On Monday OpenAI launched a separate company whose product is not a model. The new entity, the OpenAI Deployment Company, will take an initial four billion dollars in commitments and embed Forward Deployed Engineers inside client organisations to "design, build, test, and deploy production systems" connecting OpenAI models to the customer's data, tools and processes. Folded into the announcement is the acquisition of Tomoro, an applied-AI consultancy whose roughly 150 engineers (counting Tesco, Virgin Atlantic and Supercell among their clients) become the new venture's first staff once the deal closes.

The list of founding partners is the part worth staring at. BBVA, Goldman Sachs, SoftBank Corp, Warburg Pincus, B Capital, Emergence Capital, Goanna, WCAS. The investors include Advent, Bain Capital and Brookfield (led by TPG). And then, sitting in the same paragraph as the venture capital, three of the largest consulting firms on the planet: Bain & Company, Capgemini and McKinsey. This is not a typical investor roster. It looks more like the cap table of a systems integrator than of an AI lab.

That is, I think, the actual story. For three years the pitch from San Francisco has been: the model is the product, the API is the distribution, every other layer is commodity glue that customers will figure out for themselves. The new venture concedes, with a four billion dollar opening bid, that the glue is where the difficulty lives. Most enterprises cannot ship a production AI system from API access alone. Someone has to sit in their building, read their data schema, argue with their compliance team, instrument the failure modes, and stay around long enough to fix what breaks in week six. That is consulting work, and OpenAI has just stopped pretending otherwise.

The timing is uncomfortable for them. The same week DeployCo launches, Ramp's enterprise spend index shows Anthropic ahead of OpenAI in business adoption for the first time: 34.4 percent of Ramp's fifty-thousand-firm sample paying for Anthropic, against 32.3 percent for OpenAI. A year ago Anthropic was on nine percent. The lead is narrow and the methodology is partial, but the trajectory is the thing. The company that ignited the boom is no longer the default enterprise choice, and it is responding by hiring the kind of organisation it once said the boom would render unnecessary.

There's a coordination tax in any deployment that vendor benchmarks never capture; somebody has to pay it. DeployCo is OpenAI's bet that paying it directly, with engineers on the customer's floor, will let them charge for the surface area the API alone cannot reach. The interesting question is whether McKinsey and Bain see this as a partnership or as a beachhead. The previous wave of enterprise software, Salesforce and SAP and Oracle, ended up sharing the implementation pie with exactly these firms, and the implementation pie turned out to be larger than the licensing pie.

What you are watching, then, is not a strategic addition. It is OpenAI admitting that the investment math of building ever-larger models needs a different kind of revenue to support it, the kind that comes from being inside the customer's walls rather than behind a developer portal. The model business is still there. But the company has noticed, two and a half years late, that the model business by itself does not scale into a hundred billion of enterprise revenue. So it is becoming a consulting firm with a model attached, and it has brought the consulting firms along to help it do that.

Sources: