Best Neocloud / What is a neocloud?
What is a neocloud?
The GPU-first clouds behind most large AI training runs: what the word means, what they sell, who they are, and what to watch before you sign.
A neocloud is a cloud provider that specialises in renting GPUs for AI. Instead of hundreds of general services, it sells GPU compute: dedicated bare-metal servers, connected multi-node clusters, and the networking and storage that training and inference need. CoreWeave, Nebius, Lambda and Crusoe are the best-known examples.
Neocloud meaning
"Neo" means new. The word separates the GPU-first clouds that grew up during the AI build-out from the established hyperscalers: AWS, Microsoft Azure and Google Cloud. Oracle Cloud is usually counted with the hyperscalers, even though much of its recent growth is GPU capacity.
Three things make a provider a neocloud rather than just "a cloud with GPUs":
- GPUs are the product. Most revenue comes from renting accelerators, mainly NVIDIA, sometimes AMD.
- Built for big jobs. They sell whole connected clusters with InfiniBand or high-speed Ethernet fabric, often single-tenant, for months or years.
- Thin service layer. Expect Kubernetes, Slurm, storage and monitoring, not a catalog of databases, queues and app services.
What neoclouds sell
| Offer | What you get | Typical buyer |
|---|---|---|
| On-demand instances | One to eight GPUs by the hour, self-service | Developers, experiments, small fine-tunes |
| Instant clusters | Tens of GPUs, pre-connected, short terms | Startups running a training sprint |
| Reserved clusters | Hundreds to thousands of dedicated GPUs on one fabric, 6 months to 3+ years | AI labs and companies training or serving at scale |
| Rack-scale systems | GB300 NVL72 racks with 72 GPUs on one NVLink domain | Frontier training and large-model inference |
| Managed inference | Hosted model endpoints, billed per token or per GPU | Product teams that don't want to run servers |
The money is in the reserved clusters. Long contracts with AI labs and big tech let neoclouds borrow against future revenue to buy GPUs and lease data center power.
Neocloud examples
The names most buyers shortlist, with their current Blackwell Ultra status in our B300 directory:
- CoreWeave, the largest by revenue, publicly listed on Nasdaq (CRWV) since March 2025.
- Nebius, based in Amsterdam, listed on Nasdaq (NBIS), runs mostly its own data centers.
- Lambda, known for on-demand instances, 1-Click Clusters and GB300 NVL72 superclusters.
- Crusoe, which builds and runs its own AI data center campuses.
- Together AI, Fluidstack, Nscale, Firmus, TensorWave and many smaller GPU clouds. See the full list of neocloud companies.
Why neoclouds grew so fast
When demand for NVIDIA's H100 outran supply, the providers that could get GPUs racked, cabled and running quickest won the customers. Neoclouds focused on exactly that, while hyperscalers were also serving their own internal AI teams. Three things kept them growing:
- Speed to capacity. A neocloud can often deliver a dedicated cluster sooner than a hyperscaler can carve one out.
- Big anchor customers. Microsoft, Meta and OpenAI have all signed multi-billion-dollar capacity deals with neoclouds. For example, Microsoft agreed a capacity deal with Nebius on 9 September 2025 (Daloopa).
- Specialist operations. Running thousands of GPUs as one machine is a different job from running general cloud. Independent reviews such as SemiAnalysis ClusterMAX now rate neoclouds at or above hyperscalers on it.
The risks buyers should know
Neoclouds aren't risk-free. Before you sign a long contract, weigh:
- Customer concentration. Microsoft made up 62% of CoreWeave's 2024 revenue, falling to 45% in Q1 2026 (Daloopa). A provider that depends on one buyer can reprioritise fast.
- Debt and leases. Many neoclouds fund GPUs with debt and lease their data centers. Ask what happens to your cluster if the provider's financing changes.
- "Coming soon" capacity. Some quotes rely on sites still being built. Ask whether the exact hardware is running today. Our six-question checklist covers this.
How to choose one
Start with the system you need, the cluster size, the region and the start date. Then shortlist providers that can show that exact configuration running, and compare total cost over the term, not the hourly rate. If you'd rather not run that process yourself, send us one brief and we'll come back with the providers that fit.
Questions
Is CoreWeave a neocloud?
Is Oracle a neocloud?
What's the difference between a neocloud and a hyperscaler?
Which neoclouds are publicly traded?
What does neocloud mean in AI?
RackShore runs Best Neocloud and may be paid a referral fee by a provider. That doesn't change which providers are listed or how they're described. Spot a mistake? Tell us.