NVIDIA's current Blackwell Ultra rack is the GB300 NVL72, and Blackwell Ultra rack pricing is one of the murkiest numbers in cloud GPU today. Search for what it actually costs to rent and you'll find a lot of secondhand numbers and very little from the providers themselves. We checked the pricing pages that matter directly, AWS, Azure, Oracle, and CoreWeave, rather than repeating what aggregator sites report. Three of the four are flatly quote-only. The fourth has a widely cited rate that doesn't trace back to the provider's own price list at all.
TL;DR: What Is Blackwell Ultra Rack Pricing for GB300 NVL72 in 2026?
No major cloud provider publishes a self-serve hourly or rack-level price for GB300 NVL72 as of this post.
- AWS: no GB300 UltraServer line on its Capacity Blocks pricing page; GB200 UltraServer is priced at $10.582/accelerator-hr for comparison.
- Azure: no rate for ND GB300 v6 (Standard_ND128isr_GB300_v6); routes to a sales quote.
- Oracle: BM.GPU.GB300.4 shape is documented with full specs but no price; the $18.00/GPU-hr figure that circulates online traces only to third-party aggregators.
- CoreWeave: GB300 NVL72 shows "Contact sales," while the same-size GB200 NVL72 bundle is priced at $42.00/hr.
- Spheron: takes GB300 NVL72 reservations the same way, sized from one node to a full rack, with no hourly rate shown on its pricing page either.
Why Blackwell Ultra Rack Pricing Has No Hourly SKU for GB300 NVL72
GB300 NVL72 is NVIDIA's Blackwell Ultra rack: 72 B300 GPUs and 36 Grace CPUs in one liquid-cooled system, a generational step up from GB200 NVL72 in memory and dense FP4 throughput. What matters for pricing is simpler: this is a multi-million-dollar rack, not a GPU you add to a cart, and the line between the rack and the standalone GPU inside it matters more here than anywhere else. Our GB300 vs B300 breakdown covers exactly where that line falls.
AWS's own announcement of general availability states that P6e-GB300 UltraServers deliver 1.5x the GPU memory and 1.5x the FP4 compute, without sparsity, of the P6e-GB200 generation that came before them. That's a meaningful jump, and it's also recent: P6e-GB300 only went GA in December 2025, which is part of why the pricing infrastructure around it still looks unfinished. A rack this new, this expensive, and this concentrated in supply isn't something hyperscalers put behind a self-serve checkout button. It goes through the same account-team process as any large committed infrastructure purchase, the same way GB200 NVL72 did in its first year, as our GB200 NVL72 guide covers.
That reservation-only status is specific to the rack, not to Blackwell Ultra generally. A standalone B300 GPU, outside the NVL72 rack configuration, does have published rates in places. AWS lists a single-node B300 Capacity Block directly. Spheron prices B300 on-demand and spot per GPU. Our B300 Blackwell Ultra guide covers that single-GPU tier in full. GB300 NVL72 is a different purchase decision entirely: you're buying (or reserving) 72 interconnected GPUs as one system, and none of the four major providers we checked treat that as a metered, hourly product yet.
What AWS, Azure, Oracle, and CoreWeave Are Charging for GB300 Capacity
None of the four providers we checked directly publish a self-serve GB300 NVL72 rate on their own pricing pages. Here's what each one actually shows, checked against the provider's own documentation rather than a reseller's summary:
| Provider | GB300 NVL72 Status | What's Actually Published |
|---|---|---|
| AWS | No UltraServer price | No u-p6e-gb300x36 or u-p6e-gb300x72 line on the Capacity Blocks pricing page; GB200 UltraServer is $10.582/accelerator-hr |
| Azure | Sales quote only | No rate for ND GB300 v6 (Standard_ND128isr_GB300_v6) on the Linux VM pricing page |
| Oracle | Sales quote only | BM.GPU.GB300.4 shape fully specced in docs, no price shown; third-party aggregators report $18.00/GPU-hr, unconfirmed by Oracle |
| CoreWeave | "Contact sales" | GB300 NVL72 (4-GPU bundle) has no rate; the same-size GB200 NVL72 bundle is $42.00/hr |
Pricing fluctuates based on GPU availability. Third-party rates above reflect each provider's own published figures as of this post and may have changed; Spheron's B300 rate elsewhere in this post is live as of 09 Nov 2026. Check current GPU pricing for live rates.
AWS's own Capacity Blocks pricing page lists a rate for the GB200 UltraServer, $10.582 per accelerator-hour, for both the 36- and 72-GPU block sizes. There is no equivalent line for a GB300 UltraServer. The only B300 figure on that page at all is the standard, non-rack p6-b300.48xlarge Capacity Block, a single 8-GPU node rather than an NVL72 system, priced at $16.146 per accelerator-hour effective 7 October 2026.
Azure's Linux VM pricing page shows no per-hour figure anywhere for the ND GB300 v6 series (Standard_ND128isr_GB300_v6), unlike H100 or H200 SKUs, which list prices directly. GB300 inquiries route to a sales quote. Our Azure GB300 pricing breakdown walks through what that sales process actually involves.
CoreWeave's own pricing page lists GB300 NVL72, sold as a 4-GPU bundle with 279 GB of VRAM per GPU, as "Contact sales" with no rate shown at all. The same bundle size for the prior generation, GB200 NVL72, is priced directly at $42.00/hr. There's no price anywhere on that page either.
Oracle is the interesting case, and the one most worth getting right. Oracle's own compute shapes documentation lists the BM.GPU.GB300.4 shape, four B300 GPUs at 278 GB each, two 72-core Grace CPUs, and 960 GB of system memory, in full detail. But the page carries no price; it points readers to Oracle's Cost Estimator or Cloud Price List instead. Third-party GPU-price aggregators report an $18.00 per-GPU-hour on-demand rate for that shape, about $72/hr for the full 4-GPU configuration, and that number gets repeated often enough that it reads like an established fact. It isn't one. We checked Oracle's own pricing and shape documentation directly and found no trace of that figure anywhere on Oracle's own pages. Treat it as an aggregator estimate, not a confirmed Oracle rate, until Oracle publishes one itself.
Reading a GB300 Capacity Block: Minimums, Terms, and Lock-In
The one piece of GB300-adjacent pricing structure that is public, AWS's Capacity Blocks model, tells you something about what to expect even without a GB300-specific number. Capacity Blocks bill in accelerator-hours over a reserved term rather than metered, pay-as-you-go hours, which is why AWS can publish $10.582/accelerator-hr for GB200 UltraServers while still requiring you to reserve a block of time up front rather than spin one up for an afternoon. There's no reason to expect GB300 UltraServer pricing, once AWS publishes it, to work differently.
The other three providers don't get you even that far. Azure's path is an account-team conversation with no public rate card to anchor against beforehand. CoreWeave's GB300 documentation gives you the exact hardware spec, down to vCPU count, but stops short of anything you could use to estimate cost; you still contact sales to find out what it runs and under what term. Oracle is the most transparent on specs and the most opaque on getting a number: the shape page tells you precisely what BM.GPU.GB300.4 contains, then hands you off to a cost estimator tool rather than a list price, which at least suggests Oracle has an internal number it just isn't printing on that page.
The practical takeaway: budgeting for GB300 NVL72 today means a sales conversation, not a pricing page, at every one of the four providers we checked. Build in time for that conversation, and don't anchor a budget to the $18/GPU-hr figure circulating for Oracle until Oracle itself confirms it.
Reserving GB300 NVL72 Capacity on Spheron
Spheron's own pricing page lists GB300 under "Reserve," with no hourly rate shown, the same treatment it gives GB200 and Vera Rubin. That puts Spheron in the same quote-gated category as AWS, Azure, Oracle, and CoreWeave on the number itself: nobody in this market has a published GB300 NVL72 rate today, Spheron included.
Where the request process differs is who it's open to. A GB300 reservation on Spheron doesn't require an existing enterprise sales relationship the way AWS, Azure, and CoreWeave gate their GB300 access behind account teams built for their largest customers. It's also sized more flexibly: from a single 8-GPU node up to a full 72-GPU rack, rather than AWS's and CoreWeave's fixed block or bundle increments. That flexibility comes from aggregating capacity across 5+ providers on the backend, which is the same mechanism that keeps standalone B200 and B300 available on-demand while a GB300 quote is still pending.
If you need a number to drop into a budget this week rather than after a sales call, no provider in this comparison, Spheron included, gives you one for the full rack yet. What you can do in the meantime is reserve the rack and get a confirmed quote, or run the workload today on standalone B200 or B300 capacity and move to GB300 NVL72 once pricing and availability are confirmed. For planning what comes after Blackwell Ultra, see our Vera Rubin NVL72 cloud availability guide.
GB300 NVL72 pricing stays behind a sales call at every major provider we checked. Spheron reserves it the same way, from a single node to a full rack, without requiring an existing enterprise account.
Frequently Asked Questions
No provider publishes an official rack price. Our GB300 NVL72 vs GB200 NVL72 price breakdown covers the analyst estimates in detail: a physical GB300 NVL72 rack runs from $3.7M to $4.0M up to $6.0M to $6.5M, against roughly $3.1M to $3.9M for a GB200 NVL72 rack. Those are purchase-cost estimates, not cloud rental rates, since none of AWS, Azure, Oracle, or CoreWeave list an hourly or rack-level cloud price for GB300 NVL72 on their own pricing pages.
No. AWS, Azure, Oracle, and CoreWeave all treat GB300 NVL72 as reservation-only or sales-gated capacity, with no self-serve, pay-by-the-hour checkout flow on any of their pricing pages as of this post. Spheron takes the same reservation approach, sized from a single 8-GPU node up to a full 72-GPU rack, through a request form rather than an enterprise sales process.
None of the four checked directly. AWS prices the GB200 UltraServer at $10.582 per accelerator-hour but has no equivalent GB300 line. Azure and CoreWeave route GB300 inquiries to sales with no rate shown. Oracle's own shape documentation lists GB300 specs with no price; the $18.00 per-GPU-hour figure that circulates for Oracle only appears on third-party aggregator sites, not on any Oracle-published page.
Standalone B300 has a real published rate in places: AWS lists its single-node p6-b300.48xlarge Capacity Block at $16.146 per accelerator-hour, and Spheron prices B300 on-demand at $12.42/hr and spot at $7.84/hr as of 09 Nov 2026. GB300 NVL72, the full 72-GPU rack system, has no such published figure anywhere; it's sold as reserved, quote-gated capacity instead.
