SecondSourceThe Second Source on AI

GPU price rises live in new short contracts; CoreWeave’s fleet average still lags in Q3 — resolves 2026-11-16

Not yet due

GPU price rises live in new short contracts; CoreWeave’s fleet average still lags in Q3 — resolves 2026-11-16

Scheduled check: 2026-11-16

Our earlier call, under recheck · Written on Oct 18, 2026

Background

The debate is how tight AI compute really is in autumn 2026: seller presentations show sharp rises on new short contracts, while public pay-as-you-go price indexes rose far less. If only new short contracts are rising, a GPU renter’s total revenue will not keep pace at once, and anyone estimating fleet revenue from seller presentations will overshoot. SecondSource reads the premium as concentrated in new marginal contracts with the fleet average lagging, while the opposing reading says the whole fleet is being repriced.

The call

SecondSource’s reading: GPU tightness in autumn 2026 shows up as a premium tied to contract length. Short contracts of 3 to 6 months and cancellable large orders run about $40 million to $60 million per MW (a megawatt of data-center power) per year, while five-year contracts run only about $10 million to $16 million.

CoreWeave, a cloud company that rents out GPUs, says it raised prices about 25% across its products in July. But average revenue per MW across its whole fleet will not catch up with that marginal price soon; a rough calculation from second-quarter figures puts it at about $7 million to $8 million.

Our bet: third-quarter 2026 revenue, annualized and divided by end-of-quarter active power, will still be below $10 million per MW per year.

Sources behind the call

CoreWeave, Inc.(SEC EDGAR,accession 0001769628-26-000429)
“~25% price increase across SKUs in July 2026”

The receipt

Grace period of 14 days after CoreWeave’s Q3 2026 earnings press release (expected mid-November 2026). Take the quarterly revenue stated in the release times 4, divided by the end-of-quarter active power (GW) stated in the same release, to get annualized revenue per MW.

What would overturn it

If the ratio is $10 million per MW per year or more, the reading that the premium sits at the margin and the fleet average does not keep up in the near term is wrong: the whole fleet is being repriced. Section three of this piece and the core sentence of its anchor insight are withdrawn together, and the reading becomes an across-the-board price rise.