Clusters
Sixteen to five hundred and twelve GPUs. Same per-card price.
Eight-card nodes on a 3.2 Tbps InfiniBand fabric, handed over as Slurm or Kubernetes, at exactly the per-card price printed on the catalogue. No cluster premium, no minimum term beyond a month, no sales cycle: tell us the size and the card, we answer with a date.
- 3.2 Tbps InfiniBand per node
- Full NVLink inside every node
- Slurm, Kubernetes or bare nodes
- 4 regions
Sizing
Per month, per cluster, at cost.
Cards × the catalogue price. The fabric, the head node and provisioning are in it. Twelve-month terms are 10 % lower.
| Node | 16 GPUs2 nodes | 32 GPUs4 nodes | 64 GPUs8 nodes | 128 GPUs16 nodes | 256 GPUs32 nodes | 512 GPUs64 nodes |
|---|---|---|---|---|---|---|
| 8× B3002304 GB per node · NVLink 5 · 1.8 TB/s | $28,192$25,373 on 12 mo | $56,384$50,746 on 12 mo | $112,768$101,491 on 12 mo | $225,536$202,982 on 12 mo | $451,072$405,965 on 12 mo | $902,144$811,930 on 12 mo |
| 8× B2001440 GB per node · NVLink 5 · 1.8 TB/s | $22,528$20,275 on 12 mo | $45,056$40,550 on 12 mo | $90,112$81,101 on 12 mo | $180,224$162,202 on 12 mo | $360,448$324,403 on 12 mo | $720,896$648,806 on 12 mo |
| 8× H2001128 GB per node · NVLink 4 · 900 GB/s | $20,032$18,029 on 12 mo | $40,064$36,058 on 12 mo | $80,128$72,115 on 12 mo | $160,256$144,230 on 12 mo | $320,512$288,461 on 12 mo | $641,024$576,922 on 12 mo |
| 8× H100 SXM640 GB per node · NVLink 4 · 900 GB/s | $18,144$16,330 on 12 mo | $36,288$32,659 on 12 mo | $72,576$65,318 on 12 mo | $145,152$130,637 on 12 mo | $290,304$261,274 on 12 mo | $580,608$522,547 on 12 mo |
| 8× MI300X1536 GB per node · Infinity Fabric · 896 GB/s | $12,512$11,261 on 12 mo | $25,024$22,522 on 12 mo | $50,048$45,043 on 12 mo | $100,096$90,086 on 12 mo | $200,192$180,173 on 12 mo | $400,384$360,346 on 12 mo |
Memory per node: B300 2,304 GB · B200 1,440 GB · H200 1,128 GB · H100 SXM 640 GB · MI300X 1,536 GB. A 671B MoE model at FP8 needs about 806 GB: four B300 nodes hold it with cache to spare, and so do eight H200 or B200 nodes.
What a cluster is made of
Nodes, fabric, head node. Nothing virtual.
The node
Eight SXM or OAM cards on one baseboard with the complete NVLink 4, NVLink 5 or Infinity Fabric domain. CPU cores, ECC RAM and local NVMe per card as on the catalogue. Dual 100 Gbit/s Ethernet to the Internet, egress free.
The fabric
Eight NDR InfiniBand ports per node, 400 Gbit/s each, 3.2 Tbps per node, non-blocking within a 64-node fabric. RDMA end to end; NCCL, SHARP and GPUDirect configured. Your cluster is alone on its partition.
The head node
A CPU node with the Slurm controller and login, or the Kubernetes control plane, and a shared NVMe-over-fabric volume if you want one. Root everywhere; nothing you cannot replace.
Slurm
slurmd 24.11 on every node, Pyxis and Enroot for containers, NCCL topology files, a job template that puts one rank per card. The stack our own benchmarks run on.
Kubernetes
k3s or an upstream control plane with the NVIDIA GPU Operator and the network operator for RDMA. Bring your own operators, or your own distribution on bare nodes.
The price
Cards × the per-card cost sheet. The fabric ports, the head node and the provisioning are in the operations and network lines. No cluster premium, because there is no cluster margin.
How it goes
Three steps, dates included.
01Tell us the shape
Card, number of cards, region, Slurm or Kubernetes or bare, and when you need it, through a ticket in the console titled “Cluster”, from the account that ordered the first node. We answer with a date and the exact monthly amount, which is the catalogue price times the cards.
02Top up, order
The first term is taken from your balance like any server; a cluster of 64 H100 for a month is $72,576. Nothing to sign; the cluster appears in your console as its nodes.
03Hand-over
Sixteen to sixty-four cards within five working days on an existing fabric; larger clusters on a dedicated fabric in two to four weeks. You get the login node, the topology files and an NCCL all-reduce benchmark result for your fabric.
Questions
About clusters.
How is a cluster priced?
At the per-card price of the catalogue, times the number of cards, with the same term discounts. The InfiniBand fabric, the head node and the provisioning are included. A 64-card H100 cluster is 64 × $1,134 = $72,576 per month on a one-month term, $65,318.40 on twelve months.
What does a cluster node look like?
An eight-card HGX node (or an eight-module OAM node for MI300X): the full NVLink or Infinity Fabric domain between its cards, eight 400 Gbit/s InfiniBand ports for 3.2 Tbps to the fabric, dual 100 Gbit/s Ethernet, local NVMe sized per card, CPU and RAM per card as on the catalogue.
Slurm or Kubernetes?
Either. Slurm comes with a login node, a controller, Pyxis and Enroot for containers, and NCCL tuned for the fabric. Kubernetes is k3s or a vanilla control plane with the NVIDIA GPU Operator and the network operator for RDMA. You can also take bare nodes with root and build your own.
How long does it take?
Sixteen to sixty-four cards on an existing fabric: within five working days. Larger clusters are placed on a dedicated fabric; two to four weeks depending on the region and the card. Order the first eight-card node from the console, then open a ticket titled “Cluster” with the size, the card, the region and the framework, and we answer with a date.
Can I start small and grow?
Yes, on the same fabric, as long as the region has the capacity, which is why we ask for the target size up front. Nodes are added at the same per-card price and the term of the new nodes can be aligned with the existing ones.
Is storage shared across the nodes?
Each node has local NVMe. For a shared file system we add an NVMe-over-fabric volume on the InfiniBand network at the extra-NVMe price per TB; for object storage, egress is free so any external store works.
Tell us the size. We answer with a date.
From 16 to 512 cards, at the price on the catalogue, in the region you pick.