The Contract Price Tell

The easy read on this week’s compute data was that the AI boom is finally cooling. Nathaniel Whittemore, citing SemiAnalysis, gave the harder read: spot GPU rentals are falling while contract prices keep rising. SemiAnalysis’s own gloss, quoted by Whittemore, is that this is “a shift of opportunistic capacity usage towards committed production deployment.” Translation: the marginal renter is losing to the buyer who signs for years.

Underneath that price signal is a usage signal moving the same direction. Whittemore also flagged that global token volumes are now running above 30 quadrillion per month and growing roughly 14x year over year. Whatever is happening to spot rates, the meter on actual model calls is not slowing. The capacity coming off spot is going somewhere, and it is going into contracts.

The reason buyers are willing to sign is that the top of the stack is consolidating fast enough to be worth locking in. Whittemore reported that GPT-5.6 Soul on Ultra settings scored 91.9% on Terminal Bench 2.0, beating Mythos by nearly four points, and that the same model hit a roughly 11.3-hour 50% time-horizon estimate when cheating attempts were marked as failures. Whether or not you trust the benchmark, enterprise procurement does: agentic coding has moved from “demo” to “the thing our engineers refuse to give back.”

You can see the same lock-in on the demand side, in the shape of the workflows being built on top. Andrew Ambrosino, speaking publicly, described Codex writing itself a Premiere Pro extension so it could drive the app from the outside, and noted that inside OpenAI “nobody would leave the Codex app for the apps that were allegedly for these other personas.” An engineer working on agent infrastructure put the operational cost of that world plainly: unlike normal enterprise software you can leave running for five years, “agents don’t really work that way. The core thing that you’re building on is constantly changing.” Committed compute is the natural hedge against a substrate that never sits still.

Then the political layer snapped into place. Whittemore’s sharpest line of the week was that frontier AI models “are now subject to a licensing regime. It’s a licensing regime that hasn’t been passed by Congress, established in an executive order, or even fully articulated. At this moment, it is a licensing model based on the whims of Howard Lutnick.” He paired that with Apple’s petition to the Trump administration to buy memory from CXMT, a Chinese supplier currently on the Pentagon’s blacklist. Frontier compute is now something you ask Washington’s permission to source and to ship.

That is what makes the contract-versus-spot divergence load-bearing rather than a footnote. Meta reportedly told teams to stop using Codex and Claude Code on certain tasks over fears that outputs could contaminate training data, per Whittemore: even the hyperscalers are treating rival model output as a controlled substance around their own pipelines. When the biggest buyers behave like sovereigns, they do not rent by the hour.

So the story to hold this week is not that AI demand is topping. It is that the market is quietly reorganizing itself around long-dated commitments, benchmark-driven vendor lock-in, and a discretionary federal licensing layer sitting on top. The soft spot price is the sound of that transition, not a warning about it.