OpenAI just bid for the other 86% of the market

For every dollar an enterprise spends on AI software, roughly six dollars go to the people who make it work. Integrators, system implementers, change-management consultants, the engineers who sit inside a client's office for eighteen months turning a model endpoint into something that books revenue. Until this week, OpenAI was a model company. It addressed the dollar. As of Monday, through a separate legal entity capitalised at four billion dollars and seeded with a hundred and fifty engineers acquired from a UK consultancy called Tomoro, it is making a direct bid for the six.¹

This is not a product launch. It is a repositioning of where OpenAI intends to capture margin in the enterprise stack, and the choice to do it through a standalone subsidiary at a ten-billion-dollar pre-money valuation, with Goldman Sachs, Brookfield, and Bain Capital on the cap table, tells you exactly how seriously the company and its investors are taking the move.¹ The question worth asking is not whether the strategy makes sense on paper. It plainly does. The question is whether OpenAI can execute it without breaking the things that made the model business work in the first place.

The market OpenAI was actually in

Start with the ratio. The figure that most cleanly explains this announcement is the services-to-software multiple in enterprise AI deployment, which industry analysis has settled at roughly six to one.² Every dollar of model API revenue or licensed software pulls six dollars of integration, customisation, and change-management work behind it. That is not unique to AI; enterprise software has always carried a services tail. What is unique is the ratio. Traditional enterprise SaaS runs at roughly one to one, sometimes two to one for complex implementations. Six to one is closer to bespoke industrial automation than to Salesforce.

The reason the ratio is so high is the same reason ninety-five per cent of enterprise generative AI pilots show no measurable return on investment, per the MIT NANDA research published last year.² Models work. Deployment is hard. The gap between a working demo and a system that holds up against a customer service queue, a compliance regime, an existing data warehouse, and the quiet preferences of the people who actually have to use it is enormous, and that gap is where the six dollars live.

OpenAI has been competing for the one dollar against Anthropic, Google, and an increasingly capable open-weight tier. Margin in that market is under structural pressure; inference costs are falling, frontier capability gaps are narrowing, and the API business looks more like a commodity every quarter. The Deployment Company is an explicit acknowledgement that the bigger pool of profit was never in the model layer. It was in the integration layer all along, and OpenAI is no longer content to let McKinsey, Accenture, and the Big Four capture it on the back of OpenAI's own technology.

The Palantir template, with the volume turned up

The structural precedent is Palantir. The Forward Deployed Engineer model, where senior technical staff embed inside client operations for months or years rather than billing as external consultants, was Palantir's signature go-to-market and the reason it grew into government and commercial accounts that traditional software vendors could not crack.³ Palantir FDEs are not implementers in the Accenture sense. They sit in the client's office, learn the client's workflow, and build on Palantir's platform from the inside. The economics are unusual; the contracts are sticky; the competitive moat is the relationship, not the software.

OpenAI is copying the template directly. The Tomoro acquisition gives it roughly a hundred and fifty FDEs at launch, and Lightcast data shows FDE job postings up five-fold year-on-year across the sector as OpenAI, Google, and increasingly Anthropic compete for the same talent pool.³ The standalone entity structure matters here. Folding the function into the core API business would have created two problems. First, the margins are different; consulting-style services revenue has a fundamentally different shape from API revenue, and mixing them in one P&L obscures both. Second, the talent profile is different; the engineer who wants to embed at a Fortune 500 insurer for fourteen months is not the same engineer who wants to ship the next version of the responses API, and pretending otherwise loses you both.

The standalone structure is a tell. OpenAI is not adding a services team. It is building a services company.

The four billion dollar raise at a ten billion pre-money is the other tell. That is not a "let's experiment with this" cheque. That is a "we intend to build a top-five enterprise services franchise" cheque, and the investor syndicate, particularly Bain Capital sitting next to Bain & Company's direct competitive exposure, is the kind of detail that reads as deliberate. Either Bain Capital is hedging its parent firm's market, or it believes the AI services market will expand faster than it redistributes. Both readings are defensible. Neither is comfortable for the incumbents.

The neutrality problem

Here is where the Palantir analogy starts to leak. Palantir FDEs succeed in part because clients accept that they are building on Palantir's platform. The conflict of interest is priced in from the first conversation. The client has chosen Palantir; the FDE's job is to make Palantir work, not to evaluate alternatives.

OpenAI Deployment Company FDEs will face a structurally trickier conversation. Enterprise AI deployments in 2026 are increasingly multi-model. A serious financial services client will run Claude for long-context reasoning, GPT-class models for general agentic work, a Gemini model for certain multimodal tasks, and probably an open-weight model behind the firewall for anything sensitive. The integration work is precisely the question of which model goes where, and an OpenAI FDE sitting in a client's office cannot give a neutral answer to that question. The moment the head of data engineering asks whether they should also be evaluating Anthropic for the contract analysis workflow, the FDE is in a bind that an Accenture consultant simply does not have.

This does not kill the model. It caps it. The Deployment Company will win the workloads where OpenAI is unambiguously the right model, and it will lose, or be locked out of, the architectural decisions where multi-model is the right answer. Over time, sophisticated clients will figure out that they want a neutral integrator for architecture and an OpenAI FDE for OpenAI-specific implementation, which is the relationship most enterprises currently have with their cloud hyperscalers and their cloud-agnostic consultants. That is a viable business. It is not a winner-takes-most business, and the ten billion dollar valuation is priced closer to winner-takes-most than to winner-takes-its-share.

The ninety-five per cent question

The MIT NANDA figure is doing more work in this announcement than it appears to be. Ninety-five per cent of enterprise generative AI pilots showing no measurable ROI is the data point that justifies the entire Deployment Company thesis.² If the ROI problem is a model problem, then no amount of FDE labour fixes it, and OpenAI should be spending the four billion on training compute. If the ROI problem is a deployment problem, then an FDE corps that solves the deployment problem at scale captures most of the value that the model alone cannot.

OpenAI has clearly decided it is a deployment problem. That is a defensible bet. Every enterprise AI failure I have seen described in detail in 2025 and 2026 reads as a deployment failure: bad data plumbing, missing evaluation infrastructure, no change management for the human users, no monitoring once the system was live, no clear owner when it broke. The models worked fine in the demo. They failed in the wiring.

But the figure cuts both ways, and the counter-case is worth taking seriously. If deployment quality were the binding constraint, clients with access to the best implementation talent, the largest consultancies, the most expensive system integrators, should be outperforming. The evidence base for that is thin. Accenture, Deloitte, and the major SIs have been deploying generative AI for two years with their best people, and their clients are not visibly clearing the ninety-five per cent failure bar by a wide margin. That suggests either that the implementation problem is harder than even the best implementers can currently solve, in which case a hundred and fifty OpenAI FDEs are not going to solve it either, or that the ROI gap has structural causes beyond implementation quality. Both possibilities should worry an investor in the Deployment Company.

The honest reading is that deployment quality is probably necessary but not sufficient. Better FDEs will lift the success rate. They will not lift it to ninety-five per cent any time soon, and the Deployment Company will discover the same thing every consultancy has discovered, which is that some clients are uncoachable, some workflows are not yet ready for the technology, and some procurement processes destroy the value before the work even begins.

The scale problem nobody is talking about

A hundred and fifty FDEs is not a services business. It is the seed of one. McKinsey runs roughly forty-five thousand consultants. Accenture's technology and AI practice is measured in hundreds of thousands. Even Palantir, which is small by services-firm standards, employs several thousand engineers in client-facing roles after two decades of growth.²

To put it in the bluntest terms: the Deployment Company at launch can serve a handful of strategic accounts well, or a larger number of accounts badly. It cannot, by simple arithmetic, threaten the incumbent SI capacity within any planning horizon shorter than five years, and that is assuming it can hire at a rate that no professional services firm has ever sustained. The talent pool for genuinely capable FDEs is small, the five-fold posting growth has already bid up compensation, and the FDE archetype, which is a senior engineer willing to embed in someone else's office for a year, is not the modal Bay Area hire.

This is where the standalone-entity choice becomes interesting again. By spinning the Deployment Company out, OpenAI has created a vehicle that can compensate, structure, and hire differently from the parent. It can pay services-firm-style bonuses, offer client-equity-style upside, build a partnership track. It can also, crucially, acquire its way to scale, which is what the Tomoro deal already signals. Expect more acquisitions. A four-billion-dollar war chest is not for organic hiring; it is for buying the next three Tomoros, and probably one mid-sized boutique consultancy with sector-specific IP.

Even so, the headcount mathematics is brutal. To get to ten thousand FDEs, which would be roughly the scale at which the Deployment Company starts to look competitive with a single Big Four practice area, OpenAI would need to acquire or hire seventy more firms the size of Tomoro, or grow at a rate roughly five times faster than Palantir's services arm did. Neither is impossible. Both are very hard.

What Microsoft just lost

The most quietly consequential party in this announcement is Microsoft. Microsoft coined the phrase "frontier firm" and built much of its enterprise AI positioning around being the channel through which OpenAI reached the Fortune 500. Azure OpenAI Service, Copilot for the enterprise, and the broader Microsoft partner network of system integrators were supposed to be the deployment layer that turned OpenAI's models into enterprise revenue.

The Deployment Company changes that relationship in ways that are not yet being discussed openly. An OpenAI-owned entity with its own FDEs has direct visibility into the model roadmap that Microsoft's SI partners do not. It has preferential pricing that Microsoft's SI partners do not. It has first sight of enterprise requirements that, in a clean partnership, would flow back to Microsoft. The channel conflict is not hypothetical; it is built into the structure of the new entity.

Microsoft built the runway. OpenAI just bought the airline.

Microsoft's options are limited. It cannot easily build a competing FDE corps without strategic clarity about whether it is competing with OpenAI or partnering with it, and the answer to that question has been ambiguous for at least eighteen months. It can lean harder on its own model investments, which it has been doing, but that creates a different conflict: the more Microsoft promotes its own models, the more the OpenAI partnership becomes a hedge rather than a strategy. The Deployment Company makes that fork sharper, not softer.

What to watch

The next six to twelve months will tell you which of three things this is. The first possibility is that it is the move it appears to be: a serious, well-capitalised, structurally sensible bid for the services layer that grows by acquisition, builds sector-specific capability quickly, and within five years is a meaningful presence in enterprise AI deployment. In that world, the incumbents lose share but not the market, and the Deployment Company settles into a Palantir-shaped position at roughly Palantir-shaped scale.

The second possibility is that it is a defensive move dressed up as an offensive one. If the API business is commoditising faster than OpenAI has publicly admitted, the Deployment Company is a way to capture services margin to offset model-margin compression. That is still a viable business. It is a different business from the one the ten billion dollar valuation implies, and it would mean the investor syndicate is pricing growth that is actually substitution.

The third possibility is that it is a category error. If the ninety-five per cent ROI problem turns out to be structural rather than implementational, the Deployment Company spends three years discovering that no amount of FDE talent fixes the underlying issue, burns through a meaningful fraction of the four billion, and ends up either pivoted into something narrower or quietly wound down. I do not think this is the likeliest outcome, but it is more likely than the announcement's framing suggests.

The signal to watch is the second and third acquisitions. Tomoro on its own is a seed; what OpenAI buys next will tell you whether the Deployment Company is building horizontal capacity, vertical depth, or geographic reach. Each implies a different theory of where the services market is going. A second UK or European consultancy implies geographic build-out. A US-based vertical specialist, particularly in financial services or healthcare, implies a sector-depth strategy. A pure-play AI evaluations or observability firm would imply that OpenAI thinks the missing piece is post-deployment tooling rather than deployment labour itself.

The Deployment Company is the most interesting structural move any frontier lab has made in the last twelve months, and the strategic logic is sound. The execution risk is real, the scale gap is enormous, and the neutrality problem is going to cost the entity workloads it would otherwise have won. But the underlying read of the market, that the money is in the six dollars and not the one, is correct, and OpenAI is the first frontier lab to act on it at this scale. Everyone else is now playing catch-up to a thesis that was sitting in the services-ratio data the whole time.

The six dollars were always going to find an owner. OpenAI has decided it might as well be them.

Footnotes

¶