The question stopped being how much AI revenue you have and became how much of it the client keeps.

The companies that sell intelligence are cutting its price on purpose, while the consulting tier bills to manage the meter. The flip from renting to owning.

While the services tier was busy disclosing its AI revenue, the companies that sell the underlying intelligence disclosed something far more threatening to it: they are driving the cost of that intelligence down on purpose. In the same window, OpenAI unveiled its first custom chip, a processor built with Broadcom and named Jalapeño, designed specifically for running finished AI models. Broadcom's chief executive said it cuts the cost of that step by roughly half. OpenAI president Greg Brockman framed the whole effort as building the full stack, top to bottom, to make intelligence cheaper to serve.

The cloud provider said the same thing from the other direction. AWS chief executive Matt Garman told enterprises to stop reaching for the biggest, most expensive model by reflex and to pick the smallest one that does the job, and he declared that enterprise AI returns are now real, citing a room of technology chiefs in which nine in ten said they were seeing a return or a near path to one. Read the two together and the message is unmistakable. The labs and the cloud providers are collapsing the price of a unit of intelligence, and they are doing it deliberately.

That is the ground shifting under every firm whose AI line rides on rented models. A billion-dollar AI services business sits on a cost base the seller does not control and its own suppliers are working to shrink. The tier's answer has been to turn the cost into one more thing to bill for. Accenture is growing a practice, modeled on the cost-control discipline it built for cloud, to help clients optimize their spending on model usage. It is a real service answering a real anxiety. It is also selling umbrellas in the week the forecast turned dry, because the cost it proposes to manage is the one the suppliers just announced they are crushing.

So the question quietly flipped. For a year it was how much AI revenue do you have. This week it became how much of that revenue is something the client owns and keeps once the model underneath is swapped for a cheaper one. That is the one line that does not get cheaper when compute does, because its value is neither the intelligence nor the hours, but the owned, auditable result. It is also the shape of a different offer: a fixed price known before the work runs, in which the extraction step burns no tokens and the single bounded model step is priced in, leaving the client with a procedure it owns rather than a meter it rents.