GB200 NVL72 / GB200 NVL72 providers
GB200 NVL72 providers
Who says they rent out GB200 NVL72, how each one sells it, and what to ask before you sign.
Providers offering GB200 NVL72
| Provider | Status | What they say | Source |
|---|---|---|---|
| CoreWeave | Generally available | First cloud with generally available GB200 NVL72 instances. Bare metal through CoreWeave Kubernetes Service, Quantum-2 InfiniBand at 400 Gb/s per GPU, launched in US-WEST-01. | coreweave.com Feb 2025 |
| Microsoft Azure | VM series | ND GB200 v6. Each VM is 4 GPUs and 2 Grace CPUs, 18 VMs per rack, 4 × 400 Gb/s InfiniBand per VM. | learn.microsoft.com checked 10 Nov 2026 |
| AWS | Generally available | EC2 P6e-GB200 UltraServers in 36- and 72-GPU sizes, GA 9 Jul 2025 in the Dallas Local Zone, sold through EC2 Capacity Blocks for ML. | aws.amazon.com Jul 2025 |
| Google Cloud | Generally available | A4X VMs on GB200 NVL72, generally available from 29 May 2025. 28.8 Tbps of non-blocking GPU-to-GPU traffic per system over RoCE. | cloud.google.com May 2025 |
| Oracle (OCI) | Live | Thousands of liquid-cooled GB200 NVL72 GPUs live on OCI and NVIDIA DGX Cloud, with Superclusters planned beyond 100,000 Blackwell GPUs. | blogs.nvidia.com Apr 2025 |
| Nebius | Generally available | Announced the first general availability of GB200 NVL72 capacity in Europe. | nebius.com Jun 2025 |
| Lambda | Deployed racks | Has installed GB200 NVL72 racks for its cloud, including a Supermicro-built rack drawing over 110 kW at a hydrogen-powered site in Mountain View. | datacenterdynamics.com checked 10 Nov 2026 |
Built from each provider's own pages and announcements, plus trade press where noted. It shows who sells the platform, not who has free racks on your dates. Availability changes weekly, and some of the best options are operators that don't advertise at all.
Hyperscalers vs neoclouds
Hyperscalers (Azure, AWS, Google Cloud, Oracle) sell GB200 inside their wider cloud. That's simpler if your data and contracts already live there. Large blocks usually go through account teams; AWS sells P6e-GB200 through Capacity Blocks in one Local Zone (AWS).
Neoclouds (CoreWeave, Nebius, Lambda and smaller operators) are built around GPU clusters. They tend to move faster on whole-rack reservations and bare-metal access. Best Neocloud keeps a wider directory.
Racks coming back on the market
GB200 NVL72 is the most widely deployed rack-scale Blackwell system. As the first one-year and eighteen-month contracts end, and as some tenants move up to GB300, whole racks become available again. Those rarely show up on a pricing page. They move through account teams and brokers, which is where we come in.
What to ask every provider
- Whole racks with the 72-GPU NVLink domain intact, or 4-GPU slices?
- Which fabric between racks (InfiniBand, RoCE, EFA), and is it non-blocking at my size?
- Exact start date per rack, and the remedy if delivery slips.
- Rack-level uptime commitment and how failed trays are swapped.
- Bare metal, Kubernetes or Slurm, and who manages what.
- Storage throughput and data egress terms.
Questions
Which cloud has GB200 NVL72?
Who was first with GB200 NVL72?
RackShore runs this site and may be paid a referral fee by the provider you choose. Buyers pay us nothing. Spot a mistake? Tell us.