Model guide · prices from 2026-09-10 · API prices daily, full survey weekly
Cheapest way to run Mixtral 8x22B (141B MoE)
Mixtral 8x22B (141B MoE) needs ≈85 GB of VRAM at 4-bit — this week that starts at $1.20/hour (3× RTX5090, cheapest tracked walk-up price). FP8 and FP16 below. VRAM calculators stop at the memory number; this connects it to live, graded prices across 20 providers.
MoE note: Mixtral 8x22B (141B MoE) activates ~39B active per token, but the full 141B weights must fit in VRAM — memory scales with total parameters, speed with active ones.
FP16 — ≈338 GB needed
| GPU | VRAM | GPUs needed | Cheapest tracked | Verifiable A/B |
|---|---|---|---|---|
| RTX5090 | 32 GB | 11 (multi-node) | $4.41/hr | — |
| A100-SXM | 80 GB | 5 | $4.54/hr | $11.50/hr |
| RTX4090 | 24 GB | 15 (multi-node) | $5.61/hr | — |
| A100-PCIe | 80 GB | 5 | $6.75/hr | — |
| MI300X | 192 GB | 2 | $6.90/hr | — |
| H100-PCIe | 80 GB | 5 | $9.95/hr | — |
| L40S | 48 GB | 8 | $10.96/hr | — |
| H200-SXM | 141 GB | 3 | $11.97/hr | $12.87/hr |
FP8/INT8 — ≈169 GB needed
| GPU | VRAM | GPUs needed | Cheapest tracked | Verifiable A/B |
|---|---|---|---|---|
| RTX5090 | 32 GB | 6 | $2.41/hr | — |
| A100-SXM | 80 GB | 3 | $2.72/hr | $6.90/hr |
| RTX4090 | 24 GB | 8 | $2.99/hr | — |
| MI300X | 192 GB | 1 | $3.45/hr | — |
| A100-PCIe | 80 GB | 3 | $4.05/hr | — |
| L40S | 48 GB | 4 | $5.48/hr | — |
| H100-PCIe | 80 GB | 3 | $5.97/hr | — |
| B200 | 192 GB | 1 | $6.00/hr | $7.15/hr |
4-bit — ≈85 GB needed
| GPU | VRAM | GPUs needed | Cheapest tracked | Verifiable A/B |
|---|---|---|---|---|
| RTX5090 | 32 GB | 3 | $1.20/hr | — |
| RTX4090 | 24 GB | 4 | $1.50/hr | — |
| A100-SXM | 80 GB | 2 | $1.81/hr | $4.60/hr |
| A100-PCIe | 80 GB | 2 | $2.70/hr | — |
| L40S | 48 GB | 2 | $2.74/hr | — |
| MI300X | 192 GB | 1 | $3.45/hr | — |
| L4 | 24 GB | 4 | $3.48/hr | — |
| H100-PCIe | 80 GB | 2 | $3.98/hr | — |
Method: VRAM ≈ parameters × bytes/parameter × 1.2 (20% headroom for KV cache and activations at ~8K context). Longer context or high concurrency needs more; training needs far more (optimizer states: ~4–8× weights). GPU counts assume even sharding. Above one 8-GPU node, the interconnect becomes the constraint — restrict to Grade A/B listings. Grades measure what sellers publish — deliverability of the written claim, not measured performance. No fleet holds a measured (Verified) badge yet; that tier arrives with the verification harness. Prices are this week's cheapest tracked walk-up listings (teaser prices excluded); every listing with grades and sources on the GPU pages. Interactive version for any model size: Model Fit.
These prices move weekly
The Delivered Compute Report tracks them — free, every row sourced.
Subscribe free