Which GPU Clouds Offer B300 or NVL72 Reservations with a Confirmed Delivery Plan?

For a B300 or NVL72 reservation with a confirmed delivery plan, choose the provider by how it sells the system you need. For HGX B300 nodes on a fixed date, AWS EC2 Capacity Blocks for ML is the clearest public option: a block can be reserved to start up to eight weeks out, it cannot be canceled, and its supported instance types include a B300 instance. Nebius sells HGX B300 as self-service instances, so for small quantities the delivery date is the day you launch. For GB300 NVL72 racks, Nebius (through its sales team only), Together AI (reserved capacity in term bands), and GMI Cloud all quote through sales, and the one to pick is the provider that puts allocation, site readiness, milestones, and a slip remedy into the signed quote.

At GMI Cloud, our NVIDIA Blackwell lineup, including B300, GB200 NVL72, and GB300 NVL72, is in pre-order and coming soon, and the delivery window, milestones, and commitment terms are set with our sales team in the quote. As with any pre-order, the signed terms decide whether a date is a commitment, and three things show whether it will hold: specific systems allocated to your order, a site that can power and cool them, and a contract remedy if the date moves. The checks below show how to tell the difference before you sign, for any quote, ours included.

Commitment or forecast: the wording that tells you which

The fastest test is the sentence the date appears in. Most delivery language takes one of three forms.

Watch for qualifiers that pull the third form back into the first: "subject to availability," "upon receipt of hardware," "estimated," or "best effort." Each can be reasonable, and each moves the risk of a late shipment from the provider to you.

Then check which event the date refers to. "Delivery" can mean the hardware arrives at the data center, your team can log in, or the system passes acceptance. On rack-scale systems, installation and validation can separate those events, and a quote that only says "delivery" leaves the choice to the provider. Ask for all three dates, and treat the acceptance date as the one your project plan depends on.

Pin the unit before you test the date

A date only means something for a specific system. "B300" alone is ambiguous. It is the name of the Blackwell Ultra GPU, and some clouds use it for an 8-GPU HGX B300 server while others use it for the GPUs inside a GB300 NVL72 rack. "NVL72" needs its prefix too, because GB200 NVL72 and GB300 NVL72 are different racks.

• HGX B300 node. NVIDIA's HGX page lists eight Blackwell Ultra GPUs with 2.1 TB of total GPU memory, connected by NVLink 5 Switch inside the node.
• GB200 NVL72 rack. NVIDIA lists 72 Blackwell GPUs and 36 Grace CPUs in one liquid-cooled rack, with 13.4 TB of HBM3E and a single 72-GPU NVLink domain.
• GB300 NVL72 rack. NVIDIA lists 72 Blackwell Ultra GPUs and 36 Grace CPUs in a fully liquid-cooled rack, with 20 TB of GPU memory and 130 TB/s of NVLink bandwidth.

The unit changes what the date depends on. A node order depends mostly on server supply and high-density rack space at the site. A rack order also depends on liquid cooling, rack-scale power, and validation of the whole NVLink domain. Write the unit into the quote with its GPU count, CPU type, and NVLink domain size, so the date you are testing belongs to the system you plan to run.

Five pieces of evidence behind a real date

A committed date rests on a chain of steps, and each step has to finish on time. Ask for evidence of each one, with its own date and a named owner. An undocumented step leaves uncertainty in every milestone that depends on it.

1. Allocation confirmation. Written confirmation that specific systems are assigned to your order, not a share of a future shipment. On a pre-order, this is the step that turns a place in line into a delivery plan, so ask when it happens and what document you receive.
2. Site power and liquid-cooling readiness. The named data center where the systems will run, and confirmation that its power and cooling are ready for them. In NVIDIA's enterprise reference architecture for NVL72 AI factories, a full GB300 NVL72 rack can require up to 142 kW and is liquid cooled. That is a facility requirement a provider either has in place at a named site or is still building out.
3. Network and storage provisioning. Networking and storage are often provisioned separately from GPU capacity. The plan should name the fabric, the storage type and size, and the date each is ready, since a rack without its network or storage cannot run your jobs.
4. A named owner for every milestone. Each step should say who is responsible: the provider, a facility partner, or your own team. Steps without an owner are the ones that slip without anyone telling you.
5. An acceptance test that starts billing. The contract should define the test that proves the system is ready, the pass criteria, and a retest window, and it should tie billing to passing that test rather than to allocation or hardware arrival.

Allocation and site readiness carry the most weight. The first tells you the hardware is yours, and the second tells you it has somewhere to run. Missing evidence increases delivery risk, even when the contract contains a firm date.

Which sales channels can commit to a date, and how far

How a provider sells a system hints at how firm a date it can offer, before you read a single term. Read the channel as a sign of how strong the evidence is likely to be, not as a ranking of providers.

Self-serve instances. When a node is sold self-serve, the practical delivery date is the moment you launch it, as long as capacity is there. Nebius, for example, lists HGX B300 as self-service GPU instances. For small quantities the date question mostly goes away. For large quantities it comes back, and you are in a reservation conversation again.

Fixed-date capacity blocks. The clearest public example of a dated reservation is AWS EC2 Capacity Blocks for ML. AWS documentation says a Capacity Block can be reserved with a start time up to eight weeks out and cannot be canceled, and its list of supported instance types includes a B300 instance. The reserved window comes with a payment commitment: you pay for the block whether or not you use it. That trade is a useful yardstick for any quote, since a firm date usually comes with a firm commitment.

Sales-only and term-based quotes. Rack-scale systems are usually quoted through a sales team. Nebius lists GB300 NVL72 as available through its sales team only, and Together AI offers GB300 NVL72 as reserved capacity in term bands rather than by the hour. In this channel, the date is whatever the signed quote says, so the five pieces of evidence above do all the work.

Pre-orders. A pre-order records demand for systems that are not yet deliverable. The order terms determine whether its date is an estimate or a contractual commitment. Allocation and site milestones show how credible the delivery plan is. Pre-orders are where the five checks matter most, because none of the answers can be read off a product page yet.

Whatever the channel, the date you can rely on is the one in the signed contract, together with its remedy.

Clauses that make a date hold

Evidence tells you whether a date is realistic. Clauses decide what happens when it turns out not to be. Five belong in any Blackwell reservation:

• Notice of slippage. The provider tells you in writing as soon as a milestone is at risk, within an agreed notice period, and gives a revised date. The earlier the notice, the more of your own schedule you can move in time.
• Substitute hardware, and the right to reject it. If the provider offers a different system, region, or configuration instead, the contract should say whether price, commitment, and acceptance criteria change. Keep the right to decline a substitute that fails your acceptance test.
• Partial first delivery. If part of the order can arrive early, state the quantity, whether it shares an NVLink domain with what follows, and the date for the rest. Each stage needs its own acceptance event and billing start.
• A term that follows the real start. The committed term should begin at acceptance, so a late delivery does not shorten the months you agreed to pay for.
• Credits or exit if the date moves. Name the remedy: service credits after a set delay, or the right to leave the commitment without penalty after a longer one. Without a remedy, a committed date behaves like a forecast.

None of this has to be adversarial. These clauses make both sides state their assumptions, which is what a delivery plan is for.

Covering the gap while you wait

Even a well-supported date leaves a gap between signing and acceptance. The usual way to use it is to work on capacity that is available now: porting code, building and testing containers, preparing data pipelines, and running the smaller jobs that do not need a full rack. Plan that work on systems you can provision today, and confirm that any interim capacity has its own start date. Interim hardware that is also on pre-order gives you a second date to secure instead of shortening the wait.

On GMI Cloud, NVIDIA H100 and H200 are available now, from $2.60 and $3.20 per GPU-hour. If you plan interim work with us, raise it in the same conversation as your Blackwell pre-order: how the interim usage is priced, and where your data and checkpoints will sit when the Blackwell systems arrive.

Getting a written delivery plan from GMI Cloud

Our NVIDIA Blackwell lineup, including B200, B300, GB200 NVL72, and GB300 NVL72, is in pre-order and coming soon, for teams planning large-scale deployments that need maximum performance headroom. For GB200 NVL72, the GMI Cloud GPU page lists pricing from $8.00 per GPU-hour. Because the system is in pre-order, use that figure for planning, not as a rate you can provision at today. B300 is a hardware option on Prime Inference, our reserved inference service, and it is in pre-order as well.

On-demand and committed usage are priced differently, and our sales team sets up commitment options; the GPU pricing page covers current rates. GPU pricing covers compute, while networking and storage are provisioned separately based on your workload requirements. Our platform runs in the US, APAC, and EU; the quote should name the site and region for your Blackwell allocation. For a Blackwell reservation, the delivery window, milestones, acceptance event, and slip terms are set in the quote.

To make the conversation productive, bring:

• The exact system and quantity: GB200 NVL72 or GB300 NVL72 racks, or B300 capacity on Prime Inference
• Your target region and any data residency requirement
• The start date you want and the latest date that still works for your project
• The acceptance test you plan to run and the result that should start billing
• The remedies you need if a milestone moves
• Any interim work you want to run on H100 or H200 while you wait

FAQ

Is GB300 NVL72 the same as B300? No. B300 is the GPU, and it is used both in 8-GPU HGX B300 servers and inside the 72-GPU GB300 NVL72 rack. Confirm the system in the quote, not just the GPU name.

Does a pre-order guarantee a delivery date? Not by itself. A pre-order's delivery obligation comes from its agreed terms, including the date, conditions, and remedies. Allocation records and site milestones help assess whether the provider can meet that obligation.

Ask for the delivery plan in writing

Contact Sales with your unit, quantity, region, and target start date to discuss GB200 NVL72 and GB300 NVL72 pre-orders, B300 on Prime Inference, and commitment options. Ask for the delivery window, milestones, and slip terms in writing, and check them against the evidence and clauses above before you sign.