GH200 vs H100 SXM
The GH200 rents for 14% less per hour today — $3.02 against $3.50 — with 96 GB against 80 GB of memory.
Prices collected 2026-09-18 · 5 providers price both
Index median per GPU-hour, on-demand, per snapshot.
Side by side
| GH200 | H100 SXM | |
|---|---|---|
| Median rental, per GPU-hour | $3.02 | $3.50 |
| Cheapest listing | $1.67 | $1.68 |
| Per month at 720 h | $2,174 | $2,516 |
| Providers pricing it | 5 | 46 |
| Memory | 96 GB HBM3 | 80 GB HBM3 |
| Memory bandwidth | 4,000 GB/s | 3,350 GB/s |
| GB of memory per $/hr | 31.8 | 22.9 |
| Architecture | Grace Hopper | Hopper |
| Class | Datacenter | Datacenter |
| Released | 2023-08 | 2022-09 |
Same provider, both GPUs
The cleanest comparison: 5 providers price both cards, so the difference is the GPU and nothing else. The GH200 is cheaper at 2 of them.
| Provider | GH200 | H100 SXM | H100 SXM premium |
|---|---|---|---|
 Vultr | $1.67 | $3.50 | +109% |
 Lambda | $2.29 | $4.14 | +81% |
| $3.02 | $2.64 | -13% | |
 Denvr Dataworks | $3.87 | $2.45 | -37% |
 CoreWeave | $6.50 | $6.16 | -5% |
What 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 | GH200 | H100 SXM |
|---|---|---|---|
| Kimi K3 | 4-bit | 19× GPU$57.38/hr | 22× GPU$76.89/hr |
| Kimi K3 | 16-bit | 65× GPU$196.30/hr | 78× GPU$272.61/hr |
| DeepSeek V4.1-Flash | 4-bit | 4× GPU$12.08/hr | 5× GPU$17.48/hr |
| DeepSeek V4.1-Flash | 16-bit | 13× GPU$39.26/hr | 16× GPU$55.92/hr |
| Llama 4 Maverick | 4-bit | 3× GPU$9.06/hr | 4× GPU$13.98/hr |
| Llama 4 Maverick | 16-bit | 10× GPU$30.20/hr | 12× GPU$41.94/hr |
| gpt-oss-120b | 4-bit | ✓ 67.9 GB~1115 tok/s · $0.75/M | ✓ 67.9 GB~934 tok/s · $1.04/M |
| gpt-oss-120b | 16-bit | 3× GPU$9.06/hr | 4× GPU$13.98/hr |
| Llama 3.1 70B | 4-bit | ✓ 44.5 GB~81 tok/s · $10.36/M | ✓ 44.5 GB~68 tok/s · $14.28/M |
| Llama 3.1 70B | 16-bit | 2× GPU$6.04/hr | 3× GPU$10.49/hr |
| Gemma 4 31B | 4-bit | ✓ 21.2 GB~184 tok/s · $4.56/M | ✓ 21.2 GB~154 tok/s · $6.3/M |
| Gemma 4 31B | 16-bit | ✓ 65.7 GB~52 tok/s · $16.13/M | ✓ 65.7 GB~43 tok/s · $22.58/M |
| Qwen3 Coder 30B-A3B | 4-bit | ✓ 19 GB~1724 tok/s · $0.49/M | ✓ 19 GB~1444 tok/s · $0.67/M |
| Qwen3 Coder 30B-A3B | 16-bit | ✓ 62.9 GB~485 tok/s · $1.73/M | ✓ 62.9 GB~406 tok/s · $2.39/M |
| Qwen3.8 27B | 4-bit | ✓ 18.7 GB~211 tok/s · $3.98/M | ✓ 18.7 GB~176 tok/s · $5.52/M |
| Qwen3.8 27B | 16-bit | ✓ 57.6 GB~59 tok/s · $14.22/M | ✓ 57.6 GB~50 tok/s · $19.42/M |
| gpt-oss-20b | 4-bit | ✓ 13.7 GB~1580 tok/s · $0.53/M | ✓ 13.7 GB~1323 tok/s · $0.73/M |
| gpt-oss-20b | 16-bit | ✓ 43.9 GB~444 tok/s · $1.89/M | ✓ 43.9 GB~372 tok/s · $2.61/M |
| Llama 3.1 8B | 4-bit | ✓ 6.9 GB~711 tok/s · $1.18/M | ✓ 6.9 GB~596 tok/s · $1.63/M |
| Llama 3.1 8B | 16-bit | ✓ 18.4 GB~200 tok/s · $4.19/M | ✓ 18.4 GB~168 tok/s · $5.78/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 GH200 or the H100 SXM cheaper to rent?
On 2026-09-18 the GH200 has a market median of $3.02 per GPU-hour across 5 providers and the H100 SXM $3.50 across 46, so the GH200 is 14% cheaper per hour. Among the 5 providers that price both, the GH200 is cheaper at 2 and the H100 SXM at 3. 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 GH200 that is one card at $3.02/hr; on the H100 SXM one card at $3.50/hr. Estimated single-stream decode is ~81 vs ~68 tokens per second, from memory bandwidth.
What is the real difference between them?
96 GB HBM3 at 4,000 GB/s (Grace Hopper) against 80 GB HBM3 at 3,350 GB/s (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.