Model guide · prices from 2026-09-10 · API prices daily, full survey weekly
Cheapest way to run Llama 3.1 405B
Llama 3.1 405B needs ≈243 GB of VRAM at 4-bit — this week that starts at $3.21/hour (8× 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.
FP16 — ≈972 GB needed
| GPU | VRAM | GPUs needed | Cheapest tracked | Verifiable A/B |
|---|---|---|---|---|
| A100-SXM | 80 GB | 13 (multi-node) | $11.79/hr | $29.90/hr |
| RTX5090 | 32 GB | 31 (multi-node) | $12.43/hr | — |
| RTX4090 | 24 GB | 41 (multi-node) | $15.33/hr | — |
| A100-PCIe | 80 GB | 13 (multi-node) | $17.55/hr | — |
| MI300X | 192 GB | 6 | $20.70/hr | — |
| H100-PCIe | 80 GB | 13 (multi-node) | $25.87/hr | — |
| H200-SXM | 141 GB | 7 | $27.93/hr | $30.03/hr |
| L40S | 48 GB | 21 (multi-node) | $28.77/hr | — |
FP8/INT8 — ≈486 GB needed
| GPU | VRAM | GPUs needed | Cheapest tracked | Verifiable A/B |
|---|---|---|---|---|
| A100-SXM | 80 GB | 7 | $6.35/hr | $16.10/hr |
| RTX5090 | 32 GB | 16 (multi-node) | $6.42/hr | — |
| RTX4090 | 24 GB | 21 (multi-node) | $7.85/hr | — |
| A100-PCIe | 80 GB | 7 | $9.45/hr | — |
| MI300X | 192 GB | 3 | $10.35/hr | — |
| H100-PCIe | 80 GB | 7 | $13.93/hr | — |
| B300 | 288 GB | 2 | $15.00/hr | $35.60/hr |
| L40S | 48 GB | 11 (multi-node) | $15.07/hr | — |
4-bit — ≈243 GB needed
| GPU | VRAM | GPUs needed | Cheapest tracked | Verifiable A/B |
|---|---|---|---|---|
| RTX5090 | 32 GB | 8 | $3.21/hr | — |
| A100-SXM | 80 GB | 4 | $3.63/hr | $9.20/hr |
| RTX4090 | 24 GB | 11 (multi-node) | $4.11/hr | — |
| A100-PCIe | 80 GB | 4 | $5.40/hr | — |
| MI300X | 192 GB | 2 | $6.90/hr | — |
| B300 | 288 GB | 1 | $7.50/hr | $17.80/hr |
| H100-PCIe | 80 GB | 4 | $7.96/hr | — |
| H200-SXM | 141 GB | 2 | $7.98/hr | $8.58/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