What is a neocloud? GPU clouds, explained

A neocloud is a cloud company that does one thing: rent GPUs to people training and running AI models. No email, no databases, no app store. Racks of NVIDIA chips, power, cooling, and a contract. The name arrived with the AI boom, and at least eight of them now trade on public markets.

That much most explainers cover. The part that gets skipped is how the chips get paid for, and who is on the hook if the customer walks. That is where the risk in these stocks actually lives, so this post spends most of its time there.

Where did the word come from?

SemiAnalysis, the chip industry research firm, used it in an October 2024 report on "a new breed of cloud compute provider focused on offering GPU compute rental." The prefix just means new. The business is old: rent computers by the hour. What changed is that the computers are NVIDIA accelerators, the customers are AI labs, and the contracts run for years.

By August 2026 the word had a ticker of its own. Roundhill launched the first neocloud ETF, NCLD, with Nebius, CoreWeave, IREN, Hut 8 and TeraWulf as its five largest holdings.

How do neoclouds make money?

Two ways, and the second one matters more.

The storefront is on-demand rental. A customer spins up GPUs by the hour and pays a list price. At the large neoclouds this is a small share of revenue.

The business is the committed contract. A customer books a block of capacity for several years, usually on take-or-pay terms, so it pays whether or not it uses the machines. Applied Digital describes its entire backlog that way in its annual report. CoreWeave reported a revenue backlog of $104 billion at June 30, 2026, against quarterly revenue of $2.6 billion. The backlog is the business. The quarter is how much of it has been delivered so far.

The customers are the companies whose demand outran what they could build themselves: OpenAI, Meta, Anthropic, and Microsoft, which is a hyperscaler renting from a neocloud. CoreWeave's contract with Meta is worth about $21 billion through 2032. Nebius signed Microsoft for $17.4 billion over five years, rising to $19.4 billion if Microsoft takes the optional extra capacity, and it counts Meta as a customer too. IREN signed Microsoft for about $9.7 billion over five years in November 2025. TeraWulf leased roughly 401 megawatts to Anthropic for 20 years, a deal it values at about $19 billion.

How is a neocloud different from a hyperscaler?

A hyperscaler is one of the handful of companies running a general-purpose cloud at global scale: Microsoft, Amazon, Google, and a few others.

	Neocloud	Hyperscaler
Product	GPU compute, sometimes with the network fabric around it	Hundreds of services: storage, databases, apps, plus GPUs
Customers	A handful of AI labs and large software companies	Millions, from startups to governments
Revenue shape	A few multi-year contracts	Diversified, recurring, mostly not AI
Balance sheet	Chips bought with debt, secured on the chips	Chips bought with cash flow from other businesses
If AI compute prices fall	Nothing else to fall back on	Other businesses cushion it

The focus is the whole point. A neocloud can stand up a large cluster faster than a hyperscaler's internal queue and will sign a fixed price for years. The same focus means one customer's change of plans is the whole company's problem.

Which neocloud stocks are public?

Supercycle tracks 21 companies in the Neoclouds category. These are the ones where AI compute is already most of the story. The AI Exposure Score column is rendered from the live leaderboard each time the site builds, so it is current as of the site's data date.

Company	Ticker	What it rents	Exposure score	Anchor customer
CoreWeave	CRWV	GPU clusters, own and leased sites	97	Microsoft, OpenAI, Meta
Nebius Group	NBIS	GPU cloud, sites in the US and Finland	97	Microsoft, Meta
Applied Digital	APLD	Purpose-built AI data centers, leased whole	95	CoreWeave
TeraWulf	WULF	Powered sites converted from bitcoin mining	85	Anthropic, Fluidstack
Core Scientific	CORZ	GPU colocation on former mining sites	80	CoreWeave
IREN	IREN	GPU cloud on renewable-powered sites	65	Microsoft
Cipher Mining	CIFR	Powered sites, leasing to AI tenants	55	Fluidstack, backstopped by Google
Galaxy Digital	GLXY	The Helios campus in Texas	40	CoreWeave

Two patterns in that table. First, the bottom half were bitcoin miners two years ago. They already owned the two things a neocloud needs most, cheap power and permitted land, so they leased it to AI tenants and let mining fade. Second, CoreWeave appears as a customer three times. A meaningful slice of the sector is landlords to one tenant.

Nebius is the odd one out by origin: it was formed in 2024 from the international assets of Yandex, the Russian search company, after the Russian business was sold.

Who actually owns the chips?

This is the question the stock pages rarely answer. A GPU cluster costs billions, and neoclouds did not have billions. So the chips are bought with debt, and the debt is secured on the chips.

CoreWeave wrote the playbook. In May 2024 it raised a $7.5 billion facility led by Blackstone and Magnetar, with the GPUs as collateral. In 2026 it closed an $8.5 billion loan that Moody's rated investment grade, the first GPU-backed financing to get there, maturing in March 2032. Applied Digital funded its second Polaris Forge campus with $2.15 billion of senior secured notes.

Then there is the circle. NVIDIA sells the chips, and it also invests in the buyers: it held about 11.5 percent of CoreWeave's Class A shares in January 2026 and took roughly $2 billion of Nebius through warrants in March 2026. NVIDIA has also agreed to buy up to $6.3 billion of CoreWeave capacity that goes unsold through 2032. Google backstops Fluidstack, the private neocloud that anchors TeraWulf and Cipher, with credit support and equity.

The lenders own the chips until the contracts pay them off, and the chip maker owns a piece of the companies renting them. When a neocloud announces a giant contract, the next question is which lender's collateral just got safer.

What are the risks?

• One or two customers can be most of revenue. Applied Digital's named backlog is essentially all CoreWeave. If a contract gets renegotiated, the debt behind it does not shrink.
• The loans come due. GPU-backed debt across the sector matures between 2026 and 2028, from a small circle of the same lenders, and the chips have to keep earning through that window.
• The chips lose value fast. Rental rates for H100-class GPUs fell steeply from early 2024 to late 2025 as newer chips arrived. Collateral that depreciates that quickly is unusual for infrastructure debt.
• The customers can build their own. Microsoft, Meta and Google rent from neoclouds because their own sites are not ready. Every hyperscaler campus that opens is capacity a neocloud might have sold.
• Deals fall through. Core Scientific shareholders voted down CoreWeave's roughly $9 billion offer in October 2025, so consolidation is not a given.

Where do neoclouds sit in the buildout?

They are the layer between the chip and the model. Hyperscaler capex flows to NVIDIA and the server makers; neoclouds are the alternative route for the same money, funded by debt instead of cash flow. Supercycle's buildout tracker follows the physical side: IREN's Sweetwater campus in Texas is planned at 2 gigawatts, Applied Digital's Polaris Forge 1 in North Dakota has 400 megawatts contracted, and TeraWulf's Justified Data Campus in Kentucky is the site of the Anthropic lease. Each row links to its source and to the tickers exposed to it.

CoreWeave raised its 2026 capital spending guidance to between $35 billion and $39 billion in August 2026. Nebius reaffirmed $20 billion to $25 billion. Those two numbers are the size of the bet.

What to watch in neocloud stocks

The tells are public. New multi-year contracts, and who the counterparty is. Capex guidance, quarter by quarter, on the buildout tracker. Refinancing announcements as the 2026 to 2028 loans come due, and whether the rates get better or worse. And the H100 rental rate, because it is the value of the collateral under everything above.