Workhorse
HGX H100
The HGX H100 is the workhorse: the cheapest way to get a whole 8-GPU node on a dedicated 3.2 Tbps InfiniBand fabric, at $15 per node-hour on a yearly term.
Specification
- GPU memory
- 80 GB HBM3 · 3.35 TB/s
- Intra-node
- NVLink 4 · 900 GB/s
- Inter-node
- 8×400G IB NDR · 3.2 Tbps
- Node config
- 8 GPU · 15 TB NVMe · 2 TB RAM
- Best for
- Cost-efficient training ≤1K GPUs
H100 pods run 8×400 Gb/s InfiniBand NDR per node — 3.2 Tbps — rail-optimized, non-blocking, 1:1 on every tier, with SHARP in-network reduction.
Why this machine
For training runs up to roughly a thousand GPUs, H100 remains the best cost per useful FLOP — provided the fabric is real. That qualifier is where most quotes differ: an H100 node behind an oversubscribed spine is a slower machine than an H100 node on a 1:1 non-blocking fat-tree, and the price list rarely says which one you are buying.
Ours says. Every H100 pod runs 8×400 Gb/s NDR per node with dedicated leaf switches, and every cluster ships with a 72-hour burn-in report — NCCL all-reduce sweeps, HBM bandwidth, thermals under sustained load — that you sign off before you pay.
What it is for
Cost-efficient training up to ~1K GPUs
The price-performance floor for real training work, on a fabric specified in writing rather than implied.
Research blocks and fine-tuning
Whole nodes on short reserved terms with per-minute metering inside the reservation and datasets that persist between runs.
Steady-state inference
Models that fit in 80 GB, served without virtualization jitter and with 20 TB of included egress per node-month.
HGX H100 questions
How much does an HGX H100 node cost?
From $15 per 8-GPU node-hour on a yearly reservation, or from $22 monthly. That is the whole node: 8 GPUs, 15 TB of NVMe and 2 TB of RAM.
Why is this cheaper than an on-demand H100 elsewhere?
It is a reserved term rather than on-demand, and it is priced per node rather than per GPU. There is also nothing bolted on: $0 ingress, $0 inter-node traffic, 20 TB of egress included per node-month, and managed Slurm or Kubernetes at $0.
Is the InfiniBand shared with other tenants?
No. Nodes, leaf switches and storage lanes are physically dedicated per tenant at 1:1 oversubscription. There is no blended east-west fabric and no shared spine.
What is the smallest H100 cluster you sell?
Reservations start at 128 GPUs and scale past 10,000, always in whole 8-GPU nodes. Fractional GPUs would require a hypervisor, which is the thing we removed.