H200 vs GH200
The GH200 rents for 32% less per hour today — $3.02 against $4.45 — with 96 GB against 141 GB of memory.
Prices collected 2026-09-18 · 2 providers price both
Index median per GPU-hour, on-demand, per snapshot.
Side by side
| H200 | GH200 | |
|---|---|---|
| Median rental, per GPU-hour | $4.45 | $3.02 |
| Cheapest listing | $2.00 | $1.67 |
| Per month at 720 h | $3,204 | $2,174 |
| Providers pricing it | 35 | 5 |
| Memory | 141 GB HBM3e | 96 GB HBM3 |
| Memory bandwidth | 4,800 GB/s | 4,000 GB/s |
| GB of memory per $/hr | 31.7 | 31.8 |
| Architecture | Hopper | Grace Hopper |
| Class | Datacenter | Datacenter |
| Released | 2023-11 | 2023-08 |
Same provider, both GPUs
The cleanest comparison: 2 providers price both cards, so the difference is the GPU and nothing else. The H200 is cheaper at 1 of them.
CoreWeaveWhat each one runs, and what a million tokens costs
Weights + KV cache at 8K context + overhead against 90% of each card's memory. Where a model needs more than one card the cell says how many and prices the set at today's median. $/M tokens is the hourly median divided by estimated single-stream output, from memory bandwidth — an estimate for ranking, not a benchmark.
| Model | Bits | H200 | GH200 |
|---|---|---|---|
| Kimi K3 | 4-bit | 13× GPU$57.85/hr | 19× GPU$57.38/hr |
| Kimi K3 | 16-bit | 45× GPU$200.25/hr | 65× GPU$196.30/hr |
| DeepSeek V4.1-Flash | 4-bit | 3× GPU$13.35/hr | 4× GPU$12.08/hr |
| DeepSeek V4.1-Flash | 16-bit | 9× GPU$40.05/hr | 13× GPU$39.26/hr |
| Llama 4 Maverick | 4-bit | 2× GPU$8.90/hr | 3× GPU$9.06/hr |
| Llama 4 Maverick | 16-bit | 7× GPU$31.15/hr | 10× GPU$30.20/hr |
| gpt-oss-120b | 4-bit | ✓ 67.9 GB~1339 tok/s · $0.92/M | ✓ 67.9 GB~1115 tok/s · $0.75/M |
| gpt-oss-120b | 16-bit | 2× GPU$8.90/hr | 3× GPU$9.06/hr |
| Llama 3.1 70B | 4-bit | ✓ 44.5 GB~98 tok/s · $12.61/M | ✓ 44.5 GB~81 tok/s · $10.36/M |
| Llama 3.1 70B | 16-bit | 2× GPU$8.90/hr | 2× GPU$6.04/hr |
| Gemma 4 31B | 4-bit | ✓ 21.2 GB~220 tok/s · $5.62/M | ✓ 21.2 GB~184 tok/s · $4.56/M |
| Gemma 4 31B | 16-bit | ✓ 65.7 GB~62 tok/s · $19.94/M | ✓ 65.7 GB~52 tok/s · $16.13/M |
| Qwen3 Coder 30B-A3B | 4-bit | ✓ 19 GB~2069 tok/s · $0.6/M | ✓ 19 GB~1724 tok/s · $0.49/M |
| Qwen3 Coder 30B-A3B | 16-bit | ✓ 62.9 GB~582 tok/s · $2.12/M | ✓ 62.9 GB~485 tok/s · $1.73/M |
| Qwen3.8 27B | 4-bit | ✓ 18.7 GB~253 tok/s · $4.89/M | ✓ 18.7 GB~211 tok/s · $3.98/M |
| Qwen3.8 27B | 16-bit | ✓ 57.6 GB~71 tok/s · $17.41/M | ✓ 57.6 GB~59 tok/s · $14.22/M |
| gpt-oss-20b | 4-bit | ✓ 13.7 GB~1896 tok/s · $0.65/M | ✓ 13.7 GB~1580 tok/s · $0.53/M |
| gpt-oss-20b | 16-bit | ✓ 43.9 GB~533 tok/s · $2.32/M | ✓ 43.9 GB~444 tok/s · $1.89/M |
| Llama 3.1 8B | 4-bit | ✓ 6.9 GB~853 tok/s · $1.45/M | ✓ 6.9 GB~711 tok/s · $1.18/M |
| Llama 3.1 8B | 16-bit | ✓ 18.4 GB~240 tok/s · $5.15/M | ✓ 18.4 GB~200 tok/s · $4.19/M |
Buy or rent?
Hours of renting at today's median that add up to the card's approximate list price. List prices are typical quoted figures, not tracked, and do not include the server, power or the person to run it — so the break-even is a floor, not a verdict.
Questions
Is the H200 or the GH200 cheaper to rent?
On 2026-09-18 the H200 has a market median of $4.45 per GPU-hour across 35 providers and the GH200 $3.02 across 5, so the GH200 is 32% cheaper per hour. Among the 2 providers that price both, the H200 is cheaper at 1 and the GH200 at 1. Per hour is only half the answer — the table below prices a whole model on each.
Which should I rent for a 70B model?
At 4-bit, Llama 3.1 70B needs about 44.5 GB. On the H200 that is one card at $4.45/hr; on the GH200 one card at $3.02/hr. Estimated single-stream decode is ~98 vs ~81 tokens per second, from memory bandwidth.
What is the real difference between them?
141 GB HBM3e at 4,800 GB/s (Hopper) against 96 GB HBM3 at 4,000 GB/s (Grace Hopper). Memory decides what fits; bandwidth decides how fast it generates, because each token streams the active weights once. Compute matters for prompt processing and training, not for how fast text appears.