From crypto miners to AI hyperscalers
CoreWeave started in 2017 as Atlantic Crypto, mining Ethereum on GPUs. It pivoted to renting out cloud compute in 2019, after the crypto downturn made mining less attractive and the same hardware turned out to be worth more rented than run. Crusoe began by burning flared natural gas to mine Bitcoin and later expanded into AI compute; our sources do not pin down the year it turned. Nebius is the exception that spoils the pattern — it came out of Yandex's cloud and infrastructure business, not a mine.
The crypto ancestry is not a joke about the people involved. It is the reason the assets exist. Mining taught a generation of operators how to secure cheap power, get an interconnection queue position, and put a very large number of hot, fast-depreciating processors in a shed. Swap the workload and you have a neocloud.
The GPU-collateralised debt machine
The scale is easiest to see at CoreWeave, which is public and therefore has to tell us. In the quarter ended 30 June 2026, it reported revenue of $2.575bn, a net loss of $626m, capital expenditure of $9.4bn in the quarter alone, and total principal debt of $35.551bn. Against that sits a backlog — remaining performance obligations — of roughly $104bn, plus $10.0bn of undrawn revolver and delayed-draw capacity (The Register, CoreWeave Q2 2026 results). Nine billion dollars of quarterly capex against two and a half billion of quarterly revenue is not a rounding difference. It is the entire business model: borrow now, deliver contracted compute later.
The borrowing is asset-backed lending, and the asset is the chips. The borrower drops GPUs, the surrounding infrastructure and usually the assigned cash-flow rights from its customer contracts into a , and the lender takes a perfected first-priority security interest so it can seize and sell the hardware in a default. New chips are valued off the manufacturer price; used chips are against secondary-market broker pricing; both are then into a borrowing base. Reported loan-to-value ratios cluster around 50–70%, with some structures quoted as high as 70–80% where the contracts backing them are strong. Maturities are typically two to four years. Pricing is commonly cited at plus 400 to 700 basis points, or roughly 8–12% all-in. Recent reporting places Blackstone, Magnetar, Blue Owl and Coatue among the institutional capital providers in this market, generally through private-credit and structured-credit vehicles rather than plain corporate loans.
CoreWeave's own paper fits the template exactly. Earlier in 2026 it closed a $3.1bn term loan the company described as the first publicly syndicated delayed-draw facility backed by high-performance computing infrastructure. In July and August 2026 it closed a $2.6bn delayed-draw facility priced at SOFR plus 550 basis points with an approximate five-year maturity, issued by an entity named CoreWeave Financing DDTL V-V, LLC and secured on GPU and HPC infrastructure for committed customer deployments (CoreWeave). Bloomberg reported that CoreWeave had to raise the yield on that loan to get it done, and described the facility as tied to Anthropic (Bloomberg). Not every deal clears at the first price.
Equivalent totals for Nebius, Lambda and Crusoe are not something we can give you. They did not appear in the filings and financing announcements our researchers could reach. That absence is itself the point: most of the sector's leverage sits in private vehicles at private companies, and nobody outside the lenders has the full picture.
The six-year argument
Everything above rests on one number: how long a GPU keeps earning. CoreWeave's chief executive Michael Intrator has defended the company's assumption directly — "We use a six year depreciation. We believe that the GPUs will last in excess of six years, but we felt like that was a fair and reasonable" basis, he said on the All-In podcast. He has also argued publicly that older generations of GPUs retain value.
The lenders behind the same chips are not writing six-year documents. Facilities in this market are commonly structured to amortise at something like 20–30% a year, with periodic revaluation, so that the outstanding balance falls at least as fast as the collateral is expected to. That is a schedule that pays itself off in three to five years. One side of the trade is telling equity investors the asset lasts six years or more; the other side is contractually refusing to be exposed to year five. Both can be sincere. They cannot both be right about the .
The short-seller view is harsher still: Kerrisdale Capital's September 2025 report modelled net leverage peaking around 6.0x in 2025 and total debt exceeding $40bn by 2028, at which point it expected the company still would not be cash-flow positive (Kerrisdale). S&P, for its part, has been moving the other way, revising CoreWeave's outlook to positive while flagging FFO-to-debt below 12% and CFO-to-debt below 10% as the levels that would take it back to stable (S&P Global Ratings).
The Nvidia circular financing loop
On 9 September 2025, under an order form to a master services agreement dated 10 April 2023, Nvidia agreed to buy CoreWeave's residual unsold data-centre capacity through 13 April 2032, an arrangement with an initial stated value of $6.3bn (8-K). The company that manufactures the chips has promised to rent back, for seven years, whatever compute the chip buyer fails to sell to anyone else. Reuters described it as a guarantee that Nvidia will purchase any cloud capacity not sold to customers. The commercial effect is that Nvidia is underwriting value of its own hardware, which is precisely the thing lenders need underwritten before they will advance against it.
Nvidia is also a shareholder. It put $100m into CoreWeave in 2023, roughly $2bn for 24.2 million shares in the fourth quarter of 2025, and a further $2bn of Class A stock at $87.20 a share in January 2026 (8-K). Its 13F disclosed in August 2026 showed 47.2 million shares worth about $3.66bn, roughly 11% of the company (Yahoo Finance). The same pattern repeats elsewhere: about $2bn of Nvidia equity in Nebius, an equity stake in Lambda alongside a multi-year contract of around $1.5bn under which Nvidia leases back roughly 18,000 servers, making Nvidia — by that account — Lambda's largest customer.
Intrator rejects the framing. He has called the circular-investment narrative "fundamentally flawed" (CNBC), and has described the relationship with Nvidia as "symbiotic but not equal". He may be right that every leg of it is a real commercial arrangement at a real price. It is still true that one company supplies the chips, owns about a tenth of the buyer, and has signed a document making itself the buyer of last resort for the output of those chips until 2032. The word for that structure is not an insult. It is a description.
Why it matters to this crash
The reason neoclouds sit near the centre of the story is that they are the leveraged expression of it. pay for compute out of operating cash flow. The neoclouds borrow at project level against contracted leases, which means they are the part of the complex that reprices violently when anyone doubts the pace of the buildout.
That is not theory. In our 19 August 2026 dispatch we noted CoreWeave down 15.6% over five days, Nebius down 13.6%, Core Scientific down 10.2% and 20.8% over twenty days, Applied Digital down 9.3% — while Microsoft was up 24% over twenty days and Alphabet was flat (our coverage). That is not a sector selling off; it is the market sorting the complex by balance sheet. The same week we wrote that credit had not yet joined in: at 270bp, investment grade at 81bp, and Blackstone-backed QTS pricing a $3.9bn five-year data-centre bond at about 7.63% on a Baa3/BBB− rating and still drawing $10bn of indications (our coverage). Investment-grade rating, junk price, and it upsized (our coverage).
The fragility underneath is customer concentration. In the second quarter of 2026, three customers accounted for 72% of CoreWeave's revenue, contributing 36%, 26% and 10% respectively. Summaries of page 33 of the S-1 put Microsoft at 35% of 2023 revenue and 62% of 2024 revenue; the company has not broken out Microsoft and OpenAI separately in the more recent disclosures we have seen. Reported headline contracts — CoreWeave with OpenAI at $22.4bn and with Meta at $32bn, Nebius with Meta at up to $27bn over five years announced in March 2026, Nebius with Microsoft at about $9.7bn announced in 2025, Lambda with Microsoft at $1.3bn over four years announced in September 2025 — are how the backlog gets to nine and twelve figures. Several of those values reach us through secondary coverage without stated tenor or announcement date, which is worth remembering every time someone quotes a backlog number at you. A backlog is a promise from a small number of very large companies, and a debt schedule is a promise to lenders. Only one of those is enforceable against you.
What would make this dangerous
The specific chain to watch runs from demand to collateral to lenders. A visible slowdown in hyperscaler ordering, or a next-generation Nvidia part that makes the installed base uneconomic faster than six years, would push secondary-market GPU pricing down. Because facilities are marked with periodic revaluation against a borrowing base, falling secondary prices mechanically shrink how much the is allowed to owe — which forces prepayment out of cash the operator is already spending on capex. Watch for facilities repricing wider than the SOFR plus 550 CoreWeave paid in July 2026, for a deal that has to be pulled rather than merely re-yielded, and for any to shorten the six-year depreciation assumption, which would hit reported profitability directly.
The second thing to watch is the backstop itself. Nvidia's $6.3bn obligation runs to 13 April 2032 and has an initial stated value, not an unlimited one. Anything that clarifies its size, its conditions or its accounting treatment matters more to the debt than another quarter of backlog growth does.
The precedents are not reassuring, though they differ. Compute North, founded in 2017 in Eden Prairie, Minnesota, filed for Chapter 11 in Texas in September 2022 with more than $500m of liabilities and is now described simply as a bankrupt operator. Core Scientific filed on 21 December 2022 with about $4m of cash after a $434.8m net loss in the third quarter, had its plan confirmed in January 2024, emerged with debt cut by roughly $400m and about 724MW of capacity intact, and promptly repurposed part of it for AI and HPC colocation — the same sheds, the third workload. And Exodus Communications, the dot-com era's great independent hosting company, filed on 26 September 2001 with $3.5bn of assets against $4.4bn of liabilities and a market capitalisation of about $100m, having been worth north of $30bn at the peak. The business was auctioned in bankruptcy court. Investors received nothing.
The distinction that decided those outcomes was whether the physical asset outlived the financing. Power interconnections and shells did. Application-specific mining rigs did not. Whether a 2026-vintage GPU behaves more like a substation or more like an ASIC is the entire question, and it will be answered by the secondary market rather than by anyone's depreciation policy.