On May 4, 2026, within a single trading day, the two most valuable private AI companies in the world both announced they were going into the consulting business.
Anthropic announced a joint venture focusing on deploying enterprise AI services, with Blackstone, Hellman & Friedman, and Goldman Sachs as founding partners, backed by Apollo, General Atlantic, GIC, Leonard Green, and Sequoia. The Wall Street Journal first reported the venture was valued at $1.5 billion, including a $300 million commitment each from Anthropic, Blackstone, and Hellman & Friedman. Mere hours before, Bloomberg reported that OpenAI was raising funds for a similar venture called The Development Company. OpenAI's venture would operate at larger scale, raising $4 billion from 19 investors against a $10 billion valuation, with TPG, Brookfield, Advent, and Bain Capital among the named investors.
A week later, OpenAI showed its hand. It announced the acquisition of Tomoro, a UK-based AI consulting firm, to staff up its newly launched Deployment Company with approximately 150 Forward Deployed Engineers from day one. The Deployment Company is a majority-owned OpenAI subsidiary designed to embed engineers directly inside client organizations. The approach is modeled on Palantir's forward-deployed engineer strategy, which proved the real value in enterprise software is the implementation, not the license.
This is the moment the API era ended. The buyer who reads this as a normal procurement event, a new SKU from a familiar vendor, will hand over the next decade of operating leverage without noticing.
The quiet admission
For three years, the pitch from the frontier labs was that the model was the product. Buy access to the weights through an API, plug it into whatever you already had, watch productivity rise. Every CIO deck assumed this shape. Every "AI strategy" document treated model choice as the central question.
The labs themselves now disagree.
The model layer is commoditising. The application layer is fragmenting. The services layer, the part where engineers sit inside companies and make AI work, is where the margins are migrating. That sentence comes from the trade press covering the Tomoro deal, but it could just as easily be paraphrased from any of the lab fundraising memos that leaked over the last month.
Look at the math. Anthropic's business reached $30 billion in annualized revenue, and the company is in talks to raise capital at a $900 billion valuation, past OpenAI's most recent $850 billion mark. Enterprise now makes up more than 40% of OpenAI's revenue, and the company expects that portion to reach parity with consumer by the end of 2026. The growth is real. The growth is also stalling at the point where a customer has a license but no production system.
OpenAI reports that over one million businesses have used its products and APIs in recent years. The company has identified a need for tailored support to transition from pilot AI projects to full-scale implementation.
Translation: a million logos signed up, and most of them never crossed the line from demo to durable workflow. The labs spent three years assuming the integration problem would solve itself. It didn't. Now they are buying the people who can solve it, and they are using private equity capital to do it.
The Palantir lesson, applied at scale
Palantir pioneered the forward-deployed engineer model over years of defence and intelligence engagements where software had to work inside institutions too complex for remote support. The company sent its own engineers directly to intelligence agencies, military clients, and later private-sector companies because its platform was nearly unusable without heavy customisation. That operational intimacy drove Palantir's US commercial revenue to surge 133 per cent year on year.
The lesson Palantir taught the industry: the engineer in the building is the moat. The software is replaceable. The relationship, the workflow knowledge, the trust earned across six quarters of shipping inside a client's operations, that is what holds a customer for ten years. Tomoro's 150 engineers become the founding cadre of a deployment operation that will scale through further acquisitions funded by the four billion dollar war chest. The engineers will not sell software. They will sit inside enterprises and build the systems that make OpenAI's software produce business outcomes.
Now read that sentence as a buyer. An OpenAI employee, paid by OpenAI, reporting to OpenAI, will sit inside your finance function or your claims process or your supply chain for as long as it takes to design the production system. They will write the prompts. They will design the agent topology. They will pick the evaluation harness. They will know exactly where your data lives and exactly which workflows are most exposed to model substitution.
That person is loyal to the lab. They have to be. Their stock is in the lab. Their career runs through the lab. Their next promotion depends on more OpenAI tokens flowing through your systems next quarter than this one.
The procurement event that isn't
OpenAI portrays its services unit as a benefit due to its vertical integration, but you can expect CxOs to view the effort through the lens of lock-in. The acquisition of Tomoro is smart because it gives the Deployment Company a running start. It will be interesting to see how this services launch impacts its other partners.
Constellation Research said that politely. Let me say it less politely.
When the company that owns the model also owns the consultants who design your system around the model, three things happen. First, every architectural decision the consultant makes will tilt toward token volume on their employer's API. As Ramp's economist Ara Kharazian wrote about Anthropic, the company makes more money when businesses purchase more tokens, so it is incentivized to drive users to more expensive models, even when cheaper models are sufficient and faster for many tasks. The forward-deployed engineer is the human expression of that incentive.
Second, the cost of switching becomes the entire deployment, not the API contract. You can move from one frontier model to another in an afternoon. You cannot move six months of system design, prompt engineering, evaluation infrastructure, and tool integrations in an afternoon. The consulting layer is the lock-in layer. That is what the labs have figured out. That is what they are charging private equity to fund.
Third, your internal team gets quietly hollowed. The question is no longer whether OpenAI competes with the consultancies. It does. The question is whether the FDE motion produces deployments good enough to displace the in-house data science org that the customer already employs. The two-page case studies coming out of Tomoro's existing clients should make uncomfortable reading for the heads of internal AI inside Mattel, Red Bull, Tesco, and Virgin Atlantic.
Your data scientists wake up to find an OpenAI engineer in the standup. The engineer is faster, has direct access to model roadmaps your team will see in six months, and reports to a partner with $4 billion in deployment capital. What happens to the internal AI function over the next four budget cycles is not hard to predict.
The private equity hand on the wheel
The detail that should keep operators awake is not the consultancy itself. It is the capital structure.
The Deployment Company structure provides strategic distribution. Its private equity backers sponsor more than 2,000 businesses globally, giving OpenAI a captive channel into portfolio companies that are already under pressure from their PE sponsors to boost productivity. OpenAI's Frontier enterprise platform, which already counts HP, Intuit, Oracle, State Farm, and Uber as adopters and has formed alliances with BCG, McKinsey, Accenture, and Capgemini, now gets a dedicated deployment arm to convert those relationships into production systems.
Read the investor list. TPG as lead investor, Advent and Bain Capital and Brookfield as co-leads, and a long tail of names that read like a who's-who of late-stage private equity: SoftBank Corp., Goldman Sachs, Warburg Pincus, B Capital, BBVA, Emergence Capital, Goanna, WCAS. Three management consulting firms also wrote checks: Bain & Company, Capgemini, and McKinsey & Company.
The same private equity firms that own the companies the consultants will serve are funding the consultants. The same management consulting firms whose partners will be displaced are co-investing in the entity that displaces them. The 17.5% guaranteed return is worth unpacking. The average annual return of the S&P 500 over the past decade has hovered around 10-12%. A guaranteed 17.5% suggests the investors negotiated hard, and OpenAI needed this capital badly enough to agree to expensive terms.
A guaranteed return that high means the deployment work has to produce enormous, sustained, lock-in-grade revenue across the PE sponsors' portfolios. That is the contract. The forward-deployed engineer walking into your office on Monday is the operational expression of a balance sheet promise to TPG.
What this means for the buyer
If you are a CEO, founder, or operator who has been treating model selection as the strategic question, you have been answering the wrong question for two years. The labs just told you so by spending six billion dollars to fix it.
The strategic question is architectural. Who designs the system that wraps the model around your workflow? Who owns the prompts, the tool definitions, the evaluation harness, the agent boundaries, the data routing logic, the failure modes, the human-in-the-loop checkpoints? Who has the right to change vendors when the price doubles or the model regresses or a better option appears from a lab that does not yet exist?
If the answer is "the model vendor's consultancy," you have signed away the part of the stack that actually compounds.
Consider what Anthropic's parallel move tells you about how serious this is. Anthropic secured 5 gigawatts worth of computing capacity as part of an announcement with Google and Broadcom that will start to come online next year. Google said earlier this month that it plans to invest up to $40 billion in Anthropic. Many industry analysts judge that the landscape has evolved from the "Big Three" to a two-power standoff between the Anthropic camp and OpenAI. The term "Big Three" is becoming history.
Two camps. Both vertically integrated from silicon to consultant. Both willing to underwrite 17.5% guaranteed returns to private equity to plant engineers inside Fortune 500 buildings. Both betting that whoever owns the deployment layer in 2026 owns the enterprise customer through 2035.
The independence problem
There is a reason the medical world separates the diagnostic radiologist from the company that sells the MRI machine. There is a reason the auditor cannot also be the CFO. There is a reason your tax attorney does not work for the IRS. Independence of advice from product is a structural protection, designed in over centuries because the failure mode is obvious and the failure is expensive.
Enterprise AI just collapsed that separation in a single week.
The implication is not that OpenAI and Anthropic are bad actors. They are rational actors, executing a sound strategy with a clear precedent and patient capital. The implication is that the buyer who wants advice oriented to the buyer's interest, rather than to the vendor's token volume, now has to source that advice from somewhere other than the vendor.
The choices look like this. You can take the lab's engineers and accept the lock-in, in exchange for speed and access to model roadmaps. You can build a strong internal team and refuse outside help, in exchange for slower deployment and a permanent talent retention problem. Or you can work with an independent partner whose incentives are aligned with your outcomes, not with any single lab's revenue per query.
The third path is the one that holds up under five years of model churn, price changes, regulatory drift, and competitive pressure. It is also the path the lab consultancies are designed to foreclose. Close of the Tomoro acquisition is the next concrete gate. Retention of the named engineers through that close, given that the deal value of DeployCo to OpenAI sits in the headcount itself, becomes a quiet execution risk on top of the formal regulatory clearances.
The labs know the engineer is the asset. They are paying premium prices to own that asset. Every month you wait to set up your own independent architecture function is a month the lab's engineer gets closer to being the only person in your building who understands how your AI system actually works.
What an architected response looks like
A serious response has three parts.
The first is a clear ownership map. Before any forward-deployed engineer crosses your threshold, you should know exactly which artifacts belong to your company and which belong to the vendor. Prompts. Evaluation sets. Agent definitions. Tool schemas. Routing logic. Data pipelines. Memory stores. If the vendor's engineer writes it on your laptop while sitting in your conference room, who owns the IP? Who has the right to take it to a competitor's model next year? The default contract language will favor the vendor. The negotiated language can favor you, if you ask before the engagement starts, and not after.
The second is a vendor-neutral abstraction layer. Your production AI systems should call a routing layer you control, not a model endpoint a vendor controls. Anthropic Opus 4.7 today, Gemini 3.1 next quarter, a model from a lab that does not yet exist the year after that. The cost of building this routing layer is small. The cost of skipping it is the rest of your competitive life.
The third is an independent strategic function, sitting above the deployment work, that answers to you. Someone whose job is to ask whether the agent the vendor's engineer just designed is the right agent, whether the workflow the vendor's engineer just rebuilt is the right workflow, whether the token consumption the vendor's engineer just baked in is justified by the marginal value. That function cannot be staffed by the vendor. It cannot be staffed by the vendor's investors. It cannot be staffed by a consultancy that co-invested in the vendor's deployment subsidiary.
It has to be staffed by someone whose only relationship is with you.
The window is open for a short time
More than one million businesses already use OpenAI's products and APIs, and that installed base, plus the consortium's portfolio companies, becomes the obvious early target for DeployCo's first engagements. Tomoro brings approximately 150 Forward Deployed Engineers and Deployment Specialists into DeployCo from day one, giving the new entity a working delivery team on launch day rather than a roadmap to hire one.
A hundred and fifty engineers, against a million businesses, with $4 billion in capital to scale. The math says the lab's consultants will arrive at your door within the next eighteen months, or you will be told you are no longer a priority account because you did not engage early enough. Either outcome is bad. The first ends in lock-in. The second ends in being out-deployed by a competitor who took the call.
The window to set up the right architectural defenses is open now, before the engineer is in the building. Once the engineer is in the building, every architectural decision gets harder, every conversation about ownership gets more political, and every effort to maintain vendor neutrality gets framed as obstruction.
The labs ran their playbook in May 2026. Your move.
Why this needs architecture, not procurement
You cannot buy your way out of this with a tool. There is no SaaS product that fixes a misaligned incentive structure. There is no AI gateway that retroactively negotiates IP ownership. There is no procurement template that converts a forward-deployed engineer into an independent advisor.
What is required is a strategic architecture, designed for your company, that decides in advance which parts of the stack you own, which parts you rent, which parts you let a vendor staff, and which parts you never, under any circumstances, hand to the entity that profits from your token consumption. That architecture has to exist before the lab's engineer walks in. After they walk in, you are negotiating from a weaker position every week.
Agor AI Advisory exists to design that architecture and defend it through the engagement. We work for you. We do not take fees from the labs. We do not co-invest with the deployment companies. We do not retain economic interest in your model choice. Our job is to make sure the system you build in 2026 is still yours in 2030, regardless of which lab wins the two-power standoff.
The model labs spent six billion dollars in a single week to put their engineers inside your operations. You can decide what that means for your company, or you can let their engineer decide it for you.