Implied Committed Rates: What the Big Compute Deals Pay per GPU-Hour
The mega-deals that dominate the headlines ($38 billion, $300 billion) are actual negotiated committed-term prints, disclosed as lump sums rather than rates. This note backs the unit rate out of the disclosed figures, and every number it produces is a floor by construction: each assumption required to do the division biases the implied rate downward. What survives is an anchor for the rate the disclosures leave implicit. The cleanest print, Microsoft's Norway contract with Nscale, implies $2.72 per GPU-hour with every input company-disclosed. Each figure below is implied, computed by CCIR from the disclosed figures, and graded by disclosure quality. None is an observed price. A standing version of these tables is maintained on the AI commitments monitor.
01 The print inside the lump sum
The committed-term market shows itself in pieces: posted asks (our committed-term curve), financial forwards, reserved-block venues, and the lump-sum headlines. The mega-deals are the largest committed-term prints in the market. Their disclosure format states the total while leaving the unit rate implicit. When a company discloses dollars, term, and quantity for the same contract, the implied unit rate is calculation, not estimation. Where one input comes from press rather than the company, the grade says so. Where an input is missing, the rate is not computed.
This analysis lives on the research and credit surfaces. It is not part of any CCIR benchmark series, and no figure here feeds one.
02 Method: assumptions and grades
Spend is ratable across the full disclosed term. 100% of contracted hours are billed. All disclosed dollars are treated as compute, with no carve-out for power, networking, or services. Fleet size is constant over the term. 8,760 hours per year; 730 hours per month. Every one of these assumptions biases the implied rate downward. All figures are therefore floors on blended committed rates, not observed prices.
Why floors: real contracts ramp, so fewer early hours stand against the same dollars. Disclosed totals often include non-compute scope. And no fleet bills every hour. Each correction would raise the implied rate.
Chip attribution. The chip column records the generation as the parties disclosed it, or "per press" or "unspecified" where they did not. Where the generation is unspecified, the rate is fleet-blended. No performance adjustment is applied across generations: a GPU-hour is the unit, not a unit of compute output, so rates on different chips are different products, not comparable prices. Megawatt-to-GPU conversions use NVIDIA's published rack power and are always labeled as CCIR estimates.
Grades. A: dollars, term, and units all company-disclosed. B: one input from attributed press. C: talks-stage or heavy estimation. Grade-C rows do not publish, here or on the standing table. Dollar-per-GPU-hour and dollar-per-megawatt rows are separate tables and are never blended or averaged. TPU commitments are excluded entirely.
03 The GPU-hour table
| Deal | Chip | Disclosed | Implied rate | Grade |
|---|---|---|---|---|
| Microsoft–Nscale/Aker (Narvik) | GB300 disclosed | $6.2B · 5 yr · ~52,000 GPUs | $2.72/GPU-hr | A |
| Anthropic–xAI (SpaceX) | unspecified NVIDIA, no model filed | $1.25B/month through May 2029 · ~325,000 GPUs | $5.27/GPU-hr | A |
| Microsoft–IREN | GB300 disclosed | $9.7B · 5 yr · 200MW critical IT load | ~$2.2/GPU-hr CCIR estimate | AMW basis |
| Microsoft–Nebius | GB300 per press | $17.4B (up to $19.4B) · 5 yr · ">100,000" GPUs (press) | $3.97/GPU-hr ($4.43 at $19.4B) | B |
| OpenAI–AWS | GB200/GB300 disclosed | $38B · 7 yr · "hundreds of thousands" of GPUs + CPUs | $1.24–$3.10/GPU-hr (band) | B |
All rows: implied — computed by CCIR from the disclosed figures; floors per §02.
Narvik leads because it is the cleanest print in public: dollars, term, and GPU count all disclosed by the parties in a single release. $6.2B ÷ (52,000 × 8,760 × 5) = $2.72. Nothing estimated.
Anthropic–xAI is the fee-basis print. SpaceX's prospectus discloses the terms verbatim: "the customer has agreed to pay us $1.25 billion per month through May 2029," for compute capacity that "includes approximately 325,000 NVIDIA GPUs" across the COLOSSUS facilities. A monthly fee needs no term assumption: $1.25B ÷ (325,000 × 730 hours) = $5.27. Three caveats travel with the row. Either party may terminate on 90 days' notice after an initial three-month period, so the contract prices like a short-tenor rate wearing a 2029 end date. The filing names no GPU generation, so the row is not part of the GB300 set below and no spread against it is computed. And capacity ramped through May and June 2026 at a reduced fee.
IREN is graded A on the megawatt basis (§04); the GPU figure is a labeled CCIR estimate. The filing discloses 200MW of critical IT load, not a GPU count. Applying NVIDIA's published GB300 NVL72 rack figure, "full rack requiring up to 142 kW", yields roughly 101,000 GPUs as an upper bound (rack power is not critical IT load; facility IT power also carries network and storage). So the ~$2.2 rate is a floor even within the estimate.
The AWS band is a band, not a point. The companies said "hundreds of thousands" of NVIDIA GPUs plus "tens of millions" of CPUs. CCIR parameterizes their own phrase: 500,000 GPUs implies $1.24, and 200,000 implies $3.10. The CPU scope means even the band is a floor on the GPU-attributable rate.
The only identical-unit comparison in public: Narvik $2.72, IREN ~$2.2 (estimate), Nebius $3.97: all GB300-class, all five-year terms, all Microsoft as buyer. Same buyer, same silicon generation, same tenor. The residual spread is jurisdiction, scope, and counterparty.
Chart notes
GB300-class, ~5-year terms, Microsoft as buyer. Implied: computed by CCIR from the disclosed figures under the §02 assumptions, floors by construction. IREN bar is a labeled CCIR estimate from disclosed megawatts and NVIDIA's published rack power.
04 The megawatt table
| Deal | Chip | Disclosed | Implied rate | Grade |
|---|---|---|---|---|
| Microsoft–IREN | GB300 disclosed | $9.7B · 5 yr · 200MW critical IT load | $808,000/MW-month | A |
| OpenAI–Oracle (Stargate) | n/a power basis | >$300B · 5 yr · up to 4.5GW | $1.11M/MW-month | A |
All rows: implied — computed by CCIR from the disclosed figures; floors per §02.
No public conversion from gigawatts to GPU count exists for the Stargate agreement, so CCIR declines to state one. The megawatt row is the computable read. One definitional caution: IREN's denominator is disclosed critical IT load, while Stargate's "up to 4.5GW" carries no IT-versus-facility breakdown. The same "megawatt" is not necessarily the same denominator, which is why these two rows are presented, not spread-differenced.
Chart notes
Rows differ in chip generation, scope, and tenor. The chart presents them side by side and computes no spreads. The AWS entry is a parameterized band. The xAI entry is terminable on 90 days' notice after an initial three-month period, with GPU models unspecified in the filing.
05 Read against the credit record
Every GPU-backed credit prices a revenue assumption its lender cannot observe. KBRA's AI-compute framework builds on an assumed relationship between old and new compute prices. Fitch's analyses of GPU-backed paper carry renewal-rate assumptions the agency itself does not call observable. The re-leasing question (what does committed compute rent for when the anchor contract rolls?) is one those frameworks answer by assumption.
The implied rates above are a public anchor for that assumption: actual anchor-tenant contracts, reduced to unit rates, graded by disclosure quality, floored by construction. The IREN row is literal: the $9.7 billion Microsoft contract is the collateral cash flow behind whatever IREN finances next, as the Nebius and CoreWeave paper already on the compute-credit record shows in filed form.
No solvency opinion and no forecast follows from any of this. What follows is a table: CCIR maintains an implied committed rates section on the AI commitments monitor, with grade A and B rows only, the §02 assumptions block permanent, and the grade column always shown, updated as deals publish.
Sources: Aker ASA–Nscale JV release (Sep 17, 2025); SpaceX final prospectus, Form 424B4 (acc. 0001628280-26-042639, Jun 12, 2026); IREN Form 8-K Ex-99.1 (Nov 3, 2025); Nebius Form 6-K (Sep 8, 2025) with GPU count as attributed press (DCD, Reuters); OpenAI–AWS joint releases (Nov 3, 2025); OpenAI, "Five new Stargate sites" (Sep 23, 2025) and OpenAI capacity statements (Jul 2025); CoreWeave Forms 8-K (Sep 30, 2025; acc. 0001769628-26-000154, Apr 9, 2026); NVIDIA Enterprise Reference Architecture, NVL72 AI Factory components (rack power input for the labeled IREN estimate); attributed press: The New York Times (Meta–Anthropic talks). All implied rates are CCIR calculation on the cited disclosures under the §02 assumptions. They are floors on blended committed rates, not observed prices, and are not part of any CCIR benchmark series. Nothing here is a solvency opinion, a forecast, or investment advice. CCIR publishes reference data and records.