B200 vs MI355X
The B200 rents for 23% less per hour today — $6.59 against $8.60 — with 180 GB against 288 GB of memory.
Prices collected 2026-09-18 · 1 providers price both
Index median per GPU-hour, on-demand, per snapshot.
Side by side
| B200 | MI355X | |
|---|---|---|
| Median rental, per GPU-hour | $6.59 | $8.60 |
| Cheapest listing | $3.75 | $8.60 |
| Per month at 720 h | $4,745 | $6,192 |
| Providers pricing it | 26 | 1 |
| Memory | 180 GB HBM3e | 288 GB HBM3e |
| Memory bandwidth | 8,000 GB/s | 8,000 GB/s |
| GB of memory per $/hr | 27.3 | 33.5 |
| Architecture | Blackwell | CDNA 4 |
| Class | Datacenter | Datacenter |
| Released | 2024-12 | 2025-06 |
Same provider, both GPUs
The cleanest comparison: 1 provider price both cards, so the difference is the GPU and nothing else. The B200 is cheaper at 0 of them.
| Provider | B200 | MI355X | MI355X premium |
|---|---|---|---|
 Oracle Cloud | $14.00 | $8.60 | -39% |
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 | B200 | MI355X |
|---|---|---|---|
| Kimi K3 | 4-bit | 10× GPU$65.90/hr | 7× GPU$60.20/hr |
| Kimi K3 | 16-bit | 35× GPU$230.65/hr | 22× GPU$189.20/hr |
| DeepSeek V4.1-Flash | 4-bit | 2× GPU$13.18/hr | 2× GPU$17.20/hr |
| DeepSeek V4.1-Flash | 16-bit | 7× GPU$46.13/hr | 5× GPU$43.00/hr |
| Llama 4 Maverick | 4-bit | 2× GPU$13.18/hr | ✓ 228 GB~669 tok/s · $3.57/M |
| Llama 4 Maverick | 16-bit | 5× GPU$32.95/hr | 4× GPU$34.40/hr |
| gpt-oss-120b | 4-bit | ✓ 67.9 GB~2231 tok/s · $0.82/M | ✓ 67.9 GB~2231 tok/s · $1.07/M |
| gpt-oss-120b | 16-bit | 2× GPU$13.18/hr | ✓ 236.1 GB~627 tok/s · $3.81/M |
| Llama 3.1 70B | 4-bit | ✓ 44.5 GB~163 tok/s · $11.23/M | ✓ 44.5 GB~163 tok/s · $14.66/M |
| Llama 3.1 70B | 16-bit | ✓ 145.1 GB~46 tok/s · $39.79/M | ✓ 145.1 GB~46 tok/s · $51.93/M |
| Gemma 4 31B | 4-bit | ✓ 21.2 GB~367 tok/s · $4.99/M | ✓ 21.2 GB~367 tok/s · $6.51/M |
| Gemma 4 31B | 16-bit | ✓ 65.7 GB~103 tok/s · $17.77/M | ✓ 65.7 GB~103 tok/s · $23.19/M |
| Qwen3 Coder 30B-A3B | 4-bit | ✓ 19 GB~3448 tok/s · $0.53/M | ✓ 19 GB~3448 tok/s · $0.69/M |
| Qwen3 Coder 30B-A3B | 16-bit | ✓ 62.9 GB~970 tok/s · $1.89/M | ✓ 62.9 GB~970 tok/s · $2.46/M |
| Qwen3.8 27B | 4-bit | ✓ 18.7 GB~421 tok/s · $4.35/M | ✓ 18.7 GB~421 tok/s · $5.67/M |
| Qwen3.8 27B | 16-bit | ✓ 57.6 GB~119 tok/s · $15.38/M | ✓ 57.6 GB~119 tok/s · $20.07/M |
| gpt-oss-20b | 4-bit | ✓ 13.7 GB~3160 tok/s · $0.58/M | ✓ 13.7 GB~3160 tok/s · $0.76/M |
| gpt-oss-20b | 16-bit | ✓ 43.9 GB~889 tok/s · $2.06/M | ✓ 43.9 GB~889 tok/s · $2.69/M |
| Llama 3.1 8B | 4-bit | ✓ 6.9 GB~1422 tok/s · $1.29/M | ✓ 6.9 GB~1422 tok/s · $1.68/M |
| Llama 3.1 8B | 16-bit | ✓ 18.4 GB~400 tok/s · $4.58/M | ✓ 18.4 GB~400 tok/s · $5.97/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 B200 or the MI355X cheaper to rent?
On 2026-09-18 the B200 has a market median of $6.59 per GPU-hour across 26 providers and the MI355X $8.60 across 1, so the B200 is 23% cheaper per hour. Among the 1 providers that price both, the B200 is cheaper at 0 and the MI355X 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 B200 that is one card at $6.59/hr; on the MI355X one card at $8.60/hr. Estimated single-stream decode is ~163 vs ~163 tokens per second, from memory bandwidth.
What is the real difference between them?
180 GB HBM3e at 8,000 GB/s (Blackwell) against 288 GB HBM3e at 8,000 GB/s (CDNA 4). 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.