Price a Neocloud GPU Commitment: Cost per Token, Delivery Risk, Credit Backstops
Treat a multi-year GPU capacity lease as financed capacity: verify power, hardware, delivery, credit, and token economics before signature.
Power and hardware set the usable capacity
You are reviewing a multi-year GPU capacity quote. The trade-off is that you pay for energized IT load, delivery timing, and counterparty credit, not for a hardware purchase. The Tydal, Norway data center is expected to supply 133 MW of gross AI compute capacity, or about 121 MW of IT load. A debt-financed GPU cluster reaches break-even around 70% utilization; a 1,024-GPU H100 cluster moves from -$330,000 to +$340,000 monthly P&L when utilization rises from 55% to 85%.
Gross megawatts include cooling, conversion, and distribution losses; IT load is what runs accelerators. If the quote uses gross capacity, request the IT-load figure and the conversion assumption. Also request the power contract and the metering plan. That gap is the overhead you are underwriting.
Treat hardware allocation as a dated delivery claim. Neoclouds compete for early allocations of newer Nvidia generations, including Blackwell and Rubin, while Hopper-class GPUs shift to cheaper inference work. Volta's formal Nvidia Cloud Partner status gives it preferential access to Nvidia GPU allocation. Bitdeer Technologies will operate the Tydal, Norway site on renewable hydropower and configure it for Nvidia's Vera Rubin hardware; Dell is reported as the technology provider. Chip generation, operator, and power source set the date tokens can be produced.
Signed backlog is not energized capacity. CoreWeave's $99B backlog is signed before capacity exists, making energized megawatts the binding constraint. A 12-month energization gain can shift about $8M of unrecovered depreciation into revenue on a single cluster. Delivery risk is a cash-flow item that belongs in the term sheet.
Delivery and credit decide whether it pays
Neoclouds finance multi-year GPU capacity by turning 2–5-year take-or-pay contracts into debt and spreading repayment over the term. The buyer locks capacity for the term and pays even if it does not use the capacity. Your price must cover debt service, power, hardware depreciation, operations, and delivery slip. Debt service is the floor under your token price.
Anthropic signed a six-year, $10 billion compute agreement with Volta Infra Holdings. Volta started in January 2026 and closed a $300 million round at a $2.4 billion valuation on August 4, 2026. Delivery is scheduled in two phases: start December 31, 2026, complete by March 31, 2027. Bitdeer is contracted for about $4.7 billion in revenue under a 16-year colocation lease with Volta.
Credit support is the exit. Nvidia can structure neocloud deals around GPU sales, financing or guarantees, and a return tied to rental performance. In March 2026, CoreWeave's $8.5B DDTL 4.0 debt facility got Moody's A3 at SOFR+2.25%, making it the first GPU-backed financing to reach investment grade.
Build the token model from utilization, not list price. If the counterparty cannot show how contracted utilization supports break-even, do not sign the token price.
The checklist is the underwriting test
Run each item before signature, not after. Do it in writing. Each one is testable if the counterparty has the data ready. The SLA/termination item is there because late energization is the failure mode that turns a capacity deal into a financing problem. A counterparty that can answer all items in writing is showing a bankable structure; one that cannot is asking you to underwrite missing data.
- Power-to-IT load: ask for gross MW and IT MW; model only IT MW.
- Hardware allocation: confirm chip generation, partner status, and dated allocation plan.
- Lease revenue: identify the upstream colocation lease and its contracted revenue.
- Credit backstop: identify financing, guarantees, or investment-grade debt support.
- Delivery milestones: require dated energization phases, not target quarters.
- SLA/termination: confirm cure periods, termination rights, and liquidated damages for late energization.
- Cost-per-token model: test utilization, break-even, and monthly P&L at low and high load.