Neoclouds
What Does CoreWeave Do? The Neocloud Business Model in Plain English
CoreWeave buys accelerators, funds them with debt, and sells AI compute on multi-year contracts — why the model loses money while growing.
The Stack
Neoclouds — The Specialist GPU Layer
Status Tightening as of 25 Sep 2026: short AI contracts clear near $40m/MW while contracted power hits 4.2 GW. Who gets paid is the busy fleet; the stuck step is utilization, power, and concentration.
Neoclouds sit between scarce accelerators and builders who need clusters now. They own or finance GPU fleets, place them in power-ready halls, and sell capacity by the hour or by contract. This Stack layer page tracks who gets paid, what binds cash, and how that differs from hyperscalers.
Status: Tightening · Updated: 2026-09-22
Investor read: When specialist fleets reprice — as Nebius’s on-demand GPU rates do from 1 October 2026 — cash sticks with operators that keep utilization high and power delivered. The stuck step is utilization, power delivery, and customer concentration, not only chip scarcity upstream. Watch the clock on contracts versus spot.
On this page
Function
A neocloud is a specialist compute landlord. It buys or finances NVIDIA-class (and sometimes other) accelerators, stands them in leased or owned data halls, and sells cluster time to labs, startups, and enterprises that do not want to own the full stack.
That is different from a hyperscaler. A hyperscaler sells a platform — storage, networking, identity, managed services — and often builds for its own large models too. A neocloud’s product is mostly the GPU hour and the cluster contract. See Neocloud explained and Neocloud vs hyperscaler for the teaching cut; this page is the durable map of the layer.
In the AI build chain, neoclouds sit mid-stack:
- Upstream, semiconductors and packaging decide when chips can ship.
- Power and halls decide when those chips can light up — see US Markets on speed-to-power and small-MW AI labs.
- Neoclouds turn delivered megawatts and installed GPUs into billable capacity.
- Downstream, hyperscalers and large labs absorb a lot of that capacity — as customers and as competitors.
Who gets paid in this layer is whoever keeps the fleet busy at a price above the cost of chips, power, networking, and capital. Who waits is the buyer stuck on a waitlist, or the operator with dark racks and expensive debt.
Stuck step
The popular story says the bind is always “not enough GPUs.” Upstream scarcity still matters. But the cash bind inside the neocloud layer is usually one of three clocks:
1. Utilization. An idle GPU still burns depreciation, interest, and some power. Realized revenue per installed accelerator — not the rate card alone — decides whether the model works. Spot and short contracts move first; long backlog moves later.
2. Power and delivery. Chips without connected megawatts are inventory, not product. Contracted power and hall readiness are the long clock. When power slips, utilization and growth both slip.
3. Customer concentration. Large committed deals with a few lab or hyperscaler offtakers can fund growth — and can also concentrate renewal and delivery risk. Nebius’s Meta agreement (March 2026) and Microsoft capacity deal (September 2025) show the offtake path; they do not remove concentration risk.
Fresh print on the pricing clock: on 17 September 2026, Reuters reported Nebius will raise pay-as-you-go prices for selected NVIDIA GPUs from 1 October 2026 by about 17% to 21%, with some CPU instances up ~25% and some memory offerings ~41%. It is the second hike in roughly three months. The same day, CoreWeave said it continues to contract new compute at higher prices, including short-dated third-quarter contracts of about three to six months priced near $40 million per megawatt on an annualized basis, with contracted power around 4.2 GW as of 11 August 2026 (from ~3.7 GW at mid-year). Our Wire on Nebius’s October reprice reads that as the spread widening — capacity arriving while near-term prices still rise — not as demand cooling.
Caution: on-demand rate cards are not the whole backlog. Most large fleets still earn on multi-month or multi-year commitments. A hike in PAYG teaches scarcity at the margin. It does not, by itself, prove locked-in margins for every name.
So the stuck step, named plainly: keeping utilization high while power and halls arrive on time, without betting the fleet on one customer’s renewal. Chip scarcity is upstream. This layer’s money stuck step is utilization + delivery + concentration.
Role companies
Roles only — not a leaderboard, not a buy list.
| Role | Examples | What they do in this layer |
|---|---|---|
| Specialist GPU cloud / landlord | Nebius, CoreWeave | Own or finance fleets; sell on-demand and committed clusters; earn on utilization and contract price |
| Hall / hosting adjacent | Applied Digital (and similar) | Provide or partner data-hall capacity that neocloud demand needs to light up |
| Anchor offtakers (customers) | Meta, Microsoft (and other large labs) | Sign multi-year capacity that funds build-out — and concentrate who the fleet serves |
| Upstream suppliers | NVIDIA (and packaging / HBM chain) | Set when accelerators can ship; they are not the neocloud layer itself |
Always verify listing, ADR, and tax treatment for your own jurisdiction. Names here map roles. They are not recommendations.
Links
Stack desk
Explained
Wire
Markets (US doors)
Sibling Stack layers
FAQ
Is this a buy list?
No. This is a map of roles and constraints. Not investment advice. Do your own research.
What is a neocloud in one sentence?
A specialist operator that rents GPU clusters to customers who need AI compute without owning the full stack.
Why is status Tightening on 2026-09-22?
Because near-term pricing is rising while capacity is still being built: Nebius’s on-demand GPU reprice from 1 October 2026 (reported 17 Sep 2026) and CoreWeave’s higher short-dated contract pricing the same week. That is tighter economics for buyers and stronger prints for operators who keep fleets busy — until utilization or power slips.
How is this different from the hyperscaler layer?
Hyperscalers sell a broad cloud platform and often build for themselves. Neoclouds sell specialist GPU capacity. The stuck steps differ: platform share and own CapEx versus utilization, power delivery, and customer concentration. See neocloud vs hyperscaler.
What kills the neocloud model?
Chronic under-utilization, power or hall delays that leave GPUs dark, or offtakers moving committed spend back to hyperscaler capacity faster than specialists can refill the book.
Do on-demand price hikes prove the whole layer is safe?
No. PAYG teaches scarcity at the margin. Long contracts, financing cost, and delivery risk still decide who keeps the cash.
What would change the view
- Sustained on-demand and short-contract price declines across several large specialist fleets while installed capacity keeps rising — evidence demand no longer outruns delivery.
- Hyperscalers absorbing overflow so specialist utilization stays weak even when headline GPU demand looks firm.
- Power or chip delivery slips that leave new fleets dark for quarters, even with signed demand.
Sources
- Reuters, 17 Sep 2026 — Nebius PAYG GPU price hike from 1 Oct 2026; CoreWeave higher contract prices. Link
- CoreWeave / Business Wire, 17 Sep 2026 — ~$40m/MW short-dated Q3 contracts; ~4.2 GW contracted power. Link
- Nebius newsroom, 16 Mar 2026 — Meta AI infrastructure agreement (up to ~$27bn). Link
- Nebius prices page — public on-demand / commitment rate card (checked 22 Sep 2026). Link
- Nebius, 8 Sep 2025 — Microsoft multi-year AI infrastructure agreement. Link
Not investment advice. Do your own research.
Last updated: 2026-09-22 · Second Order by TKN