From $/kW-month to $/GPU-hour

Source: Compute Terminal analysis of company disclosures, October 2026.

Full-stack AI clouds (CoreWeave, Nebius, Crusoe Cloud, Lambda) own or lease the facility and the IT, and wrap it in managed Kubernetes/Slurm, storage, observability and inference services. They capture the highest $/GPU-hour and the stickiest workloads, at the cost of carrying both facility and silicon risk. Bare-metal providers (IREN's cloud segment, Voltage Park, Fluidstack's deployment model) sell dedicated clusters with a thin software layer — faster to stand up, lower price points, customers bring their own stack. Power landlords (Core Scientific, Applied Digital, Cipher, TeraWulf, CleanSpark) monetise the scarcest layer — energised shells — through 10–20-year leases priced in $/kW-month, taking real-estate economics and no GPU obsolescence risk. CleanSpark's 20-year, $6.6bn, 175 MW lease in Georgia (ready late 2027) works out at roughly $1.9m per MW per year, or ~$157 per kW-month (author arithmetic) — an order of magnitude below GPU-capacity contracts, because the tenant buys the hardware. The 2025–26 wave of hyperscaler leases shows how much value has migrated to whoever holds the electrons.

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.

• Revenue per MW, restated in tokens. At ~700 H100-class GPUs per MW of IT load (author estimate), 1,000–3,000 output tokens per second per GPU equates to roughly 22–66 trillion tokens per MW-year (author arithmetic) — the "nameplate output" of the factory. A 2026 contract at ~$20–25m per MW per year therefore prices that output at very roughly $0.3–1.1 per million tokens of capacity, before utilisation.
• Training and inference monetise differently. Training is sold as reserved capacity; inference is increasingly sold as tokens or as dedicated always-on halls. The same physical asset can serve both, but contract shape, network design and siting optimum differ.
• Token deflation is structural; GPU-hour pricing is cyclical. Each hardware generation cuts cost per token faster than it cuts $/GPU-hour, and multi-vendor silicon and hyperscaler surplus resale add downward pressure. 2026 showed that GPU-hour prices can rise for a year or more when demand outruns deliveries; underwriting should still assume token prices fall every year the asset operates.