What Is a Neocloud? How AI's GPU Landlords Rent Compute
As of October 2026: figures are dated to this month — when the semiconductor index lost $1.6 trillion in six sessions and Gartner warned 80% of neoclouds could be gone by 2030.
CoreWeave co-founder Brannin McBee said something in October 2026 that should make anyone's eyebrows move. Speaking to the Financial Times, he said the company was chasing an investment-grade credit rating — a rating, he said, that would open the door to "unlimited" pools of capital. A company that rents out computer chips wants its appetite for debt treated as a feature, and for four years the market has more or less agreed.
Welcome to the neocloud: the specialist GPU cloud. These companies sit between Nvidia's chip fabs and the AI labs, buying the world's most expensive processors and leasing them back to anyone training or running a large AI model. They are, in effect, renting artificial intelligence its brains, one GPU-hour at a time.
CoreWeave is the poster child — a former crypto miner, public since March 2025, running 250,000-plus Nvidia GPUs with a $104 billion contract backlog. Around it: IREN, the Australian bitcoin miner turned AI cloud; Nebius, a rebranded slice of Russia's Yandex; Nscale, a British challenger racing to an IPO. One of the fastest-scaling infrastructure categories ever seen — and one of the most leveraged. This explainer covers what a neocloud is, why AI created them, how the money works, and where the edifice could crack.
First, the definition: a neocloud is a one-job cloud
Forget AWS, Azure, Google Cloud. A hyperscaler is a supermarket: thousands of services sold to every enterprise. A neocloud is a butcher selling one cut of meat — GPU compute for AI training and inference, at scale. No email hosting, no databases. GPU clusters, high-bandwidth memory, and the ultra-fast networking (InfiniBand, NVLink) that lets thousands of chips act as one brain.
The category has flavours: full-stack AI clouds (CoreWeave, Nebius, Nscale) with orchestration and tooling; per-hour GPU-as-a-service shops like Lambda Labs; sovereign clouds for data-residency needs; and the converted crypto miners (IREN, Core Scientific, TeraWulf — and CoreWeave itself, founded in 2017 as a miner called Atlantic Crypto).
One thing to get right: neoclouds do not replace hyperscalers. They sit alongside them, a purpose-built compute layer for AI workloads the general cloud never served efficiently — which is why even Microsoft and Meta are major neocloud customers.
Why the AI boom created them: nobody else had the chips
The neocloud exists because of a shortage. After ChatGPT's 2022 debut, every AI lab needed the same thing — mountains of Nvidia H100s — and the hyperscalers could not hand them out fast enough, especially not without the enterprise-price baggage (compliance, support, multi-region operations) baked into every GPU-hour.
Neoclouds offered the opposite pitch: GPU compute and nothing but, reportedly two to seven times cheaper than hyperscalers, online in months rather than the three to five years a traditional data centre takes. The crypto converts had a head start — land, power, cooling, and warehouse buildings already owned, mining rigs swapped for AI GPUs.
Power is the true moat. The bottleneck in AI is grid capacity, and miners held scarce grid connections newcomers cannot get. That is why Anthropic's reported $530 billion lease scramble "increasingly depends on current and former bitcoin miners such as Hut 8, TeraWulf, IREN and Nscale."
Inside CERN's data centre — the dense, power-hungry rack halls that the modern neocloud GPU fleet echoes. Photo: Hugovanmeijeren via Wikimedia Commons, CC BY-SA 3.0.
The business model: borrow, buy, lease, repeat
The neocloud model is brutally simple, which is why it is either genius or doomed. Borrow money. Buy Nvidia GPUs — the most sought-after, fastest-depreciating hardware on Earth. Lease them to AI labs at a markup. Pledge those GPUs and the future lease revenue as collateral for the next round of borrowing. Collect the spread. Pray demand never blinks.
CoreWeave's numbers show the spread at industrial scale. It listed in March 2025 at $40 a share, and its S-1 contained one of the most alarming sentences in AI finance: over half of its $1.9 billion in revenue came from one customer — Microsoft. The backlog has since exploded to $104 billion (Q2 2026), anchored by a reported $22.4 billion OpenAI arrangement and a $21 billion Meta commitment. We checked the paper trail: debt above $17 billion in the spring, about $35 billion by mid-year, total liabilities near $72 billion by autumn against $5 billion of equity. The backlogs are real. So is the leverage.
IREN tells the same story in miniature: a five-year, $9.7 billion deal supplying Microsoft with GPU cloud from Texas, plus $3.4 billion to run managed GPU cloud for Nvidia's own AI research — while carrying over $7.4 billion in debt, up from ~$965 million a year earlier. On the August earnings call, co-founder and co-CEO Daniel Roberts teased investors with a line you can practically hear the wink through: "The headline is the new multiyear contract with a leading frontier AI lab. And to be clear, this is a new contract."
From bitcoin mines to $530 billion leases: the money arc
In 2017, crypto miners stacked GPUs to mint coins. In 2026, AI labs are signing leases so large they warp the vocabulary of corporate finance.
The crescendo arrived this month, when The Information reported that Anthropic has signed roughly $530 billion of data-centre leases — over 15 gigawatts across AWS, Google, SpaceX's compute arm, and a dozen others. Our earlier reporting covered Anthropic's reported multiyear cloud deal with IREN — unconfirmed by either side. The mega-site template is Hut 8's Louisiana project: Hut 8 develops, Fluidstack leases and operates, Anthropic supplies Google's tensor processing units. Per analyst Christian Okoye of Occam Edge, bringing it online one month early adds over $1.3 billion in revenue.
The international cast is filling in fast. Nebius — carved out of Yandex's non-Russian assets in 2024 under CEO Arkady Volozh — has a $17.4 billion Microsoft deal, a Meta contract worth up to $27 billion, and a $2 billion Nvidia investment. Nscale, the British upstart, has raised over $4.5 billion and signed a $23 billion-plus Microsoft contract ahead of a planned IPO. Lambda sits at the centre of a $35 billion cloud deal with Anthropic that Nvidia itself is backstopping. Small wonder Anthropic's turn to profitability has partners queuing to sign.
The circular-financing problem nobody can wave away
When Nvidia invests $2 billion in CoreWeave, sells CoreWeave the chips, and agrees to buy any capacity CoreWeave cannot lease — a backstop running through April 2032 — is that a chip sale, or Nvidia buying its own demand?
The loop runs tighter than most realise. Nvidia has offered select neoclouds minimum revenue guarantees on supplied GPUs, arranged over $500 billion in third-party capital (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR) to finance customer purchases, and is reportedly working with broker Howden Re on depreciation insurance for AI data-centre loans. CEO Jensen Huang calls this building an "ecosystem."
Critics call it vendor financing — the mechanism that inflated the dot-com bubble. And the critics have a figurehead: Michael Burry, who turned his Substack Cassandra Unchained into a forensic audit of the AI trade. "A chip earning more than it costs is a sign of scarcity due to memory and power shortages, not GPU die shortages and not durability," he wrote this month, calling a September Nvidia slide an "apples-to-oranges" illusion. His deeper charge: accelerated depreciation admits each GPU generation colossally devalues the last.
What happens if the demand story cracks
Every neocloud bull case rests on one assumption: AI demand keeps compounding. Strip away the jargon and the assumption looks shoddy in three places.
First, leverage. A neocloud's balance sheet is an amplifier: in good times the debt machine prints money; when utilisation or lease rates wobble, it works in reverse. CoreWeave raised its 2026 capex forecast to $39 billion while quarterly interest expense ran into the hundreds of millions — the story behind AI's $113 billion debt machine hitting a wall.
Second, concentration. The S-1 confession — half the revenue from one customer — has improved but not disappeared. A handful of labs and hyperscalers anchor nearly every backlog. If one slows or builds its own capacity, "unlimited" capital becomes musical chairs.
Third, the chips. A GPU generation lasts commercially a few years before the next one makes it look slow — Burry's depreciation point. Nvidia's Rubin and Vera architectures are already committed in financing deals, so today's premium chips are tomorrow's commodity hardware, and whoever overpaid at the peak holds the bag. Gartner's October warning — 80% of neoclouds could shut down by 2030, $3.5 trillion of AI infrastructure spending at risk — is the research establishment's way of saying the shakeout is about who, not if.
My judgment: the neocloud is a genuinely useful invention and a genuinely dangerous financial structure — the same fact, twice. The world needed specialist GPU compute, and neoclouds built it faster than anyone. But the financing that bought that speed is colossally fragile: debt stacked on GPUs stacked on backlogs stacked on the assumption the boom never pauses. When it pauses, the survivors will be the ones with creditworthy customers, real power assets, and cash, not the biggest backlogs. The backlog is a promise. The debt is a fact.
Server racks at the NOIRLab Headquarters in Tucson, Arizona — the dense hardware behind every cloud GPU. Photo: NOIRLab/NSF/AURA/T. Slovinský via Wikimedia Commons, CC BY 4.0.
Frequently asked questions
What is a neocloud?
A specialist cloud provider renting out GPU compute for AI workloads — training and running AI models — as its core business. Unlike hyperscalers (AWS, Azure, Google Cloud), a neocloud does one job: GPU-as-a-service. Names to know: CoreWeave, Nebius, Nscale, IREN, Lambda, Crusoe, Fluidstack.
How do neoclouds make money?
Borrow, buy, lease, repeat. Raise debt, buy Nvidia GPUs in bulk, lease them to AI labs at a markup, and pledge the chips and contracts as collateral for the next round. Profit is the spread between financing costs and lease income at high utilisation.
Neocloud vs hyperscaler: what's the difference?
A hyperscaler sells everything to everyone; a neocloud sells GPU compute for AI workloads — typically cheaper and faster to deploy. They complement each other: Microsoft and Meta are major neocloud customers. A neocloud can't run your email or databases.
Are neoclouds risky?
Structurally, yes. Heavy debt, a handful of giant customers, and an asset that depreciates fast as new GPU generations arrive. Critics call the financing circular; Gartner warned 80% of neoclouds could shut down by 2030. Real businesses; fragile balance sheets.
Who are the biggest neoclouds?
CoreWeave (largest, Nasdaq-listed, 250,000+ GPUs), Nebius (Amsterdam, ex-Yandex, Nvidia-backed), IREN (Australian, ex-bitcoin miner, Microsoft's $9.7 billion partner), Nscale (British, heading for an IPO), plus Lambda, Crusoe, Fluidstack, and Applied Digital.
Update, Oct. 18: the financing thesis just got its stress test
Two days after this explainer published, the financing fragility it describes went on public display: CoreWeave's five-year credit default swaps hit a record 1,100–1,195 basis points and its 9.625% 2032 notes traded near 77 cents on the dollar — the bond market pricing roughly a 63% chance of default within five years (reported in our Oct-17 sell-off coverage).
The category keeps scaling anyway: Synergy Research Group puts neocloud revenue above $25 billion in 2025 ($9 billion in Q4 alone, up 223% year over year) and forecasts ~$400 billion by 2031 — a 58% compound annual growth rate (via Nutanix). ABI Research expects 2,200+ neocloud-operated data centers by 2035, up from 558 in 2025.
Both can be true — that is the neocloud paradox this page exists to explain.
How much cheaper are neoclouds than hyperscalers?
Estimates put neocloud GPU-hours 40–70% below hyperscaler pricing; on a $1–5M foundation-model training run that is $500K–$2M in direct savings (iterate.ai). The catch: you get raw compute, not the managed services the markup buys.
When did neoclouds emerge?
The term caught on in late 2024 and became an industry-standard category through 2025, as CoreWeave (IPO early 2026), Nebius, Crusoe and Lambda scaled GPU fleets hyperscalers couldn't deliver fast enough (Saturn Cloud, Nutanix, RunPod).
Sources
Reporting drew on Reuters, The Information, Bloomberg, the Financial Times, InfoWorld, Verdict, NetworkWorld, Goldman Sachs research, Michael Burry's Substack Cassandra Unchained, Gartner, and company disclosures. Figures as reported in October 2026.
Comments
Post a Comment