Who Are the Neoclouds? Specialist GPU Clouds, and Who Gets Paid

EXPLAINED · Players · Last reviewed: 23 Oct 2026 · Next review: after CoreWeave's third-quarter 2026 results (expected November 2026)

The neoclouds are specialist GPU clouds: companies that buy or finance AI accelerators, put them in power-ready halls, and get paid when customers run reserved or on-demand GPU hours — CoreWeave and Nebius among the listed names, Lambda, Crusoe, Nscale, Fluidstack and Lightning AI among the private ones.

In short. Neoclouds sell accelerator time, not a full public cloud. CoreWeave billed $2.575 billion in the April–June 2026 quarter from a 1.5 GW live fleet against about 3.7 GW contracted. Nscale's S-1 shows $103.4 billion of signed total contract value and only $2.6 billion active. Who gets paid first is Nvidia, landlords and lenders; the stuck step is the live hall.

Why an investor cares: hyperscalers set the AI build, but they also rent extra GPU hours when their own halls are late. That overflow is the neoclouds' business. If the word itself is new, start with what a neocloud is; this page is about who they are, how each one gets paid, and why a signed book is not live hours. Roles along the same bind sit on GPU cloud providers. The layer home is Neoclouds.

Who are the neoclouds?

There is no official list. The label is industry use for specialist GPU clouds, not a regulator's category. What joins them is concentration: almost all of the revenue is accelerator time. A hyperscaler sells storage, databases and software with GPUs as one line; a neocloud lives or dies on billed GPU hours. Neocloud vs hyperscaler sets out that split.

Sort them by how cash actually arrives and four jobs appear: the listed fleet that bills committed hours; the delayed-draw borrower that draws debt as halls switch on; the contracted-before-live signer, whose take-or-pay book waits on delivery; and the energy-first or lease-backstopped operator, which gets paid as a developer or tenant as much as a cloud.

Our classification. Company figures, dated. Not a ranking.

Those seven, plus smaller hour sellers, are the usual 2026 set. Names move; the jobs do not.

How do the listed names actually get paid?

CoreWeave is the largest listed print. For the quarter ended 30 June 2026 it reported revenue of $2.575 billion, up 112%, and a net loss of $626 million. Committed contracts were 98% of revenue. Three customers were 36%, 26% and 10% of that quarter — 72% from three buyers (our sum of the 10-Q shares). Remaining performance obligations were $103.7 billion as of 30 June 2026; the earnings release called revenue backlog about $104 billion, excluding more than $25 billion of early-Q3 commitments. Active power reached 1.5 GW; contracted power was about 3.7 GW. Live megawatts bill. Contracted megawatts are the runway.

Nebius is the other listed GPU specialist with a full sales line. Group revenue was $582.3 million in the same quarter, up 454%; AI cloud was $574.9 million, about 98% of the group. Annualised run-rate revenue was $3.0 billion as of the end of June. The CEO letter the same day put customer commitments at $40 billion, raised the year-end contracted-power target to 5 GW, and said 70% of Q2 deals included prepayments covering 50–60% of associated capex. The customer pays part of the hall before the hours run. Purchases of property and equipment were $5.657 billion in the quarter. See what TCV is and remaining performance obligation.

Why do some neoclouds look huge on paper?

Because the signed book and the live fleet are different clocks.

Nscale's Form S-1, filed 18 September 2026, is the clearest labelled split. As of 31 August 2026 it reported $103.4 billion of active and contracted total contract value and only $2.6 billion active — about 25,000 GPUs live against about 461,000 active or contracted. First-half 2026 revenue was $140.6 million. The contracts are mostly take-or-pay, with a weighted-average life of about 5.7 years, but terms start when clusters are delivered. Microsoft statements of work are up to $43.8 billion through December 2033; four Anthropic agreements signed 25 August 2026 are up to $44.6 billion. Those are caps, subject to delivery, acceptance and, for Anthropic, financing. $103.4 billion is not cash this year. Take-or-pay GPU contracts explains when offtake binds.

Lambda sits on the same bind from the debt side. On 3 November 2025 it announced a multibillion Microsoft agreement covering tens of thousands of Nvidia GPUs, including GB300 NVL72 systems. On 1 October 2026 it closed a $1.008 billion delayed-draw term loan to fund three committed deployments for two investment-grade offtakers. Draws follow commissioning. The facility is not live hours. GPU financing and who finances GPU clusters map that cheque.

Crusoe's 17 September 2026 Series F was $3.9 billion at a $30.9 billion post-money valuation, with more than $140 billion of TCV, more than 6 GW contracted and 1 GW delivered. On 28 September 2026 it said it is the developer of Google's Armstrong County, Texas campus, broken ground in June 2025. Fluidstack is the tenant version: TeraWulf's 18 August 2025 exhibit put Lake Mariner leases at about 360 MW of critical IT load and $6.7 billion of contracted revenue over the initial 10-year terms, with Google's backstop about $3.2 billion. Rent starts on completion.

Lightning AI closed a merger with Voltage Park on 21 January 2026 and said the combined company had grown from $18 million to over $500 million of ARR since 2024, with more than 35,000 owned H100, B200 and GB300 GPUs. Cash still sticks when those hours bill.

Who gets paid — and where is the stuck step?

Nvidia is paid first, when the servers are bought. Who buys AI GPUs shows the hyperscalers as the largest cheque; neoclouds are the specialist buyers beside them, often on the same Blackwell and Rubin racks.

Landlords and power providers are paid next. CoreWeave's live 1.5 GW against 3.7 GW contracted is a labelled example of halls still waiting to energise. Crusoe's 1 GW live against 6 GW+ contracted is the same gap on an energy-first model. Fluidstack's Google-wrapped leases pay the landlord from completion. See how a data-centre lease works.

Lenders are paid as delayed-draw facilities amortise against contracted cash flows. Lambda's $1.008 billion loan and CoreWeave's delayed-draw term loans — $13.6 billion outstanding as of 30 June 2026 — are that cheque. Interest is a real cost: CoreWeave's net interest expense was $640 million in the quarter.

The neocloud itself is paid last among those four, when reserved or on-demand hours run on a working cluster. GPU as a Service is the billing unit; GPU utilisation is whether the fleet actually earns.

The stuck step is the live hall. Signed TCV, RPO and contracted megawatts are visibility. They become revenue as clusters are accepted and powered. That is why Nscale can show $103.4 billion against $2.6 billion active, and why CoreWeave can show about $104 billion of backlog against $2.575 billion of quarterly sales. How long it takes to build an AI data center is the clock in years.

Who is not a neocloud?

Several kinds of company pick up the label by mistake.

Our classification. Not a ranking.

Oracle is the edge case the other way: a broad cloud whose AI growth looks like a handful of huge contracts. We count it as a hyperscaler. A large listed GPU cloud is still a neocloud if almost all of the revenue is accelerator time.

Common misconceptions

• People assume “neocloud” is an official list of stocks. Actually it is an industry label for specialist GPU clouds; names change, the job does not.
• People assume a $100 billion contract book is this year's sales. Actually TCV and RPO are signed work. Nscale's $103.4 billion includes $2.6 billion active; CoreWeave's ~$104 billion backlog sat beside $2.575 billion of quarterly revenue.
• People assume CoreWeave is a hyperscaler because it is large and listed. Actually 98% of its Q2 2026 revenue was committed GPU cloud. It does not sell a broad public cloud of storage and databases.
• People assume a delayed-draw loan is cash in the bank. Actually Lambda's $1.008 billion facility funds clusters as they commission. Undrawn capacity is not live hours.

FAQ

Who are the neoclouds?
Specialist GPU clouds that buy or finance AI accelerators, put them in power-ready halls, and get paid when customers run reserved or on-demand GPU hours. In 2026 the usual names are CoreWeave and Nebius among listed operators, and Lambda, Crusoe, Nscale, Fluidstack and Lightning AI among private ones. There is no official list.

Is CoreWeave a hyperscaler?
No. In the April–June 2026 quarter, 98% of its $2.575 billion of revenue came from committed GPU cloud contracts. A hyperscaler also sells storage, databases and software. CoreWeave does not.

How is a neocloud different from a hyperscaler?
A hyperscaler sells many services. A neocloud concentrates on accelerator hours, often for those same hyperscalers and for AI labs, and funds the fleet with delayed-draw debt, customer prepayments or leases. See Neocloud vs hyperscaler.

Why is signed capacity not cash?
TCV and remaining performance obligation count contracted work, not delivered hours. Nscale's S-1 shows $103.4 billion of active and contracted TCV and $2.6 billion active as of 31 August 2026. Terms start when clusters are delivered.

Who are the main neocloud companies in 2026?
CoreWeave, Nebius, Lambda, Nscale, Crusoe, Fluidstack and Lightning AI, plus smaller developer-first hour sellers. This is a roles map, not a market-share table.

Who buys from neoclouds?
Hyperscalers and AI labs. Microsoft is the named offtaker on Nebius, Lambda and Nscale prints. Anthropic is named on Nscale's August 2026 agreements. CoreWeave's 10-Q does not name buyers; three customers were 72% of Q2 2026 revenue (our sum).

Is this a list of stocks to buy or sell?
No. Second Order Explained pages explain who does what and where the money moves. They do not rank companies or give buy or sell calls. Not investment advice. Do your own research.

Related terms

Neocloud; neocloud vs hyperscaler; GPU cloud providers; take-or-pay GPU contracts; total contract value; GPU as a Service; who are the hyperscalers.

Markets and stack doors

Layer: Neoclouds. Buyer layer: Hyperscalers. Also: AI Data Centers. Markets: US, including CoreWeave and how neocloud contracts become cash.

Get the Wire — short notes when offtake, power and live fleets move.

How we check this

Figures are each company's own, dated in Sources. Combined customer shares are labelled “our sum”. Role labels are our classification. Re-checked 23 Oct 2026. Not investment advice. Do your own research.

Last reviewed: 23 Oct 2026 · Next review: after CoreWeave's third-quarter 2026 results (expected November 2026)

Spot an error? Tell us. Editorial standards →

Sources

1. CoreWeave — Q2 2026 results (11 Aug 2026) — $2.575B revenue; ~$104B backlog; 1.5 GW active / ~3.7 GW contracted
2. CoreWeave — Form 10-Q, quarter ended 30 Jun 2026 — 98% committed; customers 36%/26%/10%; RPO $103.7B; $13.6B delayed-draw outstanding
3. Nebius Group — Q2 2026 results (12 Aug 2026) — $582.3M revenue; $5.657B PPE purchases
4. Nebius Group — CEO letter, Ex. 99.2 (12 Aug 2026) — AI cloud $574.9M; ARR $3.0B; $40B commitments; 5 GW contracted-power target
5. Nscale Limited — Form S-1 (18 Sep 2026) — TCV $103.4B vs $2.6B active as of 31 Aug 2026; H1 revenue $140.6M
6. Lambda — Microsoft agreement (3 Nov 2025) — tens of thousands of Nvidia GPUs, including GB300 NVL72
7. Lambda — $1.008B delayed-draw loan (1 Oct 2026) — three committed deployments, two investment-grade offtakers
8. Crusoe — Series F (17 Sep 2026) — $3.9B raised; $140B+ TCV; 6 GW+ contracted; 1 GW live
9. Crusoe — Google Armstrong County campus (28 Sep 2026) — broke ground June 2025
10. TeraWulf — Ex. 99.1 (18 Aug 2025) — Fluidstack Lake Mariner ~360 MW; $6.7B contracted revenue; Google backstop ~$3.2B
11. Lightning AI — Voltage Park merger (21 Jan 2026) — over $500M ARR; 35,000+ owned GPUs

S

Sean

Writes Second Order by TKN — who gets paid, where money sticks, and which step is stuck.

Author page → · Editorial standards →