How do you vet a GPU cloud provider before you wire the deposit?

A deposit on a GPU cluster or a reserved compute contract is a bet on a counterparty, not just a bet on silicon. The GPUs themselves are a commodity you can specify down to the part number. The provider standing behind them is not. Before the money moves, a buyer who cannot read a rack diagram can still ask six questions that separate a provider who can deliver and stay solvent for the term from one who cannot, and every one of them has a public, checkable answer or a refusal to answer that tells you something on its own.

Who owns the metal

Start with the simplest question: is the entity you are contracting with the same entity that owns the GPUs, or is it reselling capacity it leases or subleases from someone else? A broker layer is not automatically bad, but it adds a counterparty you cannot see and a margin stacked on top of the underlying deal. Ask directly whose balance sheet the hardware sits on, who holds title if the provider defaults, and who signs the power and colocation contracts underneath the compute you are buying. If the answer routes through more than one company before it reaches the actual data center operator, get each layer’s name and get comfortable with all of them, not just the one on your invoice.

Power, priced and dated

Power, not GPU supply, is the constraint that is actually binding capacity growth right now. Goldman Sachs Research, publishing May 20, 2026, forecasts US data center power demand more than doubling from 31 gigawatts in 2025 to 66 gigawatts in 2027, with 41 gigawatts expected in 2026.

That demand is running into a grid that cannot turn new capacity around quickly. Lawrence Berkeley National Laboratory’s “Queued Up: 2026 Edition,” covering the interconnection queue as of the end of 2025, counted roughly 8,200 projects seeking transmission interconnection, representing about 1,312 gigawatts of generation and about 749 gigawatts of storage, with a typical wait of around five years from initial application to commercial operation, and longer in the busiest zones. A provider without a signed, dated power or interconnection agreement is asking you to underwrite that queue on their behalf, on your timeline.

Ask for the contracted power figure in megawatts, the cost per kilowatt-hour, and the facility’s PUE, all dated and in writing, not described verbally on a call. For context on what a normal PUE looks like: the Uptime Institute’s 2024 Global Data Center Survey put the industry average PUE at 1.56, flat for a fifth consecutive year, down from above 2.5 in 2007 but no longer improving much across the industry as a whole. A number well above that on a facility built after 2020 is worth asking about directly.

How the hardware is financed

GPUs depreciate fast and providers borrow against them, and the two facts together are the solvency risk in this industry. CoreWeave’s own August 3, 2023 announcement disclosed a $2.3 billion debt facility, led by Magnetar Capital and Blackstone, collateralized by Nvidia H100 chips, one of the first uses of GPUs themselves as loan collateral at that scale. In a press release dated March 31, 2026, CoreWeave announced it had closed an $8.5 billion delayed draw term loan facility that it says received ratings of A3 from Moody’s and A (low) from DBRS, the first investment-grade rating the company reports for a financing secured by HPC infrastructure and an associated customer contract, priced with a floating-rate tranche at SOFR plus 2.25 percent and a fixed-rate tranche at approximately 5.9 percent.

Both facilities are public and both are collateralized by the same kind of hardware. The difference between them, roughly two and a half years apart, is what a lender was willing to charge against that collateral once the borrower had a longer track record and stronger contracted revenue behind it. Ask your prospective provider the equivalent questions: how much of their fleet is debt-financed, what the loan-to-value looks like against current hardware value rather than purchase price, and what happens to your capacity if a lender forecloses on the collateral mid-term. A provider heavily levered against fast-depreciating GPUs, with thin contracted revenue behind the loan, is a weaker credit than the marketing page suggests.

Track record, at this specific thing

Ask how long the company has operated GPU infrastructure specifically, not compute infrastructure in general, and ask for two or three referenceable customers at or above the scale you are buying who will actually take a call. A provider that pivoted into AI hosting from a different business is not automatically a red flag, but it changes what you should verify. Core Scientific, a Bitcoin miner that started hosting infrastructure in 2019, filed for Chapter 11 bankruptcy protection on December 21, 2022, citing falling crypto prices and rising energy costs, then emerged from bankruptcy in January 2024 and pivoted toward AI data center hosting. The pivot itself is not disqualifying, plenty of legitimate operators have repurposed power and space this way, but a track record that includes a bankruptcy is a fact to price in, not to skip past because the pitch deck only covers the years after the restructuring.

Concentration, in both directions

Ask what share of the provider’s revenue comes from its largest customer, and ask what share of your own compute needs this provider would represent for you. Both directions carry risk. CoreWeave’s own S-1 registration statement, filed with the SEC on March 3, 2025, disclosed that Microsoft accounted for 35 percent of its revenue in 2023 and 62 percent in 2024. That is public information the company chose, or was required, to put in its own filing, not a rumor, and it is exactly the kind of number a diligent counterparty should be willing to share or let you estimate from public filings if they are already a reporting company. If a provider will not discuss customer concentration at all, that reluctance is itself an answer. The same logic runs the other way: if you would be a large share of their book, ask what happens to your pricing and priority if a larger customer shows up next quarter.

What happens if they go under

Get this in writing before you sign, not after you need it. Bankruptcy does not automatically mean the lights go out. Cyxtera, a colocation and interconnection provider operating more than 60 data centers across more than 30 markets, filed a pre-arranged Chapter 11 in June 2023 in the District of New Jersey, backed by roughly $200 million in debtor-in-possession financing from lenders holding more than two-thirds of its term loan debt, and it continued operating all of its facilities without interruption through the process. That outcome was not automatic, it was the product of a pre-arranged filing with lender support lined up in advance.

Ask, specifically: does your contract survive assignment to a buyer or a lender who takes over the business? Is there a data escrow or export right if the provider stops operating? Who has a lien on the GPUs you are paying to use, and does your service continue if that lienholder forecloses? A provider that has thought about this will have a paragraph for it. A provider that has not is asking you to find out during the bankruptcy, which is the worst possible time to be asking for the first time.

The checklist, in the order to ask

1. Is the entity on my contract the entity that owns the hardware, and if not, who is above them and what do their contracts say.
2. What is the contracted power figure in megawatts, the price per kilowatt-hour, and the facility’s PUE, all dated in writing.
3. How much of the fleet is debt-financed, against what loan-to-value, and what happens to my capacity if a lender forecloses.
4. How long has this company operated GPU infrastructure specifically, and can I call two or three referenceable customers at my scale.
5. What share of their revenue comes from their largest customer, and what share of my needs would this provider represent for me.
6. What does the contract say happens to my capacity and my data if the provider files for bankruptcy or is acquired.

None of this requires being able to read a network diagram. It requires being willing to ask the sixth question before you wire the deposit, not after you need the answer. For what to check once the racks are actually delivered, see how to tell if a GPU cluster actually works; for the arithmetic behind what that cluster should cost to run once it passes those tests, see what it actually costs to run a GPU cluster.