NVIDIA GB300 NVL72 price per GPU-hour
Series 1 · GPU prices · NVIDIA · snapshot 7 Nov 2026
Blackwell Ultra in its rack-scale form is rented a few GPUs at a time out of a 72-GPU, liquid-cooled NVL72 system. Normalised to one GPU, the published on-demand rate is $18.00 an hour at the median, +36% against the GB200 median.
Figure 1GB300: recorded snapshots, USD per GPU-hour
| Snapshot | Median | Floor | Ceiling | Reserved 12 mo |
|---|---|---|---|---|
| 2026-11-07 | $18.00 | $13.44 | $20.00 | — |
Price trend: not enough data yet — 1 dated snapshot, and a trend needs at least 2. The chart replaces this table from the next snapshot of the GB300 series.
Reading the GB300 market
Only three sellers publish an on-demand GB300 price. Verda rents a single-GPU slice of an NVL72 tray at $13.44, the floor, and the same slice at $6.72 on spot. Oracle's BM.GPU.GB300.4 bare-metal shape, four GPUs per instance, works out to $18.00 per GPU-hour, and Fireworks lists GB300 deployments at $20.00, the ceiling.
Everyone else sells GB300 by quote. CoreWeave, Nebius, Crusoe and Together show the part on their sites without a self-serve rate, which is normal for rack-scale systems: the unit of sale is a reserved block of racks, priced on term, volume and delivery date. A three-seller range is therefore a guide to where list prices start, not to what a multi-rack contract costs.
Against its predecessor the GB300 costs $4.75 more per GPU-hour at the median (GB200 $13.25), but the GB200 range now reaches down to $8.00 as older racks fill, so the gap a buyer actually sees depends on which seller and which term. No 12-month GB300 rate is published anywhere we checked, so the reserved column stays empty.
Table 1GB300 prices by provider, USD per GPU-hour
| Provider | Offer | USD / GPU-hr | Detail | As of |
|---|---|---|---|---|
| Verda (ex-DataCrunch) | Spot | $6.72 | 1x GB300 SXM6 288GB spot instance | 2026-11-07 |
| Verda (ex-DataCrunch) | On-demand | $13.44 | 1x GB300 SXM6 288GB instance (GB300 NVL72 tray slice), 32 CPU, 225GB RAM; same rate listed for instant clusters per GPU | 2026-11-07 |
| Oracle Cloud (OCI) | On-demand | $18.00 | BM.GPU.GB300.4 (NVL72, 4x B300 279GB per instance), pay-as-you-go per GPU-hour, Oracle GPU price list (part B112140) | 2026-11-02 |
| Fireworks AI | On-demand | $20.00 | GB300 288 GB on-demand deployment GPU rate ($0.334/min) | 2026-11-07 |
Rows are the prices recorded in the tracker note for this model; 4 of the 4 come from the provider's own price page, the rest through comparison trackers. Third-party index rows are shown for context and not counted as providers. Every page behind this table: Sources, GB300.
The GB300 quotes and history are in the index download: JSON · CSV — free to reuse under CC BY 4.0; credit “Compute Census (computecensus.com)” with a link. Licence terms
Specifications
Table 2GB300 NVL72 rack (72 GPUs, 36 Grace CPUs); per-GPU memory = rack total / 72
| Memory per GPU (index) | 279 GB |
| GPUs and CPUs per rack | 72 Blackwell Ultra GPUs, 36 Grace CPUs |
| GPU memory, rack | 20 TB, up to 576 TB/s |
| CPU memory, rack | 17 TB LPDDR5X, 14 TB/s |
| FP4 Tensor Core, rack | 1,440 PFLOPS sparse | 1,080 PFLOPS dense |
| FP8/FP6 Tensor Core, rack, with sparsity | 720 PFLOPS |
| NVLink, rack | 130 TB/s |
| Cooling | Fully liquid-cooled, rack-scale |
Quoted from NVIDIA — GB300 NVL72, checked 7 Nov 2026. Vendor tensor figures are "with sparsity" unless stated; dense throughput is half.
What moves the GB300 price
- Rack-scale unit: NVIDIA specifies 72 Blackwell Ultra GPUs and 36 Grace CPUs per rack with 130 TB/s of NVLink, so a provider has to finance and fill the whole system before selling slices.
- Memory: 20 TB of GPU memory per rack (about 279 GB per GPU in the index) plus 17 TB of LPDDR5X on the Grace side — the configuration is built for very large models served from one NVLink domain.
- Low-precision throughput: 1,080 PFLOPS of dense FP4 per rack on NVIDIA's page, the figure that matters for inference economics.
- Liquid cooling only: NVIDIA describes the rack as fully liquid-cooled, which limits hosting to halls with direct liquid cooling and keeps supply tight.
Rent or buy
Rented whole, a GB300 NVL72 rack at the median per-GPU rate comes to $946,080 a month, and $706,406 at Verda's floor rate.
No dated public price for a GB300 rack appeared in any listing we checked; these systems are sold on quotation through NVIDIA partners and OEMs. The nearest reference in our data is a DGX GB200 rack listed at $8,424,412 before tax — the GB300 generation is generally priced above it. Owning one also means a liquid-cooled hall with enough power per rack; the request form below returns current offers for rack-scale systems and the sites that can host them.
Liquid-cooled AI capacity by country: AI data center census.
Other models with their own page: GB200, B200, MI355X, H200, H100, MI300X, A100, L40S, RTX 4090, B300. All models: GPU rental price index. How the figures are built: methodology.