GB200 NVL72 pricing is not what stops most buyers from booking rack-scale Grace Blackwell capacity. Every major provider prices it differently, and none of them publish a single number you can shop against another. What actually blocks a purchase is the fine print under that number: how far ahead you can book, whether the reservation guarantees capacity at all, and whether you can walk away if your workload changes. AWS, Azure, Nebius and CoreWeave structure that fine print in four genuinely different ways, and the gap between them is where the real risk lives.
TL;DR: GB200 NVL72 Pricing, Minimum Commitment and Lead Times
A GB200 NVL72 reservation means committing to one of four very different structures, not shopping a published rate.
- AWS: EC2 Capacity Blocks for P6e-GB200 UltraServers book at most 8 weeks ahead, charge the full fee upfront, and allow no cancellations once purchased.
- Azure: the ND-series GPU VM family that ND GB200 v6 belongs to is excluded from Azure's no-term Capacity Reservation, forcing a 1- or 3-year Reserved Instance with no capacity guarantee.
- CoreWeave: weighted average contract length on new capacity sat at approximately 5 years as of Q1 2026, per CFO Nitin Agrawal, against a $99.4 billion backlog.
- Nebius: GB200 NVL72 and GB300 NVL72 capacity is quoted sales-only, with no published minimum term, deposit, or unit size.
- Spheron: lists GB200 NVL72 as reservation-only with no fixed minimum term, a one-node minimum, and a quote back within one business day; see the GB200 NVL72 reservation page.
Why GB200 NVL72 Pricing Isn't the Real Blocker
A GB200 NVL72 isn't a single chip you can price like a GPU SKU. It's a full rack: 36 GB200 Superchips, each one Grace CPU paired with two Blackwell B200 GPUs over NVLink-C2C, for 72 GPUs networked into one memory domain big enough to hold a model a single 8-GPU node can't. The GB200 itself refers to the Superchip, not a standalone accelerator; nobody buys or rents a GB200 as one chip, and nobody rents an NVL72 rack as one hourly line item either. Our full architecture breakdown covers the spec sheet in detail if you need the rack power draw and interconnect numbers before anything else.
That absence of a single rack price is the first thing that trips buyers up, and it's structural, not an oversight. AWS prices P6e-GB200 UltraServers per accelerator-hour rather than per rack, as our own breakdown of that rate shows. CoreWeave has published GB200 pricing at the node level rather than the full rack. Nebius and Azure don't publish a rate at all; both route you to a sales conversation first. So "how much is a GB200 NVL72 rack" doesn't have one answer, it has four, and three of those four require you to already be in a contract negotiation before you see a number.
That's the actual blocker. A quoted rate, once you get one, is comparable across providers in a spreadsheet in about ten minutes. What isn't comparable in ten minutes is whether the number you're being quoted comes with a capacity guarantee, how far in advance you had to book it, whether you can get out of it if the project changes, and whether the provider can even deliver GB200 specifically versus GB300, which carries its own separate price gap worth checking first. Those four questions are where AWS, Azure, Nebius and CoreWeave genuinely diverge, and none of the rate cards above answer them.
Lead Times: Why Nebius, AWS and CoreWeave All Quote Differently
Ask three cloud providers how fast you can get GB200 NVL72 capacity and you'll get three different answers, because none of them are actually describing the same kind of commitment.
AWS is the most mechanical about it. EC2 Capacity Blocks for ML, the only way AWS sells P6e-GB200 UltraServer capacity, let you schedule a reservation with a start time up to 8 weeks in the future. Durations step up in fixed increments: 1-day blocks up to 14 days, then 7-day blocks up to 182 days. That's a lead-time ceiling, not a floor: you can't book further out than 8 weeks even for a project you already know is starting next quarter. Once you've booked, AWS charges the full reservation fee upfront, debited within 5 minutes to 12 hours of purchase, and releases the reservation if the charge fails. There's no walking this back afterward: cancellations aren't allowed once a Capacity Block is purchased, and on P6e-GB200 UltraServer blocks specifically, you have to terminate your instances at least 60 minutes before the block's scheduled end. AWS's standard Savings Plans and Reserved Instance discounts don't apply here either.
CoreWeave sits at the opposite end. There's no self-serve booking window at all: new capacity is sold as a negotiated contract, and on CoreWeave's Q1 2026 earnings call, CFO Nitin Agrawal told investors the company's "weighted average contract length for new capacity remaining at approximately 5 years." That length is backed by demand: CoreWeave's backlog stood at $99.4 billion at the end of the quarter, with about 36% of it expected to be recognized within 24 months and 75% within four years. A ~5-year weighted average isn't a lead time in AWS's sense; it's a statement about how long the capacity you're bidding on is already spoken for by other customers' multi-year commitments, which is its own kind of lead time if you're trying to get in behind them.
Nebius doesn't publish a number at all. Its GB300 NVL72 and Blackwell pre-order page describes the entire process as expressing "interest in pre-booking," after which its sales team "confirms the dates of your order and provides the contract." Its separate Grace Blackwell product page says the same for GB300 NVL72: available only through the Nebius sales team. There's no published minimum commitment, deposit, or term length for either product. That's not necessarily worse than AWS's 8-week ceiling or CoreWeave's 5-year average; it just means the lead time and the terms are whatever gets negotiated in that specific sales conversation, and you won't know either number until you're already talking to someone.
For a broader look at why providers can hold this much leverage on lead times right now, see our breakdown of the 2026 GPU shortage and what it's done to availability across the board.
AWS Capacity Blocks vs a Term Reservation: What You're Actually Committing To
The clearest way to see how different these commitments are is to put AWS's structure next to Azure's, because Azure actually has two different reservation products, and the one that guarantees capacity doesn't support GB200 at all.
| Dimension | AWS EC2 Capacity Block (GB200) | Azure On-Demand Capacity Reservation | Azure Reserved Instance (ND GB200 v6) |
|---|---|---|---|
| Lead time | Book up to 8 weeks ahead | Not supported for ND-series GPU VMs | Sales-quoted, no published ceiling |
| Term length | 1-day steps to 14 days, then 7-day steps to 182 days | No term required | Fixed 1-year or 3-year |
| Capacity guarantee | Yes, for the booked window | Yes, where supported | No guarantee; "Capacity priority" option only, no SLA |
| Payment | Full reservation fee upfront, no cancellations | Pay as you go | Upfront or monthly, discounted vs on-demand |
The row that matters most is the one in the middle. Azure's own documentation is direct about the split: "Capacity reservation secures capacity availability for deployment. Reserved Instances provide a billing discount only and don't guarantee capacity." That's a clean distinction, except Azure's capacity-guaranteed product, On-Demand Capacity Reservation, explicitly doesn't support the ND-series GPU VM family, which is where ND GB200 v6 lives. So a GB200 buyer on Azure can't actually reach the product that guarantees capacity. The only path to GB200 on Azure is the Reserved Instance, a 1- or 3-year commitment that, by Microsoft's own description, buys a billing discount and a "Capacity priority" option with no SLA behind it, not a guarantee that the hardware will be there.
AWS's Capacity Block is the inverse trade. You get an explicit capacity guarantee for your booked window, in exchange for an 8-week lead-time ceiling, full payment upfront, and zero flexibility to cancel. For a workload you can plan two months out and don't need to walk away from, that's a reasonable trade. For a timeline that's still moving, or where you want insurance you can get out of, it's the wrong tool, and it's also the only self-serve tool among the four providers in this post. CoreWeave and Nebius don't have a comparable public SLA to put in this table at all; both of those terms get set in the contract, not the rate card.
Questions to Ask Before You Sign
Once you're past the headline rate, these are the questions that actually determine what you're committing to:
- How far ahead can I book, and is that a ceiling or a floor? AWS's 8-week window is a hard ceiling; you can't reserve further out even if your timeline is firm. Nebius and CoreWeave have no published ceiling because the lead time is whatever the sales conversation produces.
- Does this product actually guarantee capacity, or just a price? Ask the Azure question specifically: is the SKU you need even eligible for the capacity-guaranteed product, or does it route you to a Reserved Instance that only discounts billing?
- What happens to my money if I need to cancel or shrink the commitment? AWS's answer is that you can't. Get CoreWeave's and Nebius's answers in writing before you sign, not as a verbal assurance.
- What's the real minimum unit? A rack-level commitment and a single-node commitment are very different amounts of risk, and providers don't always lead with the smaller option.
- If I need more than my reserved amount, what rate applies to the overage, and is that written into the contract or just "standard rates apply"? Our guide to negotiating a GPU cluster reservation contract goes through overage rates, exit clauses, and capacity-guarantee language in more depth than any single provider's rate card will give you.
Spheron's own GB200 NVL72 listing answers a few of these differently than the four providers above. It's reservation-only, like Nebius, Azure's Reserved Instance and CoreWeave's contracts, with no fixed hourly rate published. But the minimum unit is a single 8-GPU node rather than a full rack or a multi-year contract, there's no published minimum term, and the team responds with availability, region, and a quote within one business day of getting your GPU count, timeline, and workload. That's faster than an open-ended sales process, but it isn't AWS's self-serve Capacity Block, where you reserve and pay in minutes with no conversation at all, and a single node still typically takes a few days to come online rather than being instant. If your timeline genuinely needs AWS's 1-to-14-day self-serve window, that's a real reason to start there instead. If what you need is a quote you can act on without a rack-wide spend or a five-year contract first, Spheron's GB200 NVL72 reservation form is where that conversation starts.
None of the four provider structures in this post is wrong, and Spheron's is a fifth shape, not a verdict on the other four. They're built for different buyers: AWS for teams with a firm two-month-out date and no need to negotiate, CoreWeave for teams ready to sign multi-year capacity, Nebius and Azure's Reserved Instance for teams already deep in a sales relationship, and a smaller, quote-based reservation for teams that want to start at one node before committing to more. The mistake is comparing any two of them on price alone.
GB200 NVL72 pricing is the part you can shop. The reservation terms are the part you have to read, because that's where AWS, Azure, Nebius and CoreWeave actually differ.
Frequently Asked Questions
The GB200 NVL72 is NVIDIA's rack-scale AI system: 36 GB200 Superchips, each pairing one Grace CPU with two Blackwell B200 GPUs over NVLink-C2C, networked into 72 GPUs that share one memory domain. It's reserved as a rack-level or node-level unit, never as a single chip.
No. GB200 is the name of the Superchip module, one Grace CPU plus two B200 GPUs, and the NVL72 is the full 72-GPU rack built from 36 of those Superchips. You can't buy or rent a GB200 by itself the way you can rent a single H100 or B200.
There's no single published rack price. AWS prices its P6e-GB200 UltraServer per accelerator-hour rather than per rack, CoreWeave has published node-level GB200 pricing instead of a full-rack figure, and Nebius and Azure route GB200 pricing through a sales conversation with no published rate at all. Comparing providers means comparing the reservation structure, not one number.
No. AWS does not allow cancellations once an EC2 Capacity Block for ML is purchased, and the full reservation fee is charged upfront. For P6e-GB200 UltraServer blocks specifically, you also have to terminate your instances at least 60 minutes before the block's scheduled end time.
Not through its no-term Capacity Reservation product, which explicitly excludes the ND-series GPU VM family that ND GB200 v6 belongs to. GB200 access on Azure goes through a Reserved Instance instead, a fixed 1- or 3-year term that Microsoft's own documentation says provides a billing discount only, not a capacity guarantee.
