The NVIDIA H100 is the workhorse of large-model training in 2026. With 80GB of HBM3, fourth-gen NVLink, and Transformer Engine acceleration, it remains the default choice for 70B-class pretraining and dense inference.
Specs
| Spec | Value |
|---|---|
| Architecture | Hopper |
| VRAM | 80 GB HBM3 |
| Memory bandwidth | 3.35 TB/s |
| FP16 (with sparsity) | 1,979 TFLOPS |
| TDP | 700W (SXM5) |
| Interconnect | 4th-gen NVLink 900 GB/s |
// LIVE_LISTINGS
NVIDIA H100 hosting from 0 providers
No live NVIDIA H100 listings right now — but we can source them.
Request a quoteNeed help sizing your NVIDIA H100 cluster?
Tell us your workload and we'll route matched offers from providers with real inventory — usually within 24 hours.
Get matched quotes