The model companies just launched services firms. None of them is neutral.
Since May, every company that builds a frontier model has launched a firm to deploy it — OpenAI's $4B Deployment Company, then this month Microsoft's $2.5B, 6,000-person Frontier unit and Anthropic and Blackstone's $1.5B Ode. It is the strongest demand signal enterprise AI has produced — the model makers themselves are now betting billions that implementation, not models, is the next trillion-dollar business. But each firm is built to sell its owner's model, and that conflict matters more every month: frontier models are converging on the same capabilities while open weights close the gap at a fraction of the cost. The advantage now belongs to whoever can match the right model to each job, with no owner to answer to.
The signal is real
The companies that know most about where AI value comes from all reached the same conclusion at once: the money is in implementation, not models, and they each staked billions on it. TechCrunch put it in its headline on the Ode launch — the next trillion-dollar business is implementation, not the models.
When the people who make the models decide the returns are in helping you deploy them, the deployment gap is not a consultant's talking point. It is what the model makers are now betting their own capital on. Budgets, awareness, and urgency for this work are about to climb.
Every one of them sells its owner's model
Ode runs Claude-first. The Deployment Company is OpenAI-first. Frontier runs on Azure. Each will reach for a rival model if a client insists, but none is built to, and none has a reason to. The recommendation is set by the cap table before the first meeting.
That is how ownership works, not a scandal. A firm owned by a model company has one honest answer to "which model should we build on," and it was decided before you walked in. You get an excellent implementation of a choice someone else made for you.
The models are converging
This matters more now than a year ago. The capability gap between frontier models has narrowed to the point where, for most enterprise workflows, the model is no longer what decides whether the system works. Prompting, data access, evaluation, and workflow design decide that. Swap one frontier model for another on a well-built system and most users will not notice.
Open weights are closing what is left. Open models now rival the closed frontier on coding and reasoning at a fraction of the cost — GLM-5 matching Claude on SWE-bench, and DeepSeek, Kimi, and Qwen shipping frontier-class models you can host yourself. When the models were far apart, committing to one was a bet on capability. Now it is a bet on a single vendor's prices and roadmap.
Model-agnostic is the edge
Three things decide how well a deployment goes: which model fits the workflow, what it costs, and how fast you can move when a better or cheaper one ships. In a converging, accelerating market, those answers keep changing, and the best one is rarely the same model twice.
A firm with no model in its cap table can send each workload to whatever wins on quality and price, use open source where it fits, and switch as the frontier moves. A vendor-owned firm cannot, without undercutting the owner it exists to sell. Independence is the position that lets you get the most out of each model and pay the least for it, and keep doing both as the ground shifts.
The model companies have told the market that implementation is where the value is. They are right. They cannot also be the ones who tell you which model to use — and as the models converge, that is the answer that matters most.