OpenAI Just Admitted the Models Aren't Enough.

The company with the best models in the world just concluded that the models are not enough.

OpenAI stood up a deployment company. Four billion dollars, nineteen partners, and a workforce of engineers whose job is to sit inside your building and make the technology actually work.

I run a firm that does implementation. So read the rest of this knowing I have a dog in the fight, and read it anyway, because the conclusion holds either way.

What They Actually Built

DeployCo launched on May 12. TPG led the round with co-leads Advent, Bain Capital, and Brookfield. The founding partner list runs through B Capital, BBVA, Emergence Capital, Goanna, Goldman Sachs, SoftBank Corp., Warburg Pincus, and WCAS, with McKinsey, Bain & Company, and Capgemini also backing it. OpenAI put in $500 million of its own money.

Then it acquired Tomoro so it could start with roughly 150 Forward Deployed Engineers on day one instead of hiring them.

The Forward Deployed Engineer model is borrowed from Palantir, and the distinction that matters is physical. An FDE does not write a recommendation and hand it to your team. They embed in your operation, learn the workflow from the people who run it, and build against the actual mess rather than the version of the mess that appears in a slide.

That is a services firm. A very well capitalized one, sitting next to the model lab.

Read the Signal, Not the Press Release

Here is what a $4 billion services arm tells you about the state of enterprise AI.

If a general-purpose model were sufficient on its own, the company with the best one would sell subscriptions and go home. Distribution would be the whole game. Instead the lab that would benefit most from models-are-enough being true is spending billions to prove it is not.

The bottleneck is not capability. The bottleneck is the distance between a capable model and a business process that runs on Tuesday morning with real data, real permissions, real exception handling, and a real person accountable when it goes wrong.

That distance has a name. It is implementation, and it turns out to be expensive, unglamorous, and worth $4 billion.

Consulting Is Not the Same as Being Consulted

Watch which word DeployCo avoided.

Nobody described this as advisory. The pitch is engineers on-site, in your systems, shipping. Compare that to the last AI strategy engagement your company bought: a discovery phase, a maturity assessment, a roadmap, a use case matrix, and a recommendation to hire people who can build.

There is a reason the highest-signal move in the market skipped straight to the building part.

I have watched organizations spend two quarters and serious money producing an AI strategy document, then spend the next two quarters discovering that the document did not tell anyone what to type. The gap between capability and adoption never closes on paper. It closes when somebody who can write code sits with somebody who does the work.

What This Means If You Are the Buyer

Three things change in how you evaluate help.

Ask whether they build or advise. Not what they say in the pitch, what shows up in the statement of work. Deliverables that are documents produce documents. Deliverables that are working systems produce working systems.

Ask where they sit. On-site or embedded beats remote and quarterly, every time, for the same reason a good NCO stays with the platoon. The knowledge you need is in the hands of the people doing the job, and it does not survive translation into a requirements doc.

Ask what they leave behind. This is the one most companies skip. A partner who builds you something you cannot operate has sold you a dependency, not a capability. The exit criteria should be your team running it without them, and that should be written down at the start.

The Part Nobody at Your Company Is Saying

You are going to be offered this. Deep-pocketed implementation, staffed by people trained by the lab whose model you already use, with a Fortune 100 logo wall and a price tag to match.

Some of you should take it. If you are running a global operation with hundreds of workflows and no internal AI engineering muscle, that is a rational buy.

Most of you should not, and here is the honest test. Your problems right now are almost certainly not model problems. They are that your data is scattered, your workflows are undocumented, nobody agrees what the metric is, and three departments have three definitions of a customer.

No Forward Deployed Engineer fixes that for you. They will fix it around you, at their rate, and you will own the result without understanding it.

Build the boring stuff first. Clean data, documented workflows, one agreed definition of success. Then bring in whoever you want, and pay them to do the hard part instead of the part you should have finished yourself.

The Part That Matters

The most capable AI company on earth looked at its own product and decided it needed 150 engineers in customer buildings to make it land.

That is not a knock on the technology. It is the clearest statement anyone has made about where the work actually is.

The models are not the hard part. They never were.