterprise generative-AI pilots were delivering little or no measurable impact on the profit-and-loss statement and traced the failures not to model quality but to a learning gap, in tools that don't retain feedback, adapt to context, or fit the workflows around them. If the constraint is integration, the right firm optimizes for the integration and treats the model as an interchangeable component, which a firm earning on one vendor's consumption has reason not to do.