NVIDIA H200
NVIDIA · Server
A Hopper-architecture data-centre GPU, the first with 141GB of HBM3e at 4.8 TB/s, built for large-model training and high-throughput serving.
Specifications
| VRAM | 141GB |
|---|---|
| Memory bandwidth | 4800 GB/s |
| Power draw | 700W |
| Type | accelerator |
An enterprise GPU, typically rented by the hour rather than bought. Verify current pricing.
NVIDIA H200: common questions
- What models can the NVIDIA H200 run?
- With 141GB of VRAM it can run models up to roughly 232B parameters at a 4-bit quantisation, or smaller models with more context. Use the hardware matrix for specifics; these figures are approximate.
- How much power does the NVIDIA H200 draw?
- About 700W under load, so pair it with a power supply that has real headroom.
Strengths
- 141GB holds very large models on one GPU, with room for long context and KV cache
- 4.8 TB/s bandwidth delivers throughput no consumer card approaches
- FP8 and NVLink for multi-GPU scaling in server nodes
Weaknesses
- Data-centre only: the SXM form factor needs an HGX baseboard, not a desktop
- A 700W draw demands server-grade power and cooling
- Cost puts it far outside individual budgets, so it is realistically rented
Entry last verified 18 August 2026. Specifications and especially prices change; verify before buying.