The instinct when you compare GPU providers is to pay up for the better-capitalised name and treat the cheapest quoted rate as a bargain, and both halves deserve testing. Neoclouds like CoreWeave, Lambda and Nebius undercut AWS, Azure and Google Cloud on-demand pricing by roughly 40 to 70 per cent for identical hardware. That gap mixes credit risk with reliability, compliance, enterprise support and lock-in, so no clean credit premium is ever quoted. No rate card states the number you need: the price of the counterparty still being solvent when your term ends. This article separates the service bundle, the compute risk and the credit spread, so you can compare providers on more than price.
CoreWeave versus a hyperscaler: which carries more counterparty risk for a mid-size buyer?
Counterparty risk sorts along four axes: debt against contracted cash flow, revenue concentration, contract length, and whether the entity signing your contract actually carries the debt. Run that check and CoreWeave comes out riskier than a hyperscaler despite a $100 billion-plus backlog, because your contract becomes a line item in someone’s borrowing base.
CoreWeave has taken on more than $51 billion of debt and obligations, and its five-year CDS has risen past 1,140 basis points before settling near 1,038bp, a level S&P Global Market Intelligence classes as distress. Two-thirds of that backlog sits with investment-grade names like Microsoft, Meta, IBM and Caterpillar, and Microsoft alone was about 67 per cent of revenue. A hyperscaler’s capacity is a cost centre; a neocloud’s capacity is collateral.
Monetisable MW, revenue per constrained megawatt a year, makes the point better than the headline backlog. Powered-shell landlords monetise about $1.2 million to $2.3 million per MW a year and hand GPU obsolescence to the tenant, while full-stack neoclouds monetise $7 million to $13 million per MW, a roughly five-fold spread on the same megawatt. CoreWeave comes in near $18.5 million per MW on active power but compresses to about $5.3 million across contracted power, a gap that tracks how much value rests on the operator staying solvent.
Who owns the silicon decides who holds the depreciation. In a powered-shell deal the tenant owns the GPUs, leaving you a resale problem; in a full-stack deal the operator owns them and prices depreciation into the rate, leaving you a counterparty. That is the fourth axis at work: who actually owns the GPUs you are renting matters more than the name on the contract. Claimed utilisation cannot be verified from outside, so revenue per MW is the closest proxy.
Scale changes the shape of counterparty risk without removing it. Oracle keeps an investment-grade rating while its five-year CDS runs around 282 basis points and its $18 billion Jupiter project debt trades at 89 to 91 cents after force majeure. Nscale is the concentration case: 85 per cent of its $103.4 billion book sits with Microsoft and Anthropic, and Microsoft walked away from a UK deal over power delays.
The hyperscaler side carries its own premium. Part of the higher rate is a service bundle: reliability, compliance, enterprise support and lock-in, none of which the bare neocloud quote includes. The rest is the credit question. Before you sign, test solvency the way a bondholder would.
Reserved multi-year commitment versus on-demand: which risk profile fits a 200-person SaaS company?
Once you accept that the risk lives in the financing structure, the next question is the shape of the exposure you carry. The two exposures are different in kind: a multi-year reservation locks credit and price risk for the whole term, while on-demand pricing is a price risk you renew every hour, with the credit risk sitting on the provider’s own balance sheet. A reservation typically cuts 25 to 50 per cent off on-demand.
For a smaller SaaS team the swap cost usually binds harder than the rate. Migrating a production inference service can run to about a quarter of a team’s time, and a three-year commitment locks you to hardware two generations old at exit. Used H100s that sold for $40,000 in late 2023 traded at $6,000 to $22,000 by mid-2026, an 85 per cent decline in under three years. The middle ground is short reservations with renewal rights, rolling capacity and on-demand burst, so no single provider carries the whole path.
A leveraged GPU cloud concentrates hardware, debt and customer contracts in one balance sheet, so one refinancing failure can take down the whole inference path. A decentralised pool of independent operators spreads hardware across many small owners with no long contract: worse credit risk per operator, lower correlation, and residual value staying with whoever chose the hardware.
If you bill in AUD against a USD commitment, the reservation is also a currency bet. Check the three clocks line up before committing: customer contract, asset useful life and financing term. And optionality is the alternative to clauses, because shortening the term achieves more than any enforcement language.
How much of a premium is a better-capitalised provider worth, and what is the catch in a cheap rate?
That framing sets up the number in the title: what the safer name is worth. Start with the arithmetic. The premium is the annual rate differential multiplied by committed megawatt-years; the expected loss is the probability of failure over the term multiplied by migration and downtime cost. You pay the premium where the second exceeds the first. As a working rule, a 15 to 25 per cent premium on a three-year commitment buys a migration contingency you would otherwise self-fund.
Benchmark any multi-year quote against the forward curve. Silicon Data’s daily H100 index, its 36-month curve, and cash-settled H100 and B200 rental futures on ICE and CME test a multi-year quote against year-three prices. A rate below the year-three forward means the provider expects to be cheaper than the market later, an efficiency claim or a subsidy, and subsidies here usually come from the financing structure. So benchmark the rate the way the futures market does.
That is the catch in cheap multi-year rates. A low rate is often a provider buying backlog to support a borrowing base, which turns your commitment into collateral and makes the discount a function of your own credit support. The rate is also a residual-value bet placed on your behalf, untested through a full depreciation cycle. Cheap rates exist because Goldman Sachs counts $489 billion of 2026 AI-related debt issuance plus roughly $3 trillion in off-balance-sheet commitments.
The credit spread is not confined to neoclouds. NVIDIA has guaranteed up to $108.5 billion of leases that ratings agencies count as nothing, protection aimed at the financing rather than your service level. Oracle’s BBB-rated debt trades like junk too. As of this writing, no major neocloud has defaulted, so the premium is priced against a scenario the market has not tested.
The premium question comes down to arithmetic, and the premium itself is narrower than most buyers assume: a lower probability that your contract becomes a line item in an enforcement. Read the rate card as a financing signal rather than a price, and start with the credit evidence behind the choice.
Frequently Asked Questions
Will my GPU provider actually go bust before my contract ends?
No one can put a number on a single provider, but the risk is not evenly spread. The failures to watch for are refinancing events, not demand collapse: a provider whose debt matures inside your term and whose contracted revenue will not cover the refinancing is the one that breaks. Track maturities, customer concentration and revenue per megawatt, and treat a near-term refinancing wall as the signal that matters.
If a GPU provider has no traded bonds or CDS, how do I judge its credit risk?
Work from the closest available proxies. Ask for debt maturity dates, the lenders’ covenants, and the share of revenue coming from your three largest co-tenants. Revenue per contracted megawatt against the debt secured on that capacity tells you how much slack exists, and a provider that will not disclose any of it has told you something too. Treat private credit marks and vendor financing terms as substitutes for a public spread.
Is uptime the same thing as solvency?
No. Uptime measures whether the service is running today; solvency measures whether the operator can keep it running for the rest of your term. A provider can post excellent availability for two years and still fail at a refinancing, and the outage that follows is not a service incident but an enforcement. Ask separately about operational track record and balance sheet durability, and weight the second by the length of the commitment.
What contract terms actually reduce counterparty risk?
The useful ones are assignment and step-in rights, a data and model export obligation on a fixed schedule, and the right to terminate early without penalty if a defined credit event occurs. Escrow of configuration and pipeline artefacts matters more than a service credit. Note that shortening the term achieves more than any clause, because a clause is only worth what an administrator chooses to honour once enforcement starts.
Is a one-year reservation always safer than a three-year one?
Safer on credit, not always safer overall. A short term reduces the exposure you carry but hands the provider the renewal conversation, and renewal pricing on scarce capacity is where the saving disappears. Compare total cost across the workload life, including the renewal rate you would plausibly face, not the headline rate on the term you are signing. Match the term to how far ahead you can forecast the workload.
What does it really cost to migrate if I have to leave mid-term?
Engineering time is the smallest line. Plan for a quarter of a team’s time to re-platform a production inference service, then add egress on stored data, re-qualification of model versions against new hardware, and the revenue lost during cutover. Data gravity compounds it: the longer a pipeline runs in one place, the more of the cost sits in things nobody documented. Estimate it before signing, not after.
What happens to my data and model weights if the provider fails?
Usually nothing immediately, but access becomes the problem. In an administration the estate controls access to the systems your checkpoints sit on, and retrieval becomes a negotiation rather than a right. Keep an independent copy of weights and training data outside the provider, export on a schedule, and confirm in writing that your data is not pledged as collateral.
Should an Australian buyer hedge the USD exposure on a multi-year reservation?
Yes, or price it in and accept it consciously. A three-year USD commitment against AUD revenue adds a currency exposure on top of the credit exposure, and the two can move together in a risk-off market. A layered hedge covering the first twelve months and rolled forward is usually enough; full-term hedging is expensive and often locks in a rate worse than the risk it removes.
At what spend does this analysis become worth the effort?
Roughly where the annual rate differential exceeds a week of engineering time. Below that, the cheaper provider wins on arithmetic almost regardless of credit quality, so sign short, keep the workload portable and revisit. Above it, the expected-loss calculation and the diligence behind it pay for themselves quickly, particularly when the commitment runs past the provider’s next debt maturity.
Does my contract survive if the provider is acquired or restructured?
It survives on paper and weakens in practice. Assignment clauses decide whether the contract can transfer without consent, so check whether a change of control requires your approval. A restructuring usually brings cost cuts before service improvements, and a new owner may reprice renewals or retire the older hardware your workload depends on. Get consent rights and a termination trigger written in before you sign.
Are all cheap multi-year rates a warning sign?
No, and treating every discount as a subsidy will cost you real money. Some providers are genuinely cheaper because they own the shell, hold long power contracts at fixed prices, or run newer silicon that needs fewer megawatts per unit of throughput. The test is whether the rate sits below the year-three forward: above it you are buying efficiency, well below it you should ask who is funding the difference and why.
