NVIDIA GB200 NVL72 rental price per GPU-hour
Series 1 · GPU prices · NVIDIA · snapshot 16 Oct 2026
GB200 is rented in slices of a 72-GPU, liquid-cooled rack, and normalised to one GPU the median rate is $13.25 an hour. That is the most expensive hour in the index, and it has fallen sharply this year.
Figure 1GB200: median rental price and published range, USD per GPU-hour, 2026-02-01 to 2026-10-14
§1Reading the GB200 market
GMI Cloud lists GB200 on demand at $8.00 per GPU-hour, marked available now — the first sub-$10 on-demand GB200 price the tracker has seen. CoreWeave's long-standing listing is $10.50 ($42 an hour for a four-GPU instance). Oracle bare metal, at ~$16.00, is the best-corroborated mid-point, and Azure's ND128isr GB200 v6 costs $108.16 an hour for four GPUs: $27.04 each, the ceiling.
The index median of $13.25 is the midpoint of the two confirmed on-demand configurations, down from an Oracle-anchored $16.00 one cycle earlier. Over the series the median has dropped from $18.50 in Feb 2026 to $13.25 — −28% in under a year; each new on-demand listing, most recently GMI Cloud's, has pulled it lower.
Only one term price exists in the record: AWS's 1-year reserved rate of $10.58, 20% below the on-demand median but above GMI's on-demand listing.
Table 1GB200 prices by provider, USD per GPU-hour
| Provider | Offer | USD / GPU-hr | Detail | As of |
|---|---|---|---|---|
| GMI Cloud | On-demand | $8.00 | listed as available now | 2026-09-14 |
| CoreWeave | On-demand | $10.50 | $42/hr for a 4-GPU instance | 2026-09-14 |
| AWS | Reserved, 12 mo | $10.58 | 1-year reserved | 2026-09-14 |
| Oracle Cloud | On-demand | ~$16.00 | bare metal | 2026-09-14 |
| Microsoft Azure | On-demand | $27.04 | ND128isr GB200 v6, $108.16/hr for 4 GPUs | 2026-09-14 |
Rows are the prices recorded in the tracker note for this model; 3 of the 5 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, GB200.
§2Specifications
Table 2GB200 NVL72 rack (one Grace Blackwell Superchip in brackets)
| Memory per GPU (index) | 186 GB |
| Configuration | 36 Grace CPUs, 72 Blackwell GPUs (Superchip: 1 CPU, 2 GPUs) |
| GPU memory | 13.4 TB HBM3E (Superchip: 372 GB) |
| Memory bandwidth | 576 TB/s (Superchip: 16 TB/s) |
| FP8/FP6 Tensor Core, with sparsity | 720 PFLOPS (Superchip: 20 PFLOPS) |
| NVLink bandwidth | 130 TB/s (Superchip: 3.6 TB/s) |
| Cooling | Rack-scale, liquid-cooled design |
Quoted from NVIDIA — GB200 NVL72, checked 17 Oct 2026. Vendor tensor figures are "with sparsity" unless stated; dense throughput is half.
§3What moves the GB200 price
- Rack-scale unit: NVIDIA specifies the NVL72 as 36 Grace CPUs and 72 Blackwell GPUs sharing 130 TB/s of NVLink; providers sell slices, but they must finance and fill the whole rack.
- Liquid cooling: NVIDIA describes a rack-scale, liquid-cooled design, so only facilities with direct liquid cooling can host it — a hard supply constraint.
- Memory per GPU: 186 GB in the index (372 GB per two-GPU Superchip), the largest of the NVIDIA parts tracked here.
- Few sellers: four on-demand providers and one reserved price make each new listing move the median.
§4Rent or buy
Rented whole, an NVL72 rack at the median per-GPU rate comes to $696,420 a month; at GMI's $8.00 it is $420,480, and at the AWS 1-year rate $556,085.
Buying means buying a rack, not a GPU, and we hold no cited rack price — so none is printed. Ownership also needs a liquid-cooled hall with enough power for the whole rack. If owning is the plan, start with a site: the AI campus census lists liquid-cooled capacity by country.
Liquid-cooled capacity by country: AI data center census.
Other models with their own page: B200, MI355X, H200, H100, MI300X, A100, L40S, RTX 4090. All models: GPU rental price index. How the figures are built: methodology.