Can I reserve a multi-node cluster?

Yes. Talk to our team about reserved capacity for multi-node runs; on-demand instances remain available for bursts.

Which GPU should I train on?

H100 SXM nodes are the default for most training. H200 and B200 add memory capacity and bandwidth for long context and very large models.

Per second, with a 60-second minimum. Reserved capacity is billed at the reserved rate for the term.

LLM training

Train foundation models without waiting for a cluster

From a single 8-GPU node to multi-node runs on a 3.2 Tbps InfiniBand fabric — reserved when you need guaranteed capacity, on demand when you are still finding the recipe.

Launch a GPU Talk to an engineer

Recommended start

H100 SXM ×8 Node $19.92/hr

3.2 Tbps

InfiniBand per node in a training cluster

8× H100

NVLink GPUs per HGX node

180 GB

HBM3e on a single B200

The hard parts

What makes this workload expensive

The bill is rarely about the GPU price alone. These are the problems we design around.

Capacity you can actually get

Large runs stall on procurement: quotes, waitlists, and multi-year minimums before the first token is trained.

Interconnect that keeps GPUs busy

All-reduce traffic grows with model size. A slow fabric turns expensive GPUs into idle ones.

Failures at scale

The more nodes in a job, the more often one of them fails. Lost hours come from restarts, not from compute.

How Reviosa handles it

Built for the shape of the job

HGX nodes with NVLink

8× H100 SXM nodes with NVLink inside the node, plus single B200 and H200 instances for memory-hungry layers and experiments.

3.2 Tbps InfiniBand

Nodes in a training cluster share a non-blocking InfiniBand fabric, so NCCL collectives scale across nodes.

Checkpoints on NVMe

Block volumes attach in seconds and survive instance restarts — write sharded checkpoints locally, sync to object storage.

Reserved or per-second

Reserve capacity for the long run; burst extra nodes on demand and pay only for the seconds they run.

Recommended hardware

Where to start, where to scale

H100 SXM ×8 Node

Start here

640 GB VRAM · 8× NVIDIA H100

$19.92/hr

B200 SXM

180 GB VRAM · 1× NVIDIA B200

$5.79/hr

H200 SXM

141 GB VRAM · 1× NVIDIA H200

$4.39/hr

GPU Cloud Storage & Networking

Playbook

From first run to production

1
Prototype the recipe on a single H100 or H200 instance.
2
Scale to one 8× H100 node and tune throughput per GPU.
3
Reserve a multi-node cluster for the full run; keep sharded checkpoints on NVMe volumes.
4
Release the capacity the moment the run finishes — per-second billing stops with it.