H100 SXM vs H100 PCIe
The H100 PCIe rents for 18% less per hour today — $2.88 against $3.50 — same 80 GB of memory, 2,000 vs 3,350 GB/s.
Prices collected 2026-09-18 · 8 providers price both
Index median per GPU-hour, on-demand, per snapshot.
Side by side
| H100 SXM | H100 PCIe | |
|---|---|---|
| Median rental, per GPU-hour | $3.50 | $2.88 |
| Cheapest listing | $1.68 | $1.59 |
| Per month at 720 h | $2,516 | $2,071 |
| Providers pricing it | 46 | 9 |
| Memory | 80 GB HBM3 | 80 GB HBM2e |
| Memory bandwidth | 3,350 GB/s | 2,000 GB/s |
| GB of memory per $/hr | 22.9 | 27.8 |
| Architecture | Hopper | Hopper |
| Class | Datacenter | Datacenter |
| Released | 2022-09 | 2022-09 |
Same provider, both GPUs
The cleanest comparison: 8 providers price both cards, so the difference is the GPU and nothing else. The H100 SXM is cheaper at 0 of them.
| Provider | H100 SXM | H100 PCIe | H100 PCIe premium |
|---|---|---|---|
 RunPod Community Cloud | $2.69 | $1.99 | -26% |
 Vast.ai | $3.45 | $2.88 | -17% |
 RunPod Secure Cloud | $3.49 | $2.89 | -17% |
 Beam | $3.50 | $1.83 | -48% |
 Cyfuture AI | $3.56 | $2.99 | -16% |
 GPU.ai | $3.69 | $2.43 | -34% |
 io.net | $3.76 | $1.59 | -58% |
 Lambda | $4.14 | $3.29 | -21% |
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 | H100 SXM | H100 PCIe |
|---|---|---|---|
| Kimi K3 | 4-bit | 22× GPU$76.89/hr | 22× GPU$63.27/hr |
| Kimi K3 | 16-bit | 78× GPU$272.61/hr | 78× GPU$224.34/hr |
| DeepSeek V4.1-Flash | 4-bit | 5× GPU$17.48/hr | 5× GPU$14.38/hr |
| DeepSeek V4.1-Flash | 16-bit | 16× GPU$55.92/hr | 16× GPU$46.02/hr |
| Llama 4 Maverick | 4-bit | 4× GPU$13.98/hr | 4× GPU$11.50/hr |
| Llama 4 Maverick | 16-bit | 12× GPU$41.94/hr | 12× GPU$34.51/hr |
| gpt-oss-120b | 4-bit | ✓ 67.9 GB~934 tok/s · $1.04/M | ✓ 67.9 GB~558 tok/s · $1.43/M |
| gpt-oss-120b | 16-bit | 4× GPU$13.98/hr | 4× GPU$11.50/hr |
| Llama 3.1 70B | 4-bit | ✓ 44.5 GB~68 tok/s · $14.28/M | ✓ 44.5 GB~41 tok/s · $19.49/M |
| Llama 3.1 70B | 16-bit | 3× GPU$10.49/hr | 3× GPU$8.63/hr |
| Gemma 4 31B | 4-bit | ✓ 21.2 GB~154 tok/s · $6.3/M | ✓ 21.2 GB~92 tok/s · $8.68/M |
| Gemma 4 31B | 16-bit | ✓ 65.7 GB~43 tok/s · $22.58/M | ✓ 65.7 GB~26 tok/s · $30.73/M |
| Qwen3 Coder 30B-A3B | 4-bit | ✓ 19 GB~1444 tok/s · $0.67/M | ✓ 19 GB~862 tok/s · $0.93/M |
| Qwen3 Coder 30B-A3B | 16-bit | ✓ 62.9 GB~406 tok/s · $2.39/M | ✓ 62.9 GB~242 tok/s · $3.3/M |
| Qwen3.8 27B | 4-bit | ✓ 18.7 GB~176 tok/s · $5.52/M | ✓ 18.7 GB~105 tok/s · $7.61/M |
| Qwen3.8 27B | 16-bit | ✓ 57.6 GB~50 tok/s · $19.42/M | ✓ 57.6 GB~30 tok/s · $26.63/M |
| gpt-oss-20b | 4-bit | ✓ 13.7 GB~1323 tok/s · $0.73/M | ✓ 13.7 GB~790 tok/s · $1.01/M |
| gpt-oss-20b | 16-bit | ✓ 43.9 GB~372 tok/s · $2.61/M | ✓ 43.9 GB~222 tok/s · $3.6/M |
| Llama 3.1 8B | 4-bit | ✓ 6.9 GB~596 tok/s · $1.63/M | ✓ 6.9 GB~356 tok/s · $2.24/M |
| Llama 3.1 8B | 16-bit | ✓ 18.4 GB~168 tok/s · $5.78/M | ✓ 18.4 GB~100 tok/s · $7.99/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 H100 SXM or the H100 PCIe cheaper to rent?
On 2026-09-18 the H100 SXM has a market median of $3.50 per GPU-hour across 46 providers and the H100 PCIe $2.88 across 9, so the H100 PCIe is 18% cheaper per hour. Among the 8 providers that price both, the H100 SXM is cheaper at 0 and the H100 PCIe at 8. 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 H100 SXM that is one card at $3.50/hr; on the H100 PCIe one card at $2.88/hr. Estimated single-stream decode is ~68 vs ~41 tokens per second, from memory bandwidth.
What is the real difference between them?
80 GB HBM3 at 3,350 GB/s (Hopper) against 80 GB HBM2e at 2,000 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.