Take-or-pay compute contracts

The industry's commercial core is the take-or-pay reserved-capacity contract: a customer commits to pay for a defined block of GPUs (or MW) for two to five-plus years, whether or not it uses them. CoreWeave reports ~98% of its revenue on committed take-or-pay terms and a backlog of ~$104bn at Q2 2026, with its top three customers accounting for 36%, 26% and 10% of Q2 revenue. Nebius describes ~$40bn of committed backlog from investment-grade customers that it can borrow against, and Oracle's remaining performance obligations reached ~$638bn — more than half attributable to a single customer, OpenAI (Jul 2026). Contract structures increasingly include prepayments that part-fund the capex: Nebius expects more than $9bn of customer prepayments in 2026, and 70% of its Q2 deals carried prepayments covering 50–60% of associated capex. Third parties now stand behind the contracts too: Google's lease guarantees behind ~$3.2bn of TeraWulf and Cipher/Fluidstack obligations (Jul 2026); Broadcom's backstop of more than 80% of a $35bn financing structure for Anthropic compute capacity (Reuters, 1 October 2026); and NVIDIA's six-year cloud-service agreements with neoclouds, which totalled $36bn at 26 July 2026 and effectively give operators a revenue floor. On-demand and spot capacity clears at the margin and sets the visible price, but it is a minority of revenue for every scaled operator.

These contracts are the collateral of the sector's entire credit stack: lenders underwrite the counterparty's obligation to pay, not the GPUs. That is why the composition of a backlog — investment-grade share, tenor, concentration, prepayment mechanics, termination rights — has become the single most scrutinised disclosure in the sector. The 2026 deal prints make the point directly: within CoreWeave alone, a single strong-counterparty facility priced at SOFR+225 (A3) in March, while a pool of shorter lab and trading-firm contracts priced at SOFR+550 at 97 (Ba2/BB+) in August.

The pivotal design choice of the modern GPU loan is that lenders and rating agencies underwrite the counterparty, not the silicon. In the template transaction, CoreWeave's $8.5bn DDTL 4.0, a bankruptcy-remote SPV owns the GPUs and holds an assignment of the receivables from a take-or-pay contract with an investment-grade customer (reported as Meta). Debt service is swept from contract cash flows through a waterfall, and the GPUs serve only as a recovery floor. The loan is rated A3/A(low) because the customer's obligation to pay is A-grade, not because anyone believes five-year-old accelerators keep investment-grade value. Lenders use 3–4-year GPU depreciation in their underwriting (Forbes, 14 Oct 2026).

The same issuer shows how far contract quality moves the price. In March CoreWeave paid SOFR+225 for a single strong-contract SPV. In August it paid SOFR+550 at an issue price of 97 for DDTL 5.5, a pool of contracts with AI labs and trading firms averaging about three years against a roughly five-year loan, with a 1.35× DSCR test. The deal had been talked at SOFR+425–450 at 99 and widened before pricing. That is a gap of about 325bp before the discount, or about 385bp with the 3-point discount spread over five years (author calculation). Across the market, contract-anchored investment-grade GPU debt priced at SOFR+213–300 (about 5.9–6.0% fixed) before the hike, against SOFR+450–550 for BB contract-backed debt: a gap of about 250–325bp. On tens of billions of dollars, that spread is the largest single competitive variable in the industry.

The simplest test of a GPU borrower is contract cover: contracted backlog divided by debt. It is crude, because backlog is gross revenue over five to twenty years while debt is principal, but it shows direction. On our arithmetic, CoreWeave's cover fell to about 2.9× at Q2 2026 ($104bn backlog against about $35.6bn of debt principal), from about 4.0× at Q1. Between Q1 and Q2 it added about $10.5bn of principal against about $4.6bn of new backlog. Hut 8's River Bend bonds are covered about 2.2× by the $7bn base lease, or 5.4× including options. Applied Digital sits at about 3.5×, and Riot, with a ~$9.1bn, 20-year lab lease and little debt, above 15× (all author calculations from company disclosures). Miner-converters with 15–20-year leases show high gross cover. Their real risk is construction and delivery, which is what Galaxy's 9.875% coupon prices.

The sector's maturity wall is not yet measurable from public data. We could not trace widely repeated figures for total drawn debt and 2029–31 maturities to a primary source, and we do not use them. A partial rebuild from 2026-vintage instruments alone shows the heaviest buckets in 2031 (about $10.8bn) and 2032 (about $10.6bn) (author calculation). It excludes CoreWeave's pre-2026 facilities and 2025 miner notes. Many facilities also amortise from contracted cash flows (DDTL 4.0, Lambda's term loan, Galaxy's 4% a year, Oracle Michigan's 13-year amortisation), so final maturity overstates refinancing risk. The true balloon is what remains after contract amortisation. It matters most where loan tenor exceeds contract tenor, as in DDTL 5.5. A reliable schedule needs each issuer's 10-Q debt note.

Buyers have learned to write physics into SLAs. Modern capacity contracts specify delivered NCCL all-reduce bandwidth at defined message sizes and scales, sustained tokens per second on reference models, ceilings on interruption rates and time-to-replace guarantees, alongside classical uptime. Operators publish goodput dashboards, and sophisticated customers audit them. Two clusters with identical hardware can differ by 20–30% in useful output. That difference comes from reliability engineering, scheduling and network tuning, and it is the durable, software-shaped moat in a business whose hardware anyone can eventually buy. Its value has risen with contract pricing: at ~$20–25m per MW per year, a 10-point goodput gap on a 100 MW cluster is worth roughly $200–250m of annual revenue (author calculation).

Table 8. What buyers now write into capacity SLAs

Source: composite of market practice reported by operators and large buyers (Jul 2026); exact numbers are deal-specific and not public.