← All hardware
The memory-upgraded Hopper GPU that remains the AI workhorse.
Pros
- 76% more memory than H100 at similar power
- Broad availability across every cloud
- Strong price/performance for inference
Cons
- Hopper generation now superseded by Blackwell
- Lower FP4 capability than Blackwell
- Still expensive ($30k-$40k) per GPU
✓ Where it shines / best for
- Mainstream LLM inference and serving
- Fine-tuning and mid-scale training
- Enterprises upgrading existing H100 fleets
✕ Not the best fit for
- Frontier-scale training needing FP4 efficiency
- Teams wanting the latest generation performance
- Ultra-low-latency single-token inference niches
Features
- ✓ 141 GB HBM3e with 4.8 TB/s bandwidth
- ✓ 16,896 CUDA cores, 528 4th-gen Tensor Cores
- ✓ FP8 Transformer Engine, 3,958 TFLOPS FP8
- ✓ NVLink 4.0 at 900 GB/s per GPU
- ✓ Available in SXM and PCIe (NVL) form factors
- ✓ Drop-in compatible with H100 HGX systems
- ✓ Mature CUDA / TensorRT-LLM ecosystem
- ✓ 700W TDP, air- or liquid-coolable
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| H200 GPU (purchase) | $30,000-$40,000 | one-time | Street price per GPU outright |
| Cloud rental | $1.45-$13.78 | per GPU-hour | Wide range by provider/billing: budget neoclouds from ~$1.99-$2.50/hr, on-demand average ~$4.20/hr, up to ~$10-13.78/hr on hyperscalers; 8-GPU nodes ~$80-85/hr |
| DGX H200 system (8x SXM) | ~$350,000-$500,000 | one-time | Complete-system estimate (some 8-GPU listings quoted ~$315k) |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| power | 700W TDP |
| memory | 141 GB HBM3e, 4.8 TB/s |
| nvlink | 4.0, 900 GB/s per GPU |
| cuda_cores | 16,896 |
| architecture | Hopper (TSMC 4N) |
| tensor_cores | 528 (4th-gen) |
| fp8_performance | 3,958 TFLOPS |
Sponsored
A full review is being generated for this product and will appear here shortly.