The GPU-for-Rent Economy: Inside the Neocloud Gold Rush
Specialist GPU clouds like CoreWeave, Nebius, and Nscale are growing 200 percent a year on hyperscaler mega-deals and chip-backed debt. The open question is what the collateral will be worth when the loans come due.
Five years ago, none of these companies mattered. CoreWeave was a crypto miner in New Jersey. Nebius was the discarded shell of a Russian search giant. Crusoe was burning flare gas in oil fields. Today they anchor a new tier of the cloud industry: the neoclouds, purpose-built landlords for Nvidia silicon, and one of the fastest-growing infrastructure markets ever measured.
Synergy Research Group puts full-year 2025 neocloud revenue above $25 billion, with the fourth quarter alone hitting $9 billion, up 223 percent year over year. Synergy forecasts the segment will approach $400 billion by 2031, a sustained 58 percent compound annual growth rate.
Neocloud revenue passed $25 billion in 2025 and is forecast to approach $400 billion by 2031, growing 58 percent a year, according to Synergy Research Group.
“Traditional hyperscale systems were conceived around generalized elasticity, whereas AI workloads impose far more rigid constraints, particularly around parallelism, locality and compute concentration,” Synergy chief analyst Jeremy Duke said in the firm’s April 2026 forecast. Translation: the big clouds were built for everything, and AI wants something narrower, denser, and faster to stand up.
Rent the chips, mortgage the chips
The business model is closer to commercial real estate than to classic cloud computing. A neocloud signs a multi-year take-or-pay contract with a large customer, often before the data hall exists. It then borrows against that contract, and against the GPUs themselves, to buy the hardware. Rental payments service the debt; whatever is left is margin.
CoreWeave pioneered the financing template, raising a $2.3 billion GPU-collateralized facility in 2023 and a $7.6 billion follow-on in 2024, with Blackstone, Magnetar, and Carlyle among the lenders. The sector has kept borrowing: CoreWeave’s total debt reached $35 billion as of June 30, 2026, according to Capacity, and Quartz has documented a broader stack of GPU-collateralized loans across the industry, with a refinancing wall concentrated between 2026 and 2028.
The chips are the collateral, which makes the whole edifice a bet on how long a GPU stays valuable. More on that below.
The scoreboard
The 2025-2026 growth figures are the kind that make even dot-com veterans blink. CoreWeave reported second-quarter 2026 revenue of $2.58 billion, up 112 percent year over year, according to CNBC, and the company says its revenue backlog stands at $104 billion, excluding more than $25 billion in commitments signed early in the third quarter. Nebius reported first-half 2026 revenue of $981.3 million, up 529 percent from a year earlier, and guided to $3.0–3.4 billion for the full year with an exit run rate of $7–9 billion, per its SEC filings.
| Neocloud | Headline 2026 figure | Marquee commitment |
|---|---|---|
| CoreWeave | $2.58B Q2 revenue, +112% YoY | ~$22.4B from OpenAI |
| Nebius | H1 revenue +529% YoY | $17.4B Microsoft deal |
| Nscale | Pre-IPO, building out | $14B Microsoft, ~200K GB300s |
| Crusoe | ~$30B valuation talks | 900 MW Abilene campus for Microsoft |
| Lambda | >$1.5B expected revenue | $3B pre-IPO raise at ~$12B |
| Together AI | $1B annualized revenue | $800M raise at $8.3B valuation |
Lambda is expected to top $1.5 billion in revenue this year while seeking up to $3 billion at a $12 billion-plus valuation, according to Tech Funding News. Together AI crossed $1 billion in annualized revenue in February 2026 and raised $800 million at an $8.3 billion valuation in July, per Sacra. Crusoe, per multiple reports, has been in talks to raise about $3 billion at a valuation near $30 billion, roughly triple its level eight months earlier.
Microsoft and OpenAI, anchor tenants
Strip away the branding and most neocloud revenue traces back to a handful of names. OpenAI has contracted up to roughly $22.4 billion of capacity from CoreWeave across a March 2025 deal and two expansions, according to Sacra’s tally of company disclosures. Microsoft, meanwhile, has become the sector’s universal anchor tenant: a five-year, $17.4 billion contract with Nebius (expandable to $19.4 billion) signed in September 2025, according to CNBC; a $14 billion deal with the UK’s Nscale for about 200,000 Nvidia GB300 GPUs across Europe and Texas, per CNBC; and a 900 megawatt Crusoe campus in Abilene announced in March 2026, part of a site Crusoe says will eventually reach 2.1 gigawatts. Lambda, too, counts Microsoft among its anchor customers, according to Tech Funding News.
That concentration cuts both ways. Investment-grade counterparties make the debt financeable; Nebius touts more than $40 billion in additional contracted revenue from customers such as Microsoft and Meta. But if two or three buyers pause, the growth story pauses with them.
How fast does a GPU die?
The sector’s most contentious spreadsheet cell is depreciation. Hyperscalers and CoreWeave book GPU server life at five to six years. Skeptics, most loudly investor Michael Burry, argue the true economic life is closer to two to three years, and that the gap will overstate big-tech earnings by a cumulative $176 billion between 2026 and 2028, according to CNBC’s coverage of the debate. Amazon has already trimmed the useful life of a subset of its servers from six years to five, citing the pace of AI development.
The market data sits uncomfortably in the middle. Rental rates for Nvidia’s H100 fell from roughly $8 per hour at the 2023 peak to under $2.50 by mid-2025 on Silicon Data’s rental index, a steep decline in earning power even as the physical chips keep running. Defenders counter that older accelerators, including 2020-era A100s, still rent out profitably for inference and fine-tuning, and that software optimization extends useful life. Both things can be true: the chips last six years, but their peak earnings may not last three.
Boom, with a maturity date
The bull case is straightforward: AI compute demand still exceeds supply, contracts are backed by the most creditworthy companies on Earth, and Synergy’s numbers describe a market compounding at nearly 60 percent. The bear case is equally clean: this is a capital-intensive, debt-heavy business renting a depreciating asset to a customer base you can count on one hand, with Nvidia simultaneously acting as supplier, investor, and occasional backstop, a circularity that draws comparisons to 1990s vendor financing.
Estimates diverge, and honest analysis says so: GPU life of two years or six, a 2031 market of $50 billion or $400 billion. Somewhere between those ranges lies the answer to whether the neoclouds are the next hyperscalers or the fiber glut of the 2020s. The refinancing window of 2026-2028 will start telling us which.