GH200
The GH200 is a cloud GPU with 96 GB of memory and 700 W TDP. It is currently available from 2 providers starting at $2.29/hr, with a market median of $2.29/hr across 4 offerings.
Best deal
Hardware specifications
- Architecture
- Hopper
- VRAM
- 96 GB
- Memory type
- HBM3
- Memory bandwidth
- 4,000 GB/s
- FP16
- 989 TFLOPS
- FP32
- 67 TFLOPS
- TDP
- 700 W
Same across all providers.
Market comparison
All-time price history
Offerings
| Provider | Cloud | Count | vCPU | RAM | Region | Per GPU hr | SPOT | Total/hr | |
|---|---|---|---|---|---|---|---|---|---|
HardwareHQ 0.0% 30dReferral link | — | ×1 | — | 432 GB | — | $2.29/hr | — | $2.29/hr | Launch ↗ |
GPU Finder 0.0% 30d | — | ×1 | — | — | dulles-usa-3 | $2.29/hr | — | $2.29/hr | Launch ↗ |
GPU Finder 0.0% 30d | — | ×1 | — | — | dulles-usa-3 | $2.29/hr | — | $2.29/hr | Launch ↗ |
GPU Finder 0.0% 30d | — | ×1 | — | — | us-default, us-east-3 | $2.29/hr | — | $2.29/hr | Launch ↗ |
LLMs that fit
| Model | Params | 16-bit | 8-bit | 4-bit |
|---|---|---|---|---|
GLM-5.2 Z.ai | 753B | Needs >8 GPUs | Needs >8 GPUs | Barely fits on 8× |
DeepSeek R1 DeepSeek | 671B 37B active | Not released | Needs >8 GPUs | 7× · 14.1% spare |
DeepSeek V4 Flash DeepSeek | 284B | Not released | Not released | 3× · 57.6 GB free |
Solar Open2 250B Upstage | 250B 15B active | Barely fits on 8× | 5× · 84 GB free | 3× · 57.6 GB free |
Qwen3 235B-A22B Alibaba | 235B 22B active | Barely fits on 8× | Barely fits on 4× | 3× · 70.8 GB free |
Laguna-S 2.1 Poolside | 118B 8B active | Barely fits on 4× | Barely fits on 2× | 2× · 84 GB free |
gpt-oss-120b OpenAI | 117B 5.1B active | Not released | Not released | Barely fits on 1× |
Llama 3.1 70B Meta | 70B | Barely fits on 2× | Barely fits on 1× | 1× · 42 GB free |
Qwen3.6 35B-A3B Alibaba | 35B 3B active | 2× · 88.8 GB free | 1× · 39.6 GB free | 1× · 62.4 GB free |
Qwen3 32B Alibaba | 32.8B | Barely fits on 1× | 1× · 44.4 GB free | 1× · 66 GB free |
Qwen2.5 14B Alibaba | 14B | 1× · 60.6 GB free | 1× · 76.7 GB free | 1× · 84.6 GB free |
Llama 3.1 8B Meta | 8B | 1× · 75.1 GB free | 1× · 84.6 GB free | 1× · 89.3 GB free |
Estimates: model weights plus 20% for KV cache and runtime overhead, at short context lengths. Long contexts and large batches need more. Rows marked “tight” hold the weights but not that full margin. A ×N figure is the total VRAM across N of these GPUs and makes no claim about interconnect throughput.
Similar GPUs
See more GPUs like this in
Frequently asked questions
How much does the GH200 cost per hour?
The cheapest GH200 offering starts at $2.29/hr, and the market median is $2.29/hr.
Which cloud providers offer the GH200?
We track GH200 offerings from HardwareHQ, GPU Finder.
How much VRAM does the GH200 have?
The GH200 has 96 GB of VRAM.
What LLMs can run on the GH200?
With 96 GB of VRAM, the GH200 can run 14 of our tracked models in 4-bit quantization, including GLM-5.2, DeepSeek R1, DeepSeek V4 Flash. 10 models fit in 16-bit precision.
What's the biggest LLM I can run on multiple GH200s?
With 8× GH200 (768 GB total), the largest tracked model that fits in 16-bit is Qwen3.6 35B-A3B (35B params), which needs 2× GPUs.
Where is the GH200 cheapest?
The cheapest GH200 offering in our catalog is HardwareHQ at $2.29/hr per GPU-hour (×1).