What is a neocloud? GPU clouds beyond the hyperscalers
A neocloud is a cloud provider built around GPUs for AI. What they sell, how they differ from AWS, the numbers behind CoreWeave and Nebius, and the risks.
A neocloud is a cloud provider built mainly to rent out GPUs and other accelerators for AI training and inference, rather than the broad catalogue of services a hyperscaler like AWS, Microsoft Azure or Google Cloud sells. Forrester describes neoclouds as "AI infrastructure cloud platforms purpose-built to deliver high-performance infrastructure for genAI", with a primary focus on GPUs and accelerators. CoreWeave and Nebius are the best-known listed examples.
Below: what neoclouds actually sell, the published numbers of the two largest, how they differ from a hyperscaler, and what to check before you sign with one.
What a neocloud sells: GPUs, networking and capacity
A hyperscaler sells hundreds of services, and GPU instances are one shelf among many. A neocloud turns that around. Its product is the accelerator itself, and everything else exists to keep the GPUs busy:
- GPU capacity first. Whole servers or large slices of them, usually with the newest NVIDIA data centre cards, often on bare metal or a thin virtualisation layer so the customer gets close to full performance.
- Fast networking between servers. Training a large model spreads work across hundreds or thousands of GPUs, so the network between machines matters as much as the cards.
- Power and cooling sized for GPUs. Racks are planned per kilowatt. Our own 8 × RTX PRO 6000 server alone draws about 6.5 kW with all eight GPUs under full load, measured at the wall.
- Long contracts. The largest customers reserve capacity for years. This is how neoclouds finance the hardware: the contract comes first, then the debt, then the data centre.
The word is not a formal category, and lists of who counts differ between analysts. GPU marketplaces that rent single cards by the hour, and small providers that rent out whole servers, are often grouped in as well.
The two largest, by their own filings
The numbers below come from the companies' own reports to investors and the US Securities and Exchange Commission.
CoreWeave (USA). In its IPO prospectus, filed on 3 March 2025, CoreWeave said that at the end of 2024 its 32 data centres ran more than 250,000 GPUs on more than 360 MW of active power. For the second quarter of 2026 it reported revenue of $2,575 million, up from $1,212 million a year earlier, a net loss of $626 million, a revenue backlog of about $104 billion, 1.5 GW of active power and about 3.7 GW of contracted power. In April 2026, DatacenterDynamics reported a further contract with Meta worth about $21 billion, running to December 2032.
Nebius (headquartered in Amsterdam). For the second quarter of 2026 Nebius Group reported revenue of $582.3 million, up from $105.1 million a year earlier, of which $574.9 million came from its AI cloud. Its loss from continuing operations was $190.4 million. Its first data centre is in Mäntsälä, Finland, which the company announced in October 2024 it would triple to 75 MW.
Two things stand out. Revenue grows very fast, and so do losses and debt, because the GPUs are bought years before the contracts pay for them. That trade-off is the business model, not a side effect.
How a neocloud differs from AWS, Azure and Google Cloud
The same GPU can cost very different amounts depending on who rents it out. On 12 October 2026, AWS listed its g7e.48xlarge instance, with 8 × RTX PRO 6000 Blackwell Server Edition, at $33.14 an hour on demand in US East. On 15 October 2026, RunPod's pricing table listed the same card at $2.49 per GPU-hour, or $19.92 an hour for eight. We compared these and other paths in detail in AI server cost in 2026.
The hyperscaler price is not simply a markup. It buys things a neocloud usually does not offer:
| Hyperscaler | Neocloud | |
|---|---|---|
| What it sells | Hundreds of services: databases, storage, identity, analytics, GPUs | GPU compute, storage and networking around it |
| GPU access | Virtual machines inside the provider's cloud | Often bare metal or thin VMs |
| Regions | Many, worldwide | Fewer, sometimes one or two countries |
| Typical customer | Companies already running everything on that cloud | AI labs, model builders, teams that need GPUs and little else |
If your data, identity management and compliance setup already live in AWS, moving the GPU work elsewhere has a real cost in engineering time. If you only need GPUs, paying for the rest of the cloud is the expensive part.
What to check before you rent from one
Some of the risks are visible in the largest neocloud's own filings, and the rest are simple questions to ask any provider:
- Customer concentration. CoreWeave's prospectus said its top two customers made up about 77% of its 2024 revenue. A provider that depends on a few large tenants can change priorities quickly when one of them grows or leaves.
- Debt and capital spending. GPUs are bought with borrowed money against future contracts. Ask how long the provider has been running and who owns the hardware you will use.
- Where the servers are. "European company" and "servers in Europe" are not the same thing. Ask for the data centre, the country and the company that operates it.
- Shared or dedicated. Many GPU clouds share a physical server between customers. If your data or model weights are sensitive, ask whether you get the whole machine and who else has access to it.
- What happens at the end. Ask how you get your data out, how fast, and whether egress traffic is charged.
Where a small neocloud in the EU fits
The large neoclouds are built for customers who reserve thousands of GPUs for years. Most teams need far less: one machine for a private model, a fine-tuning job, or inference for a product. For that, the choice is mostly about the size of the job and how long it runs.
| If you... | The kind of provider that usually fits | Why |
|---|---|---|
| need thousands of GPUs for training | A large neocloud (CoreWeave, Nebius and similar) | They build and contract capacity at that scale |
| run everything in one hyperscaler already | That hyperscaler's GPU instances | Identity, networking and compliance stay in one place |
| need GPUs for hours or days, or are testing | An hourly GPU cloud or marketplace | You pay only while it runs |
| need one 8-GPU machine for months, in the EU, not shared | A dedicated GPU server, such as ours near Sofia | Whole machine, root access, fixed monthly price from $1.70 per GPU-hour, about $9,930 a month |
NEOXIS is one of the small ones: a single 8 × RTX PRO 6000 server with 2 × EPYC 9754, 1.5 TB of RAM, 2 × 4 TB NVMe and a 1 Gbps unmetered uplink, in a data centre near Sofia, Bulgaria. The limits are honest ones: one location, one configuration, and no hyperscaler services around it. If that fits, the pricing page has the terms.
Questions
Is a neocloud the same as a GPU cloud? In practice, mostly yes. "GPU cloud" describes the product; "neocloud" is the industry term for the new companies built around it, as opposed to the hyperscalers that added GPUs to an existing cloud.
Are neoclouds cheaper than AWS? For the GPU alone, published prices usually are, as the RunPod and AWS figures above show. For a whole system that also needs managed databases, identity and many regions, the comparison is less simple.
Is RunPod a neocloud? Lists differ. RunPod, Lambda and similar GPU clouds are often included, but there is no official register, so treat any list as one analyst's view.
Can a single server be a neocloud? The term usually describes companies at data centre scale. A single dedicated server is better described as dedicated GPU hosting, which is what we offer.
Sources
All pages were checked on 15 October 2026 unless stated otherwise.
- Forrester: Introducing Neocloud and NeoPaaS: the next frontiers of the AI-native cloud, Charlie Dai and Lee Sustar, 12 August 2026.
- CoreWeave: Form S-1 registration statement, filed 3 March 2025, data centres, GPUs, active power and top-two-customer share for 2024.
- CoreWeave: CoreWeave Reports Strong Second Quarter 2026 Results, 11 August 2026.
- DatacenterDynamics: Meta signs $21bn AI infrastructure contract with CoreWeave, April 2026.
- Nebius Group: second quarter 2026 results, Form 6-K exhibit 99.1, August 2026.
- Nebius: Nebius to triple capacity at Finland data center to 75 MW, October 2024.
- RunPod: pricing, RTX Pro 6000 96 GB at $2.49 per hour in the on-demand table.
- AWS: EC2 G7e instances and the EC2 on-demand price list, g7e.48xlarge, Linux, US East (N. Virginia), checked on 12 October 2026.
- NEOXIS: pricing and server specifications; power draw measured on our server.
