GB300 NVL72 Cloud Pricing: What a Blackwell Ultra Rack Actually Costs
NVIDIA's GB300 NVL72 pairs 72 B300 "Blackwell Ultra" GPUs with 36 Grace CPUs in a single liquid-cooled rack. Here's what's actually confirmed about what one costs, and why most numbers you'll see quoted don't agree with each other.
The honest answer
There is no single agreed price for a GB300 NVL72 rack, and no on-demand rate published by any major cloud provider at all. Capex estimates from third-party analysts range from roughly $3.7M to $6.5M per rack, depending on configuration and who's estimating. Loop Capital analyst Ananda Baruah put a baseline rack at $3.7M to $4.0M, while Tom's Hardware supply-chain reporting put an inference-optimized configuration at $6.0M to $6.5M. Those two estimates are roughly 1.7x apart, which tells you these are analyst estimates, not published vendor list prices.
We went further than most coverage of this and checked whether the underlying B300 GPU has a real on-demand rate anywhere, since a rack rate should be traceable back to a per-GPU one. It doesn't, not at AWS, Azure, or Google Cloud. Details below.
Why You Can't Get a Straight Answer on Price
We checked AWS, Azure, and Google Cloud directly against their own pricing pages and announcements. None of the three publishes a standard, self-serve on-demand rate for B300, the GPU a GB300 NVL72 rack is built from. Access is gated the same way at all three:
| Provider | Real instance | How you actually get access |
|---|---|---|
| AWS | p6-b300.48xlarge | Capacity Blocks or Savings Plans only. AWS's own text: "for on-demand reservation, reach out to your account manager." |
| Azure | Standard_ND128isr_GB300_v6 | No published rate. Microsoft's spec page links out to the pricing calculator instead of listing one. |
| Google Cloud | A4X Max | Reserved capacity required to provision at all. Google's own guidance rules out standard on-demand. |
Contact for quote at all three. Checked directly against each provider's own site, not a secondary source.
This is exactly the situation the Quote Benchmarker exists for. If nobody publishes a rate, the number your rep gives you is the only number you have to go on, and you have no easy way to tell if it's reasonable. We can still tell you how it compares to what similar capacity has actually sold for.
Check a quote →GB300 vs GB200: The Chip-Level Difference
We don't have a full rack-level spec sheet for either NVL72 configuration, but the chip-level difference (same specs already shown on /compare/b300-vs-b200) drives the rack-level difference directly. Confirmed against Microsoft's own Azure specs: both GB300 and GB200 NVL72 racks use 4 GPUs per VM, 18 VMs per rack, 72 GPUs total, not 8 GPUs per VM as an earlier version of this page stated.
| Spec (per GPU) | B300 (GB300 NVL72) | B200 (GB200 NVL72) |
|---|---|---|
| Memory | 288GB HBM3e | 192GB HBM3e |
| TDP | 1400W | 1000W |
| Interconnect | NVLink 5.0 (1800 GB/s) | NVLink 5.0 (1800 GB/s) |
| FP8 (dense) | 7,000 TFLOPS | 4,500 TFLOPS |
| GPUs per VM (Azure) | 4 | 4 |
Per-GPU specs from GPUAdvisor's tracked dataset. GPU-per-VM count confirmed directly against Microsoft Learn's ND GB300-v6 and ND GB200-v6 series pages. A GB300/GB200 NVL72 rack contains 72 of the respective GPU plus 36 Grace CPUs across 18 VMs. Other rack-level aggregate specs (total memory, total power draw) are not independently verified here, so we don't state them as a single number.
Why Pricing Is Hard to Pin Down
A GB300 NVL72 rack isn't a single SKU. Configuration (networking, storage, liquid-cooling infrastructure, warranty terms) changes the price, and most public estimates come from supply-chain analysts pricing bare hardware, not vendors publishing a list price. That's why estimates for the same rack can differ by nearly 2x.
On the cloud rental side, rack-scale on-demand billing doesn't exist yet for this generation. Every provider we checked requires a capacity reservation, a Capacity Block, or a direct conversation with an account manager. That's a real, structural gap right now, not a gap in our tracking.
FAQ
Is GB300 NVL72 available on-demand yet?
No, not at AWS, Azure, or Google Cloud, and not even the underlying B300 GPU on its own. All three require a capacity reservation, a Capacity Block, or a direct conversation with sales. See the table above for what each provider actually requires.
How much more does GB300 cost than GB200?
It depends which estimate you use. The lower-end analyst estimate (Loop Capital) implies roughly a 10 to 30% premium over GB200's ~$3.1M to $3.9M estimated rack cost. The higher-end estimate (Tom's Hardware, for an inference-optimized configuration) implies close to double. There isn't a single confirmed multiplier.
What comes after GB300?
NVIDIA's next architecture, Vera Rubin, is expected after Blackwell Ultra. See our Rubin 2027 preview for what's publicly known so far (no pricing or specs yet).
Access-model and instance-name findings checked directly against AWS, Azure, and Google Cloud's own pricing pages and announcements, September 2026. Rack-level capex estimates are third-party analyst figures (Loop Capital, Tom's Hardware, SemiAnalysis), not vendor list prices or GPUAdvisor-verified numbers. All specifications approximate. GPUAdvisor has no commercial relationship with any provider listed. No sponsored rankings.