Model guide · prices from 2026-09-10 · API prices daily, full survey weekly
Cheapest way to run Llama 4 Maverick (400B MoE)
Llama 4 Maverick (400B MoE) needs ≈240 GB of VRAM at 4-bit — this week that starts at $2.72/hour (3× A100-SXM, 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: Llama 4 Maverick (400B MoE) activates ~17B active per token, but the full 400B weights must fit in VRAM — memory scales with total parameters, speed with active ones.
FP16 — ≈960 GB needed
| GPU | VRAM | GPUs needed | Cheapest tracked | Verifiable A/B |
|---|---|---|---|---|
| A100-SXM | 80 GB | 12 (multi-node) | $10.88/hr | $27.60/hr |
| RTX5090 | 32 GB | 30 (multi-node) | $12.03/hr | — |
| RTX4090 | 24 GB | 40 (multi-node) | $14.96/hr | — |
| A100-PCIe | 80 GB | 12 (multi-node) | $16.20/hr | — |
| MI300X | 192 GB | 5 | $17.25/hr | — |
| H100-PCIe | 80 GB | 12 (multi-node) | $23.88/hr | — |
| L40S | 48 GB | 20 (multi-node) | $27.40/hr | — |
| H200-SXM | 141 GB | 7 | $27.93/hr | $30.03/hr |
FP8/INT8 — ≈480 GB needed
| GPU | VRAM | GPUs needed | Cheapest tracked | Verifiable A/B |
|---|---|---|---|---|
| A100-SXM | 80 GB | 6 | $5.44/hr | $13.80/hr |
| RTX5090 | 32 GB | 15 (multi-node) | $6.02/hr | — |
| RTX4090 | 24 GB | 20 (multi-node) | $7.48/hr | — |
| A100-PCIe | 80 GB | 6 | $8.10/hr | — |
| MI300X | 192 GB | 3 | $10.35/hr | — |
| H100-PCIe | 80 GB | 6 | $11.94/hr | — |
| L40S | 48 GB | 10 (multi-node) | $13.70/hr | — |
| B300 | 288 GB | 2 | $15.00/hr | $35.60/hr |
4-bit — ≈240 GB needed
| GPU | VRAM | GPUs needed | Cheapest tracked | Verifiable A/B |
|---|---|---|---|---|
| A100-SXM | 80 GB | 3 | $2.72/hr | $6.90/hr |
| RTX5090 | 32 GB | 8 | $3.21/hr | — |
| RTX4090 | 24 GB | 10 (multi-node) | $3.74/hr | — |
| A100-PCIe | 80 GB | 3 | $4.05/hr | — |
| H100-PCIe | 80 GB | 3 | $5.97/hr | — |
| L40S | 48 GB | 5 | $6.85/hr | — |
| MI300X | 192 GB | 2 | $6.90/hr | — |
| B300 | 288 GB | 1 | $7.50/hr | $17.80/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