GPU price check · updated 2026-09-10
GB200-NVL72 rental prices
GB200-NVL72 rents for $10.50–$27.04 per GPU-hour on demand across 3 sellers this week. The spread is not noise — it is mostly the network fabric, which determines what the chip can actually do.
Fair-price band (disclosed training-class, Grade A/B): $10.50–$16.00 per GPU-hour on demand. Below that band, ask what the seller is not telling you about the fabric; above it, you are paying for something the rate card should be able to name.
Committed pricing (3 published rates tracked): from $10.58 per GPU-hour (AWS, 1-182 day block). Most committed pricing is quote-gated — the published rates below are the exception, and their spread against on-demand is the basis the futures market can't see. All terms in the table.
Is a GB200-NVL72 quote a good deal? Read it in ten seconds
| If you're quoted… | What it usually means | Do |
|---|---|---|
| below $10.50 | Cheaper than any disclosed training-class listing — typically unstated or shared fabric, not the same good cheaper. | Fine for batch/inference; for training, get the interconnect in writing first. |
| $10.50–$16.00 | Inside the fair band for disclosed training-class GB200-NVL72. | Reasonable walk-up price — proceed, confirm the grade. |
| above $16.00 | A premium over the median. | Worth it only if the fabric, SLA, and incident record beat the pack — verify, or negotiate down. |
Check a specific quote against your workload, utilization, and committed alternatives on the deal check. Procurement guidance from published prices — not investment advice, and not a view on any futures contract.
What fits on GB200-NVL72 (192 GB)
| Model | FP16 | FP8/INT8 | 4-bit |
|---|---|---|---|
| 8B class (Llama 3.1 8B) | 1 GPU | 1 GPU | 1 GPU |
| 32B class (Qwen 32B) | 1 GPU | 1 GPU | 1 GPU |
| 70B class (Llama 3.3 70B) | 1 GPU | 1 GPU | 1 GPU |
| 123B class (Mistral Large) | 2 GPUs | 1 GPU | 1 GPU |
| 405B class (Llama 3.1 405B) | 6 GPUs | 3 GPUs | 2 GPUs |
| 671B MoE class (DeepSeek-R1) | 9 GPUs (multi-node) | 5 GPUs | 3 GPUs |
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. Find the cheapest fit across all GPUs: Model Fit.
Every listing we track
| Provider | SKU | $/GPU-hr | Terms | Grade | Conf | Fabric (as published) | Src |
|---|---|---|---|---|---|---|---|
| CoreWeave | NVIDIA GB200 NVL72 (gb200-4x) North America, Europe | $10.50 | on-demand · min: docs state NVL72 must be deployed as full racks of 18 nodes | B | publ | 400 Gb/s NDR InfiniBand, NVIDIA Quantum-2 fabric | ↗vendor |
| AWS | u-p6e-gb200x72 (1-182 day block) | $10.58 | capacity-block · min: 1-182 day block | U | unkn | not published | ↗vendor |
| Microsoft Azure | Standard_ND128isr_NDR_GB200_v6 (3yr) | $11.90 | reserved · min: 3yr | A | publ | InfiniBand NDR: 4x 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand connections per VM (one 400 Gb/s NIC per GPU); 28.8 Tb/s scale-out networking per 72-GPU rack (Microsoft Learn) | ↗API |
| Oracle Cloud (OCI) | BM.GPU.GB200.4 | $16.00 | on-demand (PAY_AS_YOU_GO) | A | publ | '4x400 Gb/sec InfiniBand' = 1,600 Gbps per node (Oracle price-list shape table) | ↗API |
| Microsoft Azure | Standard_ND128isr_NDR_GB200_v6 (1yr) | $17.31 | reserved · min: 1yr | A | publ | InfiniBand NDR: 4x 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand connections per VM (one 400 Gb/s NIC per GPU); 28.8 Tb/s scale-out networking per 72-GPU rack (Microsoft Learn) | ↗API |
| Microsoft Azure | Standard_ND128isr_NDR_GB200_v6 eastus, eastus2 | $27.04 | on-demand | A | publ | InfiniBand NDR: 4x 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand connections per VM (one 400 Gb/s NIC per GPU); 28.8 Tb/s scale-out networking per 72-GPU rack (Microsoft Learn) | ↗API |
Grades summarize deliverability from the seller's own published specs — methodology. 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. A missing fabric field is recorded as unknown, never guessed. Weekly context in the Delivered Compute Report.
Source classes: API = provider's live pricing API · vendor = provider's published page · mkt API = marketplace API · 3rd-party = aggregator (provider publishes no price on its own pages) · quote = quote-gated or "starting at", not a walk-up price. Quote-class prices are shown in tables but excluded from headline ranges, spreads, and fair-price bands.
Method notes
Fair-price band = lowest to median price among Grade A/B on-demand listings — capacity whose published specs support training-class work. Marketplace listings (Vast.ai and similar) are pooled with on-demand in headline stats because both are walk-up prices; they carry confidence "mark" and usually cannot state their fabric. History: weekly snapshots begin September 2026; price-over-time charts appear as the series accumulates. Prices exclude storage, egress, and bundled vCPU/RAM differences — check the source link for the all-in picture before renting.
Prices move weekly. Get the report.
Headline prices, deliverability grades, and one teardown a week — free.
Subscribe free