Neoclouds vs Hyperscalers: GPU Cloud Business Models Compared

Neoclouds — CoreWeave, Nebius, Lambda, Crusoe — are GPU-only cloud providers that rent Nvidia compute 40-85% below AWS, Azure, and Google Cloud list prices, but they finance that growth with debt and multi-year take-or-pay contracts rather than the diversified revenue and cash reserves that hyperscalers use. Both models sell GPU cycles; they differ in what else is on the menu, who is on the other side of the balance sheet, and how much counterparty risk a buyer takes on. Below is the data on pricing, financing, and where each model fits.

Key takeaways

• Neocloud on-demand H100 pricing runs about $4.17/hr median versus roughly $7.89/hr averaged across AWS, Azure, and Google Cloud — an 89% premium for hyperscaler-branded, identical Nvidia silicon, per Thunder Compute and GPUSmith trackers.
• The four largest hyperscalers plus Oracle are guiding to $660-690 billion in combined 2026 capital expenditure, roughly 77% above 2025’s ~$410 billion, per Futurum Group.
• CoreWeave’s total debt reached $35.6 billion by June 30, 2026, against FY2025 revenue of $5.13 billion — a debt-to-revenue ratio no hyperscaler carries against its cloud segment.
• Nebius signed $46 billion in combined capacity deals with Microsoft and Meta in 2026 and grew revenue 454% year over year to $582.3 million in Q2, with ARR reaching $3 billion.
• Synergy Research forecasts the neocloud market growing from roughly $25 billion in 2025 revenue to nearly $400 billion by 2031 — a 58% compound annual growth rate — even as hyperscalers grow their share of total data center capacity to 67% over the same period.
• The GPU depreciation debate (5-6 year hyperscaler schedules versus a 2-3 year real-world chip cycle) matters more for neoclouds, whose debt is often collateralized directly against the depreciating hardware.

What separates a neocloud from a hyperscaler

A hyperscaler — AWS, Microsoft Azure, Google Cloud — sells hundreds of managed services (databases, serverless functions, identity, compliance tooling) across dozens of global regions, with GPU instances as one line item among many. A neocloud sells almost nothing but GPU compute, high-bandwidth InfiniBand networking, and fast local storage, configured specifically for AI training and inference. That narrower scope is the entire cost advantage: less software overhead, thinner margins by design, and a sales model built around large, negotiated capacity commitments rather than self-serve billing.

The other structural difference is the balance sheet behind the compute. Hyperscalers fund GPU purchases from existing operating cash flow measured in the tens of billions per quarter. Neoclouds — with a handful of years of operating history and no other product line — fund the same GPU purchases with project debt, GPU-collateralized loans, and vendor financing, then race to convert that hardware into billable, contracted revenue before the next depreciation cycle and loan covenant test arrive.

The players in 2026

Provider	Type	Scale signal	Headline pricing
AWS / Azure / Google Cloud	Hyperscaler	$660-690B combined 2026 capex (top 5 incl. Oracle)	H100 on-demand $6.88-12.29/hr
CoreWeave	Neocloud	FY2025 revenue $5.13B (+168%); Q1 2026 revenue $2.08B (+112%)	H100 PCIe ~$4.25/hr; reserved ~$1.45/hr
Nebius	Neocloud	Q2 2026 revenue $582.3M (+454%); ARR $3B	Contract-priced; avg. deal yield >$20M/MW
Lambda	Neocloud	H100 SXM $3.99-4.29/hr with 99.9% uptime SLA	H100 from $2.49/hr (no egress fees)
Crusoe	Neocloud	~5 GW contracted AI infrastructure capacity (June 2026); $30B valuation	Build-to-suit and cloud, contract-priced

For a head-to-head on the mid-market neoclouds specifically, see our RunPod vs Lambda vs Vast.ai comparison and the live rate table on the GPU price index.

Pricing: the gap that defines the category

GPU / metric	AWS	Azure	Google Cloud	Neocloud (typical)
H100 on-demand	$6.88/hr	$12.29/hr	$10.98/hr	~$2.49-4.25/hr (Lambda, CoreWeave)
H100, reserved/committed	Lower with 1-3yr terms	Lower with 1-3yr terms	Lower with 1-3yr terms	~$1.45/hr (CoreWeave reserved)
Egress fees	Standard hyperscaler rates	Standard hyperscaler rates	Standard hyperscaler rates	Often $0 (Lambda, RunPod)
Service catalog	Hundreds of services	Hundreds of services	Hundreds of services	GPU compute, storage, networking only

The 89% average premium hyperscalers charge for the same Nvidia chip is not simply markup — it also prices in the surrounding platform, contractual SLAs backed by a much larger company, compliance certifications across more jurisdictions, and existing procurement relationships that make switching costs low for enterprise buyers already inside AWS, Azure, or Google Cloud. Whether that premium is worth paying is the central buyer question this guide works through below. Related context on the GPU rental side: our GPU cloud comparison of 15 providers and AI inference vs. training infrastructure guide.

How neoclouds finance hypergrowth

Neoclouds close the gap between “GPUs ordered” and “GPUs paying for themselves” with three overlapping mechanisms:

1. GPU-collateralized debt. Lenders accept the GPUs themselves as security. CoreWeave has roughly $7.5 billion of debt collateralized this way; the loan covenants are tied to assumptions about GPU resale and residual value that analysts expect could be tested as early as 2027 if depreciation runs faster than modeled, per Capacity Media.
2. Take-or-pay capacity contracts. A hyperscaler or AI lab commits to pay for reserved capacity regardless of utilization, which the neocloud then uses as bankable collateral. Nvidia itself backstops $6.3 billion of CoreWeave capacity this way; Nebius’s $27 billion Meta deal and $17-19.4 billion Microsoft deal, both running through roughly 2030-2031, work on similar logic.
3. Circular vendor financing. Nvidia has taken direct equity stakes in the neoclouds it also supplies chips to and sells capacity from (a $2 billion CoreWeave investment at $87.20/share, alongside the take-or-pay backstop) — a structure critics call circular financing, because the chip vendor’s revenue and its customer’s ability to pay for chips become linked.

Capex and balance sheet snapshot

Metric	Hyperscalers (combined, 2026)	CoreWeave	Nebius	Crusoe
2026 capex / capital raised	$660-690B (Microsoft, Amazon, Google, Meta, Oracle)	$35-39B capex guidance	Contracted revenue basis, not capex-disclosed	$3B+ raised Sept. 2026 at $30B valuation
Revenue signal	Cloud segments profitable, capex from operating cash flow	FY2025 $5.13B revenue; Q1 2026 $2.08B	Q2 2026 $582.3M revenue; ARR $3B	~5GW contracted AI infrastructure capacity
Debt load	Investment-grade, diversified across businesses	$35.6B total debt (as of June 30, 2026)	Financed via contracted revenue + equity	Equity-heavy so far
Anchor customer contracts	Internal (own AI products) plus external cloud customers	Nvidia $6.3B take-or-pay backstop	Meta $27B (5yr), Microsoft $17-19.4B	Microsoft (Abilene, TX, 900MW+)

Hyperscaler capex is funded largely from existing operating cash flow and spread across mature, profitable businesses, which is why rating agencies treat Microsoft, Amazon, and Google debt very differently from CoreWeave’s GPU-collateralized loans. For the financing mechanics in more depth, see our guide to how AI data centers get financed and, for the public-market angle, our REIT analysis (note: most GPU-focused neoclouds are not REITs — they are financed as leveraged infrastructure operators, not real estate).

The bear case: depreciation and concentration risk

The loudest structural risk to the neocloud model is accounting, not demand. Hyperscalers depreciate AI servers, including GPUs, over 5-6 years. Investor Michael Burry argues the real economic life is closer to 2-3 years given Nvidia’s product cadence, and that the gap understates industry-wide depreciation by an estimated $176 billion through 2026-2028, per CNBC’s coverage of the dispute. Hyperscalers can absorb a shorter real-world useful life inside diversified, profitable balance sheets. Neoclouds cannot as easily: debt collateralized against the GPUs themselves assumes today’s residual-value and useful-life estimates hold, and a faster write-down cycle would hit loan covenants directly.

The second risk is customer concentration. A large share of neocloud revenue sits inside a handful of take-or-pay contracts with a small number of hyperscaler and AI-lab counterparties (Nvidia, Microsoft, Meta, OpenAI). That is a feature when those counterparties are investment-grade and the contracts run five-plus years — it is also a single point of failure if any one of them pulls back capacity plans.

Reliability, compliance, and workload fit

CoreWeave and Nebius now deploy Nvidia’s latest Blackwell GB300 generation at scale, in some cases ahead of hyperscaler general availability in a given region, and both run dedicated infrastructure with published SLAs rather than best-effort marketplace capacity. That closes much of the reliability gap that used to separate “neocloud” from “hyperscaler” in a buyer’s risk assessment. What has not closed: the breadth of managed services, the number of compliance certifications held across jurisdictions, and the depth of enterprise support relationships that come from a hyperscaler account team already managing a customer’s non-AI workloads.

Market trajectory

Neocloud revenue reached roughly $9 billion in Q4 2025 alone (up 223% year over year) and exceeded $25 billion for the full year, with Synergy Research projecting growth to nearly $400 billion by 2031 — a 58% compound annual growth rate that assumes AI compute demand keeps outstripping hyperscaler supply. At the same time, hyperscalers are projected to grow their share of total worldwide data center capacity from 48% at the end of 2025 to 67% by 2031, since the same capital wave is funding hyperscaler-built AI regions alongside independent neocloud campuses. Both trends can be true together: neoclouds are growing off a small base inside a market that hyperscalers still dominate by absolute capacity.

What to do

Run the decision on three variables, not price alone. First, workload duration and switching cost: short-lived training runs and price-sensitive inference tolerate a newer, narrower-service counterparty better than a multi-year production deployment does. Second, counterparty durability: check whether the neocloud’s largest contracts are with investment-grade customers on take-or-pay terms (Nebius and CoreWeave both disclose this) versus shorter, more discretionary commitments. Third, total cost including the platform you are not buying — a hyperscaler premium of roughly 89% on raw GPU-hours partly funds compliance certifications, global regions, and managed services that a pure-play neocloud does not offer; price that gap against what your workload actually needs before deciding it is “just markup.” For capacity outside pure GPU rental — colocation, powered land, and build-to-suit — the broader data center catalog and colocation price index cover the underlying real estate and power side of the same capacity crunch, and our quote request tool routes to operators directly.