← All hardware
The Hopper workhorse that powered the LLM boom, now widely available
Pros
- Widely available, no allocation wait
- Mature CUDA software stack
- Price has dropped sharply
- Proven at scale
Cons
- Two generations behind Blackwell
- 80GB limits largest models
- Less efficient per watt than newer parts
- Being phased out of new builds
✓ Where it shines / best for
- Mainstream LLM training and fine-tuning
- Production generative AI inference
- Enterprise AI and HPC clusters
✕ Not the best fit for
- Frontier trillion-parameter training where Blackwell racks excel
- Edge or on-device inference
- Hobbyist budgets
Features
- ✓ AI inference
- ✓ Data-center scale
- ✓ NVLink
- ✓ LLM Training
- ✓ FP8
- ✓ HBM3
- ✓ Transformer Engine
- ✓ Mig
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| H100 80GB (street price) | $25,000-$30,000 | one-time | Per-GPU street price for SXM/PCIe via OEMs and resellers |
| DGX H100 system (8x H100) | $250,000-$400,000 | one-time | 8-GPU appliance via NVIDIA partners |
| Cloud rental (per H100) | ~$2.00-$4.00 | per hour | On-demand neocloud rate; reserved/spot below $2/hr in 2026 |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| use | Data center training and inference |
| power | 350-700W |
| memory | 80GB HBM3, 3.35 TB/s |
| performance | ~4 PFLOPS FP8 (sparse) |
| architecture | Hopper (TSMC 4N) |
Sponsored
A full review is being generated for this product and will appear here shortly.