Neocloud Brief · neocloud stocks · what is a neocloud · neocloud ETF · what is a neocloud
What is a neocloud? GPU clouds vs hyperscalers, explained
A neocloud rents high-end GPUs for AI work — and undercuts the hyperscalers by roughly half. The interesting part isn’t the discount. It’s the utilization cliff that decides who survives it.
01The 30-second read
Fast read
A neocloud is a cloud provider that specializes almost exclusively in renting high-end GPUs — GPU-as-a-Service — for AI training and inference, rather than selling general-purpose cloud.
They exist because of an arbitrage: the same Nvidia H100 that lists for $6.88–$12.29 per GPU-hour on hyperscaler on-demand instances rents for roughly $2.40–$6.16 on neoclouds. Median on-demand H100 pricing runs about $4.17/hr on dedicated clouds vs $7.89/hr on hyperscalers — an ~89% premium.
That discount is the product. But the business is brutally capital-intensive: you buy the GPUs, finance them with debt, and then need roughly 75% utilization just to cover costs. On a 1,024-GPU H100 cluster, the gap between 55% and 85% utilization is −$330k vs +$340k per month.
02The definition, plainly
A neocloud — short for “new cloud” — is a cloud provider built around accelerated compute. Its customers are AI labs and enterprises buying GPU capacity: cluster-hours for training runs, capacity for inference at scale. The company owns or contracts the GPUs, the data-center capacity, and usually the orchestration software on top.
Three things it is not:
Not a hyperscaler
- AWS, Azure, and Google Cloud sell hundreds of services and treat GPUs as one product line among many
- A neocloud’s entire business is the GPU
Not a colocation REIT
- A colo landlord rents space and power
- A neocloud rents compute — and carries the GPU’s depreciation and obsolescence risk
And not a GPU marketplace: brokers like Vast.ai and RunPod-style pools rent idle capacity from many owners. A neocloud owns the fleet and signs the contract.
03Neocloud vs hyperscaler vs colocation
| Hyperscaler | Neocloud | Colocation | |
|---|---|---|---|
| Sells | Everything: compute, storage, databases, networking, AI services | GPU capacity + software | Space, power, cooling |
| GPU price | Premium $6.88–12.29/hr (H100) | Discount ~$2.40–6.16/hr (H100) | N/A — customer brings hardware |
| Hardware risk | Owned, spread across a vast fleet | Owned, concentrated in AI silicon | Customer’s |
| Who it serves | Everyone | AI labs, AI-native startups, enterprises | Anyone needing a facility |
| Where value sits | Breadth + ecosystem lock-in | Scarcity + price/performance + software | Real estate + power contracts |
04The pricing gap, in numbers
| GPU | Neocloud (dedicated) | Hyperscaler (on-demand) | Note |
|---|---|---|---|
| H100 | $2.43–$6.16 /GPU-hr | $6.88–$12.29 /GPU-hr | Median ~$4.17 vs ~$7.89 — roughly an 89% premium |
| B200 | Median ~$6.11 /GPU-hr ($3.44 low, Vast.ai) | Up to $16.11 /GPU-hr (Google Cloud) | Roughly double H100 pricing |
Ranges compiled from published rate cards and pricing indices (Spheron, Silicon Data, IntuitionLabs, Thunder Compute, VESSL). Per-GPU-hour, on-demand; rates move constantly.
Why the gap exists: hyperscalers carry the overhead of a general-purpose platform and price GPUs as a premium product. Neoclouds strip that down — fewer services, narrower support, purpose-built networking — and pass the savings to customers whose workloads are pure compute. The trade is price for specialization.
05How a neocloud actually makes money
The revenue formula is short. Everything dangerous lives in the last two terms.
1 · Utilization is a cliff, not a slope
- GPUs earn only when rented — but depreciate whether or not anyone is using them
- 1,024-GPU H100 cluster: 55% utilization ≈ −$330k/month; 85% ≈ +$340k/month
- Rule of thumb: about 75% utilization to cover costs; bare-metal gross margins 55–65% before depreciation
2 · Depreciation is the quiet risk
- Nvidia’s architecture cadence runs 18–24 months (Hopper → Blackwell → Rubin)
- CoreWeave extended its accounting depreciation from four years to six — flattering reported earnings vs. how fast the hardware loses competitive value
3 · It’s financed with debt
- More than $20B of sector loans are collateralized by Nvidia accelerators
- When collateral is depreciating hardware and revenue comes from a few large customers, lenders watch utilization as closely as backlog
4 · The backlog is the counterargument
- Contracted multi-year obligations give unusual visibility; sector backlogs run into the $100B+ range
- Full-stack neoclouds can monetize ~$7M–$13M per megawatt — several times what a bare landlord earns for the same power
06The players
The public universe splits into three layers — cloud (CoreWeave, Nebius), campus/power (Applied Digital, IREN, Hut 8, TeraWulf, Cipher Mining), and the supply side (memory, silicon).
See the full list with contract anchors at neocloud stocks, and the deep pages for CRWV, IREN, NBIS, and DRAM.
07What to watch
Bull signals
- Utilization holding at or above the ~75% cost-coverage line as new capacity lands
- Contract pricing firming — spot softened in 2026 while contract and Blackwell prices rose
- Backlogs converting on schedule: contracted megawatts becoming live megawatts
- Financing still available at reasonable rates against GPU collateral
Bear signals
- Utilization slipping as capacity floods in faster than demand
- Depreciation catching up with reality — writedowns or shortened useful lives
- GPU collateral values falling faster than debt amortizes
- Customer concentration: one renegotiation becomes a repricing
- Hyperscaler capex cuts, or more efficient models reducing compute intensity
08FAQ
What is a neocloud in simple terms?
A cloud company that rents high-end GPUs for AI work. Instead of buying Nvidia hardware and building a data center yourself, you rent capacity by the hour from a specialist — usually for about half of what a hyperscaler charges.
Is a neocloud the same as a hyperscaler?
No. Hyperscalers sell breadth (hundreds of services, GPUs as one line item). Neoclouds sell one thing: accelerated compute, plus the software to run it efficiently.
Why are neoclouds cheaper?
They strip out the general-purpose platform overhead and specialize the hardware, networking, and support around AI workloads. The trade is price for specialization — less breadth, fewer services, tighter focus.
Do neoclouds make money?
At the margin, yes — bare-metal gross margins run 55–65% before depreciation. At the bottom line, mostly not yet: heavy capex, debt service, and depreciation outrun revenue while the fleet is still being built. Utilization and depreciation schedules decide who becomes durably profitable.
Do neocloud stocks depend on Nvidia?
Heavily. Nvidia supplies the GPUs, sets the architecture cadence that determines how fast hardware ages, and in some cases invests in or backstops its own customers — a circularity worth understanding before investing.
Sources & method
- Pricing: published rate cards and third-party indices as reported Sep 2026 — Spheron (GPU cloud pricing comparison 2026), Silicon Data (B200/H100 index), IntuitionLabs (data-center GPU pricing index), Thunder Compute (AI GPU rental market trends), VESSL (hyperscalers vs neoclouds), GPUSmith (H100/H200/B200 rental prices)
- Unit economics: American Compute / ModulEdge (utilization math: 55% vs 85% on a 1,024-GPU H100 cluster; 55–65% bare-metal gross margins before depreciation), The Diligence Stack (business models; ~$7M–$13M revenue per MW), Zettabyte (debt-financed GPU cloud unit economics), Aethir (>$20B of sector loans collateralized by Nvidia accelerators), Luminix (depreciation schedules; 18–24 month architecture cadence)
- Market size: Synergy Research Group (>$25B neocloud revenue in 2025)
- All figures are third-party claims, approximate, and point-in-time; not independently verified by us
Get the brief
Occasional notes on the AI infrastructure buildout — new pages, corrections, and the events that actually move the story. No spam, unsubscribe in one click.