Neocloud.
Banks with GPUs.

Anthropic committed about $80 billion to neoclouds in a single week this summer, if you count a reported $35 billion deal with Lambda. That's how badly the AI labs want computing power, and how happy they are to rent it.

A neocloud is the landlord in that deal. It buys Nvidia GPUs by the hundred thousand, packs them into data centers built for AI, and rents the computing power to whoever needs it: OpenAI, Anthropic, Meta, Microsoft. Even the richest tech companies can't build fast enough on their own, and renting keeps the bill off their balance sheets. The neocloud takes on the build, the GPUs, and the debt.

Demand isn't the worry. GPU rental prices have held up instead of crashing, and CoreWeave raised its prices about 25% in July. The worry is the debt. CoreWeave and Nebius sold almost $10 billion of convertible debt between them in August and September, with the 10-year Treasury yield above 5% for the first time since 2007.

That's the real business. A neocloud borrows money for years to buy chips that stay top of the line for maybe three or four, then rents them out under contracts that have to pay it all back. Neocloud stocks rise and fall on that spread.

Who carries the GPU risk decides where each of these neocloud stocks sits:

• Operators: own the GPUs and run the clouds. The most growth, and the most debt.
• Landlords: former bitcoin miners that own the power and the buildings and lease them out for 10 to 20 years. Less upside, longer contracts.
• Private: two big names still waiting on a ticker.

These companies own the GPUs. When AI demand moves, they move first, and they carry the risk that today's chips are worth a lot less in three years.

CoreWeave NASDAQ: CRWV

CoreWeave is the yardstick for the whole group. Q2 revenue more than doubled to $2.58 billion, backlog sits around $104 billion, and contracted power has grown to about 4.2 gigawatts. In September it said short three-to-six-month deals signed this quarter were pricing at about $40 million per megawatt a year, which says demand is still running ahead of supply.

The same day, it launched what became $4.2 billion of convertible debt, plus a new share-sale program. Net interest was $640 million in Q2, up from $267 million a year earlier, and the company lost $626 million. CoreWeave is the best revenue machine in the group, financed with the most expensive money in the group.

Nebius NASDAQ: NBIS

Nebius used to be the low-debt alternative to CoreWeave. In August it sold about $5.75 billion of convertible notes, so that's no longer the pitch.

The pitch now is growth. Q2 revenue from its AI cloud was $575 million, up 514%, with long contracts from Meta and Microsoft that run into the early 2030s, and Meta has agreed to buy any unsold capacity on some clusters. It's the fastest grower among the public operators, and the market has priced it that way.

IREN NASDAQ: IREN

IREN just showed the rest of the group how to borrow. It financed the GPUs for its $9.7 billion Microsoft contract with $3.6 billion of investment-grade debt at 6%, a cost of money most neoclouds can't touch. It also signed a contract with an unnamed frontier AI lab and a five-year, $3.4 billion cloud deal with Nvidia.

The former bitcoin miner is quitting mining by year-end. It lost $703 million in the fiscal year to June, mostly on non-cash write-downs from the pivot. If the cheap financing keeps coming, IREN has the cost advantage that decides this business.

Nscale NYSE: NSCL, pending

London-based Nscale filed to list on the NYSE in September, and its filing is the clearest look yet at neocloud economics. It reports about $103 billion of contracted revenue, with roughly 85% of it coming from just two customers, Microsoft and Anthropic. Only about $2.6 billion of that contract value was active by the end of August. First-half revenue was $141 million against a $1 billion loss.

That's the whole neocloud model in one filing: huge contracts, tiny current revenue and a giant bill to build the gap. When Nscale prices, every stock above gets repriced against it.

Former bitcoin miners that spent years collecting the scarce stuff: land with power, grid connections and substations. They lease powered buildings to AI tenants on 10- to 20-year contracts and leave the GPU risk to someone else.

Applied Digital NASDAQ: APLD

Applied Digital started as CoreWeave's landlord in North Dakota and has moved up the tenant list. This spring it signed two $7.5 billion, 15-year leases with an investment-grade hyperscaler, then a third lease worth about $5.2 billion, bringing contracted capacity to about 1.4 gigawatts across five campuses. A landlord is only as good as its tenants, and Applied Digital's just got a lot better.

TeraWulf NASDAQ: WULF

Anthropic signed with TeraWulf directly: a roughly $19 billion, 20-year deal for about 400 megawatts at a converted industrial site in Kentucky. Its earlier leases with Fluidstack come with Google guaranteeing $3.2 billion of the rent. When the biggest names in AI back your tenants, borrowing gets cheaper, and TeraWulf is reportedly lining up about $3.5 billion of debt for the Kentucky campus.

Hut 8 NASDAQ: HUT

Hut 8 has about $19.6 billion of 15-year leases at its Beacon Point campus in Texas, and the tenant, according to the Wall Street Journal and the FT, is Nvidia itself. Hut 8 only calls it a high-investment-grade company. The reports say Nvidia supplies the chips and Lambda resells part of the capacity to Anthropic. Behind that sits an 8.7-gigawatt development pipeline, the biggest land bank relative to market value among the miners.

Cipher Digital NASDAQ: CIFR

Cipher rebranded from Cipher Mining this year, and its best asset is a tenant: Amazon Web Services, on a 15-year, 300-megawatt lease. Another 300 megawatts goes to Fluidstack with Google behind it. It's smaller than Hut 8 and TeraWulf, but it has the one hyperscaler tenant most landlords would trade their pipeline for.

Their funding rounds set the private comps, and their listings will be the next big price signals for neocloud stocks.

Crusoe

Crusoe builds Stargate. Its Abilene, Texas campus for Oracle and OpenAI was the first Stargate site. In September it raised $3.9 billion at a $30.9 billion valuation, and Bloomberg reported the same month a roughly $13 billion, five-year compute deal with trading firm Jane Street. If that's confirmed, it's a customer outside the usual AI labs, and a sign that compute demand is spreading beyond them.

Lambda

Lambda is the Nvidia-backed GPU cloud that reportedly landed a $35 billion Anthropic deal at the end of August, running on capacity Nvidia reportedly leases from Hut 8. Reports say it's lining up banks for an IPO as soon as 2027. When it comes, it'll show what the market pays for a neocloud this tangled up with Nvidia.

Strip out the AI, and a neocloud is an aircraft lessor.

Borrow money, buy expensive equipment with a limited useful life, rent it to customers on multi-year contracts, and earn the spread. The winners in that business aren't the companies with the most planes. They're the ones with the cheapest money and the best tenants, because when every lessor flies the same jets, the cost of capital is the only edge left.

GPUs are planes that go stale in four years instead of twenty-five. So the number that matters is what a neocloud pays to borrow compared with what its customers pay to rent. IREN borrowing at 6% against a Microsoft contract is a moat. A neocloud borrowing at 9% against a startup's contract is a countdown.

Today the market values neocloud stocks like fast-growing tech, and leasing companies trade at a fraction of that. As that gap closes in 2027, the neoclouds that lock in cheap, long, investment-grade money now will look like banks with GPUs. The rest are just carrying the debt.