H100 · HOPPER · 8-GPU VM
$27.35/hr $3.42/GPU-hr · billed per minute
8× NVIDIA H100 80GB on a single VM — 640GB of pooled HBM, 208 vCPU and 880GB RAM. On-demand at $27.35/hr, Spot at $17.78/hr, and the meter runs per minute until you stop or delete the instance. Hopper full node — frontier fine-tunes.
| Model | 8-GPU VM | Per GPU |
|---|---|---|
| On-demand | $27.35/hr | $3.42/GPU-hr |
| Spot (35% off) | $17.78/hr | $2.22/GPU-hr |
| Per-minute meter | $0.4558/min | $0.2963/min on Spot |
| 24/7 for a month | $19,966/mo | 730 hours at the on-demand rate |
EU and APAC regions add a 15% data-residency premium; GLOBAL and US list at the base rate. Egress is metered per GB on the Standard or Premium network tier, and boot disk is billed separately.
GPUs
8× NVIDIA H100 80GB
VRAM
640GB pooled (80GB per GPU)
vCPU
208
RAM
880GB
Memory budget
640GB of pooled HBM across 8 80GB cards. Rule of thumb at FP8 — about one byte per parameter — 70B+ dense checkpoints fit with room for KV cache, and the node is positioned for frontier fine-tunes.
Typical loads
Multi-GPU fine-tunes, 70B+ dense serving and parallel eval harnesses. Fine on: torch, JAX, raw CUDA, axolotl, vLLM, sglang.
Billing math
$27.35/hr is $0.4558/min while the VM runs; the meter accrues per minute from launch until you stop or delete the instance. Spot bills $17.78/hr and stops automatically on preemption.
The 8× NVIDIA H100 80GB VM is $27.35/hr on-demand for the whole node — $3.42 per GPU-hour. Spot is 35% off at $17.78/hr ($2.22/GPU-hr). GLOBAL and US regions list at the base rate; EU and APAC add a 15% residency premium.
Per minute of runtime: $0.4558/min on-demand ($0.2963/min on Spot). Metering opens at launch and accrues until you stop or delete the instance, and your credit balance is checked before launch — an insufficient balance is refused up front.
Yes — it is in the live launch catalog at /console/vms, launchable On-Demand or on Spot, in GLOBAL / US / EU / APAC regions, with Ubuntu and CUDA base images, boot-disk sizing and local NVMe. Spot may be reclaimed; billing stops automatically when preemption is detected.
640GB of pooled HBM across 8 80GB cards, 208 vCPU and 880GB RAM. Rule of thumb at FP8 — about one byte per parameter — puts 70B+ dense checkpoints here with KV-cache headroom, and the catalog positions the node for frontier fine-tunes. Fine on torch, JAX, raw CUDA, axolotl, vLLM and sglang.