When you sign a multi-year GPU reservation you are buying uptime. The contract is priced as years of guaranteed capacity at a fixed rate, and the risk you are told you are accepting is hardware failure or a service-level slip. That is the wrong risk to have at the top of your list.
The binding constraint on uptime sits in the provider’s credit agreement. A debt maturity arrives on a fixed calendar date. Neocloud facilities run about five years while the pledged customer contracts average closer to three, and the gap is refilled by future demand at future prices. Here is the maturity wall in context. By the end you can size a commitment so you are not the anchor whose departure destabilises the financing, and run a solvency read before signing.
Why do neocloud loans run about five years when the pledged customer contracts average closer to three?
Lenders underwrite the multi-year useful life of the hardware, so they want a loan long enough to amortise the asset. Customers refuse to commit beyond three years at a fixed price against silicon that changes every year. The facility is written long and secured short, and the gap is refilled by renewals at prices nobody knows yet.
Put plainly: the loan is longer than the contract. Lenders size facilities to expected contracted revenue rather than the neocloud’s balance sheet or the resale value of the GPUs (Clifford Chance). Each renewal of a pledged contract is itself a refinancing event at an unknown price, so model it as rolling refinancing, where each renewal is a maturity in its own right. As PIMCO puts it, leases can outlast GPUs, GPUs can outlast customer contracts, and debt can outlast both.
The maturities are not spread out. CoreWeave has US$4.4bn due in the rest of 2026 and US$6.2bn in 2027 (Marcus Today). MUFG, Morgan Stanley, JPMorgan, Goldman Sachs and the private-credit funds appear across multiple names, so correlated maturities mean correlated refinancing conditions.
Cross-default provisions do the rest. A breach on one facility can cascade across the structure, turning a credit event into an operational one for every customer inside it (Aethir). Oracle shows the distance: force majeure was served on the 2.45GW Project Jupiter campus while US$18bn of project debt remained outstanding (SiliconANGLE). The obligation outlived the project. A downgrade from BBB pushes investment-grade-only holders into forced sales (CNBC). CoreWeave’s bonds trade at 12 to 14.5 per cent and its CDS implies a 50 per cent chance of default over five years, which is what the spread is trying to price.
How should you size a commitment so you are not the anchor customer who destabilises the structure?
Size the commitment by its proportion of the borrowing base. A commitment that is a material share of the pledged borrowing base or of the provider’s revenue, or that trips a facility’s concentration limits, gets named in the credit agreement (Matchpoint Partners). Being named turns your renewal or departure into a covenant event; below the thresholds you stay anonymous, which is protective.
Offtaker quality sets a provider’s cost of capital. Hyperscaler-backed deals price around SOFR+225, while a three-year contract with re-lease optionality prices around SOFR+550, a 325-basis-point gap on the same borrower and silicon (Data Gravity). That gap is the cost of being named: a weak offtaker is written into the credit agreement and prices the facility wide, so a buyer who is not investment grade either pays for a guarantee or becomes the reason the structure costs more.
The sizing logic therefore runs like this: keep any single commitment small enough that its non-renewal does not move the borrowing base, and cap aggregate exposure to one provider, one financing structure and one common counterparty. The honest tension: concentration is what makes the debt financeable and what makes a single departure destabilising. That trade-off only resolves against the rest of the structure; the wider neocloud credit risk picture sets out how the pieces fit.
Nscale is the cautionary case. Microsoft and Anthropic made up 85 per cent of US$103.4bn in contract value, and Microsoft pulled out of the Loughton site weeks before a planned NYSE listing (The Next Web). The anchor left, and the structure did not have a second anchor. That is what the premium is actually buying.
How do you assess a GPU cloud provider’s solvency before a three-year reservation?
So you ask four questions before signing. When does each facility mature, and what is the amortisation profile? Who are the lenders, and how concentrated is that group across the sector? What is the covenant package, and how much headroom is there against the current quarter? Is the entity signing your contract the entity carrying the debt, and if not, what is the guarantee path?
For a private operator there are no public filings, so the documents come from what the company discloses under NDA: audited accounts, the facility term sheet, a lender presentation and the credit agreement’s key definitions. Third parties fill the gaps, including Moody’s and S&P reports where the debt is rated, private-credit fund reporting where a vehicle is the borrower, and the public facilities of listed peers as a structural proxy.
Weigh the answers together. A 2027 maturity with one lender and thin headroom is a different counterparty from a 2030 maturity across ten banks, even at identical leverage. CoreWeave is the disclosure-rich example: US$35.6bn of debt, US$16.3bn of operating lease liabilities and a US$103.7bn backlog that is 72 per cent concentrated in three customers (StockTitan). Oracle runs 282 basis points of CDS against a BBB rating (CNBC) and carries US$638bn of remaining performance obligations (UBS).
Two structural red flags are decisive: cross-default provisions that reach facilities you cannot see, and a renewal cycle that depends on a single contract. If your revenue is Australian dollars and your obligation is in US dollars, test enforceability, because transition rights are only worth what the governing law and the lender’s acknowledgement make them. That is the route from a missed refinancing to a lost workload.
A commitment is a pledge, and its renewal feeds the refinancing maths. Proportion decides whether it is load-bearing. Leave with a test: stay below the concentration thresholds that make a contract named, and run the four questions before signing. The mechanisms behind each, and the levers you actually hold, are set out in the full picture across the cluster.
The cost of getting it wrong is: the anchor leaves, the structure has no second one, and the loss lands on workloads and the lender.
Frequently Asked Questions
Does a provider refinancing at a higher interest rate affect my contract after I sign?
Yes, though not through the SLA. A refinancing at a wider spread lifts the provider’s interest cost against a contract price that is already fixed, and that squeeze tends to surface at renewal as a higher price, slower capex or thinner support. The contract itself survives on paper. What changes is the provider’s willingness and ability to keep performing at the price you signed.
What actually happens to my workloads if my provider misses a debt maturity?
A missed maturity does not switch off GPUs the next morning. It usually moves through a standstill or forbearance, then covenant waivers, then enforcement or administration, and customers experience it as renewals that cannot be signed, sudden price rises and eventually a lender or administrator stepping into the operating entity. The gap between the first waiver and a lost workload is measured in weeks and months, not years.
Is my service level agreement still enforceable if my provider’s lender takes control of the GPUs?
Usually not in the way you expect. Your SLA is a contract with the operating company, not with its lender. If the lender enforces security over the GPUs, control can move to an entity that never signed your agreement, so the answer depends on who actually owns the hardware and whether your step-in or transition rights bind the lender as well as the operator.
Why does customer concentration matter to me if I am only a small customer?
Because you inherit the consequence without being named. A single anchor leaving can trip concentration covenants, force a repricing or push the facility back to lenders, and every smaller contract sitting behind it is affected. Nscale is the case: Microsoft and Anthropic accounted for roughly 85% of $103.4B in contract value, and Microsoft exited the Loughton site weeks before a planned NYSE listing.
What is a borrowing base, and why would my GPU contract be pledged into one?
A borrowing base is the pool of contracted cash flows a lender will advance against. Only contracts that meet its tests on tenor, counterparty credit quality and termination rights are counted, which is why a provider asks for detailed terms before signing and why your reservation becomes part of the security package. Your contract is not just revenue. It is collateral.
How far ahead should a neocloud have its refinancing sorted before a maturity?
Most operators should be in market twelve to eighteen months before a maturity, with a lender group and term sheet visible nine to twelve months out. A syndication typically takes three to six months in an open market. If a provider cannot name its lenders or show a term sheet inside that window, treat it as a warning rather than a timing detail.
Do Australian buyers carry extra risk on a USD-denominated GPU contract?
Two extra layers, yes. Your obligation is in US dollars while your revenue is in Australian dollars, so an adverse move in the exchange rate raises the real cost of a fixed commitment, and hedging three years out is neither free nor perfect. The second is enforceability: step-in and transition rights are only as good as the foreign governing law and the lender’s acknowledgement of them.
Is a longer, cheaper contract worth the concentration risk?
Sometimes, but only when the discount pays you for being named and for sitting through a maturity. A three-year commitment at a lower rate can be rational if the contract carries termination rights on a credit event, step-in and transition assistance, and clear data export. Without those protections you have accepted a discount in exchange for someone else’s refinancing risk.
How do I tell whether a provider’s debt load is actually a problem?
Compare the debt to contracted backlog and, more importantly, to the maturity ladder. CoreWeave carries $35.6B of debt against $103.7B of backlog, but 72% of that backlog sits with three customers, so the number is only as good as those renewals. Debt matched to long contracts is a different proposition from debt that matures before the contracts do.
What early warning signs suggest refinancing trouble before it reaches the SLA?
Watch the credit signals rather than the service dashboard: a negative rating outlook, widening credit default swaps (Oracle traded at 282bp against a BBB rating), a changing lender group, covenant waivers, going-concern language in audited accounts and discounting on renewals. Any one is a prompt to ask questions. Two together should change how you structure the next commitment.
Can I insure against my GPU provider failing?
Rarely in a useful form. Standard business interruption and equipment cover responds to hardware failure and outages, not to a counterparty running out of refinancing options, and trade credit insurance is typically short-dated and excludes named entities. The cheaper protection is structural: staged commitments, short initial terms with renewal options, portable data and exit rights written into the contract.
