Pricing reference
Six public prints for the same class of rack, spanning 3.4x.
Public list or reported prices from other sellers, gathered for orientation. They are not GPU SLAX prices, not quotes, and not multi-year committed rates.
| Seller and product | USD per GPU-hour | Basis | Source | As of | Confidence |
|---|---|---|---|---|---|
| GMI CloudGB200 NVL72 | from $8.00per GPU-hour | Advertised starting rate, reserved or on-demandMarketed as a floor. GMI states that racks move to committed contracts. Not re-verified on the live page. | GMI Cloud pricing page and blog | Medium confidence | |
| CoreWeaveGB200 NVL72, 4-GPU instance | $10.50per GPU-hour | On-demand, North America. $42.00 per instance-hour divided by 4 GPUsPer-GPU figure is our division of the instance price by GPU count. CoreWeave advertises committed discounts of up to 60 percent but publishes no committed rate card. | CoreWeave GPU pricing page | High confidence | |
| AWSGB200 UltraServer capacity block (u-p6e-gb200x72) | $10.58per GPU-hour | EC2 Capacity Blocks for ML. $761.904 per block-hour divided by 72 acceleratorsShort-term reservation, not a multi-year contract. AWS reprices blocks on supply and demand; next update October 2026. | AWS Capacity Blocks for ML pricing | High confidence | |
| Oracle CloudGB200 NVL72 (BM.GPU.GB200.4) | $16.00per GPU-hour | Pay-as-you-go list, metric GPU per hourOnly hyperscaler with a self-serve GB200 NVL72 rate. Annual-commit discounts are not public. | OCI Price List API, part number B110979 | High confidence | |
| Oracle CloudGB300 NVL72 (BM.GPU.GB300.4) | $18.00per GPU-hour | Pay-as-you-go list, metric GPU per hourThe only primary GB300 NVL72 hourly print we found. Roughly a 12.5 percent premium over Oracle's GB200 rate. | OCI Price List API, part number B112140 | High confidence | |
| Microsoft AzureND GB200 v6, 4-GPU VM (Standard_ND128isr_NDR_GB200_v6) | $27.04per GPU-hour | Retail Prices API consumption meter, Linux, US regions. $108.16 per VM-hour divided by 4 GPUsA consumption meter exposed by the pricing API, not a figure on Azure's marketing page. Highest published GB200 rate found. Azure publishes no GB300 price. | Azure Retail Prices API | High confidence |
Methodology
- We include a seller only when it prints a GB200 or GB300 NVL72 rate on its own pricing page or pricing API. Sellers that say contact sales are not listed, and we do not attribute a price to them.
- Where a seller prices a multi-GPU instance, the per-GPU figure is the instance price divided by its GPU count. Divisions are noted in the basis column.
- Every row carries the source name, a link, and the date we last checked it. High confidence means the figure was read from the seller's live page or API on that date. Medium means it rests on a dated secondary page and was not re-verified.
- All figures are on-demand, pay-as-you-go or short-term block rates. No seller publishes a one-, two- or three-year committed price for a full NVL72 rack. Committed discounts of 35 to 60 percent are advertised by some sellers without a rate card.
- Azure's figure is a consumption meter exposed by its Retail Prices API, not a marketing-page list price. Azure publishes no GB300 price, and we show none.
How to read the table against the book
GPU SLAX tickers are multi-year, take-or-pay contracts for whole racks. The public prints above are hourly rates for capacity you can release at will. A committed rack should clear below the on-demand prints and above the owner's cost of running it. Where inside that band is what the order book is for.
Rack arithmetic: one NVL72 rack is 72 GPUs, or 630,720 GPU-hours per year. At $10.00 per GPU-hour that is $6.31 million of notional per rack-year and a $1.26 million deposit at 20 percent. Every price on the exchange shows its notional, per-day cost and deposit next to it.
GB300 prints carry a premium of roughly 10 to 14 percent over GB200 at the sellers that list both.