GPU-hour economics: what a GPU earns and costs

A GB300-class megawatt costs roughly $45–55m all-in to stand up (author estimate, Jul 2026). Sold on a 2025-vintage long-dated hyperscaler contract — IREN–Microsoft works out at ~$9.7m per MW per year — that implies a ~4–5-year gross payback before opex and interest. Sold at 2026 terms of ~$20–25m per MW per year (Nebius's four large Q2 deals; IREN's recent three-year contracts), gross payback falls to roughly two to two-and-a-half years; Nebius puts payback on its Q2 deals at one year ten months, helped by prepayments that cover 50–60% of the associated capex. Short-dated (3–6-month) capacity has cleared at ~$40m per MW at CoreWeave and $40–50m at Nebius. Adjusted-EBITDA margins flatter the picture because depreciation is the business: the true economic question is whether contracted cash flows repay the silicon before the next architecture reprices it, and what the asset earns when it is re-contracted. This is why contract cover matters more than any other metric, and why the market treated CoreWeave's March 2026 milestone — the first investment-grade-rated, $8.5bn GPU-backed loan — as a category event: it moved the marginal cost of capital for the whole sector.

The supply arc continues regardless. Blackwell Ultra (GB300) drove NVIDIA's Q2 FY2027 ramp, and Vera Rubin began production shipments in August 2026, with racks running at CoreWeave, Google Cloud, Azure, OCI and Nebius. Operators must therefore decide whether to buy the end of one curve or the start of the next — and every Rubin rack eventually lowers the clearing price of the H100 hour.

Competitive structure is also shifting above the neoclouds. On 1 July 2026 Meta launched Meta Compute, renting slices of its own build-out — a signal that hyperscalers may become sellers as well as buyers of the same capacity; CoreWeave and Nebius fell 13–15% on the announcement. The bull read: it validates the rental model, and Meta still leases from neoclouds where it lacks sites. The bear read: the largest buyers can toggle to competitors whenever internal demand dips, which caps long-run pricing power for undifferentiated capacity. Either way, the episode reinforced the sector's core lesson — durable value accrues to contracted power, deployment speed and operational excellence, not to merely owning GPUs.

A cluster's commercial output is not its nameplate FLOPs. Two ratios describe reality. Goodput is the share of wall-clock time that does useful, non-repeated work: what survives after failures, restarts, checkpoint overhead and scheduling gaps. Model FLOPs Utilisation (MFU) is stricter. It is the share of theoretical peak FLOPs that becomes actual model mathematics, additionally discounting communication stalls, memory-bandwidth waits and imperfect kernels. World-class frontier training achieves ~90%+ goodput and ~38–43% MFU on Hopper-class BF16. Even excellent operators therefore convert well under half of the silicon they bought into sellable mathematics. Closing that gap is what operators sell.

Source: Compute Terminal illustrative decomposition (author estimate), calibrated to published frontier runs (Meta Llama 3 405B: BF16 MFU 38–43%, >90% effective training time on 16,384 H100s). Larger NVLink domains and FP4/FP8 shift achievable MFU on Blackwell-class systems.