What is a neocloud, and how is it different from a hyperscaler?
A neocloud is a cloud provider built specifically to rent out GPUs for AI training and inference. Examples include CoreWeave, Lambda, Nebius, Crusoe and Nscale. Compared with a hyperscaler such as AWS, Microsoft Azure or Google Cloud, a neocloud offers a narrower product, mainly GPU servers, high-speed clusters, storage and Kubernetes or Slurm, usually at a lower price per GPU-hour and with newer GPUs available sooner.
Where neoclouds fit among GPU suppliers
GPU capacity reaches buyers through four kinds of supplier:
| Supplier | What it sells | Typical contract | Managed services |
|---|---|---|---|
| Hyperscaler | GPU VMs inside a full cloud: databases, identity, networking, managed AI services | On-demand, savings plans, capacity reservations | Hundreds |
| Neocloud | GPU VMs and bare metal, InfiniBand clusters, storage, managed Kubernetes or Slurm | On-demand for some GPUs, reserved clusters, multi-year contracts | A few, focused on AI |
| GPU marketplace | Capacity from many hosts behind one interface | Mostly on-demand, some reserved | Few or none |
| Bare-metal GPU host | Physical GPU servers rented by the month | Monthly or longer | None |
The GPU marketplace vs managed GPU Kubernetes post covers marketplaces in detail.
The word “neocloud” has no formal definition. Network World describes neoclouds as specialized cloud providers built entirely around high-performance GPUs, and the term is used loosely for any GPU-first provider, from large public companies to small regional clouds.
What you get from a neocloud
- GPUs first. The newest NVIDIA systems, such as B300 and GB300 NVL72, often appear at neoclouds before they are broadly available on demand at hyperscalers. See where to rent NVIDIA B300 for a dated snapshot.
- Cluster networking. Reserved clusters typically come with InfiniBand or RoCE between nodes for multi-node training.
- A smaller platform. Expect compute, block and object storage, networking and a managed Kubernetes or Slurm service. Expect fewer managed databases, identity integrations and compliance services than at a hyperscaler.
- Contract terms. Much of the capacity is sold as reserved clusters with minimum terms, sizes and sometimes prepayment.
How prices compare
Price is the main reason teams choose a neocloud. Uptime Institute, as reported by Network World in June 2025, found an average on-demand price of $98 per hour for an 8-GPU DGX H100 instance at hyperscalers against $34 at neoclouds, about 66% less. AWS cut its P5 on-demand prices by up to 45% shortly after, which narrowed the gap but did not close it. In Cloudfleet’s snapshot of public list prices on 4 November 2026, an H100 cost $6.88 per GPU-hour at AWS and $11.06 at Google Cloud, against $2.45 to $4.89 at the GPU clouds listed. The full, dated table is on the GPU Kubernetes page.
Compare totals, not hourly rates. Storage, egress, support and the CPU and memory bundled with each GPU differ between providers.
Hyperscalers rent from neoclouds too
Neoclouds do not only compete with hyperscalers, they also sell to them. In September 2025, Nebius announced a multi-year agreement to provide dedicated GPU capacity to Microsoft from a new data center in Vineland, New Jersey, reported at about $17.4 billion through 2031. Large AI labs and hyperscalers lock up much of the newest capacity this way, which is one reason cloud GPUs are hard to get on demand.
Risks to weigh
- Counterparty risk. Many neoclouds are young, growing fast and financed with debt. A multi-year prepaid contract depends on the provider still operating at the end of it.
- Concentration. A few very large customers can account for much of a neocloud’s capacity and revenue, which affects how much capacity smaller customers can get.
- Fewer services. Anything the neocloud does not offer, such as a managed database or a particular identity integration, you run yourself or keep at another provider.
- Location and jurisdiction. Check which legal entity operates the data center and where support staff sit, not only the region name. For EU options, see where to rent GPUs in Europe.
- Lock-in. Proprietary APIs and storage formats make it harder to move. Standard Kubernetes and portable storage keep the exit open, as covered in how to avoid cloud vendor lock-in.
How to compare offers
- GPU model, memory per GPU and the price unit: per GPU or per 8-GPU node.
- Interconnect: NVLink inside the node, InfiniBand or RoCE between nodes, and bandwidth per GPU.
- Storage throughput for datasets and checkpoints, and egress charges.
- Term, minimum size, prepayment, delivery date and what happens if delivery slips.
- Hardware failure: how fast a failed GPU is replaced and whether downtime extends the term.
- How the nodes are handed over: VMs in an account, bare metal with SSH, or a managed cluster.
Whether to rent at all, or to buy GPUs, is a separate calculation covered in should you buy or rent GPUs.
Neocloud GPUs on Cloudfleet
Cloudfleet joins GPU nodes from neoclouds such as Lambda and Nebius, from hyperscalers and from your own servers into one managed Kubernetes cluster, so workloads keep the same manifests wherever the GPUs are. The GPU Kubernetes page lists current providers, and the Cloudfleet team helps source reserved capacity that you contract directly with the provider.