Which Providers Will Discuss a Smaller GB300 NVL72 Allocation Before You Reserve a Full Rack?

Start with providers that either publish a unit smaller than a rack or quote GB300 NVL72 through a sales conversation where the unit is open. Together AI says on its GB300 NVL72 page that you can rent "any count, from a single node up to a full rack or multiple racks," priced per GPU or per rack. Nebius says its GB300 NVL72 is "available through the sales team only," so the allocation size is part of that conversation. AWS sells its NVL72-based P6e systems as UltraServers; for the GB200 generation, its P6 page lists 36-GPU and 72-GPU UltraServer types and says the individual four-GPU instances are only available inside UltraServers. At GMI Cloud, GB300 NVL72 is on pre-order and coming soon, like the rest of our NVIDIA Blackwell lineup, and the allocation unit, isolation, and expansion path are set in your quote.

Whichever provider you talk to, ask for two things in writing: a defined NVLink partition, stated as a number of compute trays on a named rack, and a contractual path from that partition to the rest of the same rack. A smaller allocation without the second part is a separate small deployment. You would still have to find, validate, and move into a full rack later.

What a partial GB300 NVL72 allocation is made of

NVIDIA's enterprise reference architecture for NVL72 AI factories builds a GB300 NVL72 rack from 18 compute trays, each with four Blackwell Ultra GPUs and two Grace CPUs, joined by nine NVLink Switch trays that give all 72 GPUs non-blocking connectivity at 130 TB/s in aggregate. The same document lists 1,152 GB of HBM per compute tray, or 288 GB per GPU. The compute tray, with its four GPUs, is the hardware building block; the smallest unit a provider rents is a separate commercial choice.

A provider turns trays into a sellable unit with NVLink partitioning. NVIDIA's Mission Control administration guide for GB200 and GB300 NVL72 systems describes partitions as hardware isolation domains inside one rack: half the compute trays can sit in one partition and half in another, and trays in different partitions can't establish multi-node NVLink connections with each other. Out of the box, the rack runs as one default partition holding all 72 GPUs.

A job larger than the allocated GPU count needs more capacity or a different parallel layout. A group spanning separate NVLink partitions can use a scale-out path only if the allocation and software support that topology.

How isolation works when other tenants share the rack

NVLink partitioning isolates the GPU-to-GPU fabric. NVIDIA's guide is direct about its limits: partitioning is "one piece of the puzzle" for multi-tenancy, and it "does not solve the isolation problem for other non-NVLink networks or storage." Several other parts of a shared rack need their own answers.

• Scale-out and storage networks. Each compute tray has its own network adapters, but the switches and storage behind them are shared infrastructure. Ask how your traffic is separated from other tenants on those paths.
• Multicast resources. NVIDIA's guide says GB200 and GB300 systems support up to 1,024 multicast teams, a global resource shared across all partitions, and recommends allocating fewer to smaller partitions. Ask how many your partition receives.
• Shared rack hardware. The NVLink Switch trays, power shelves, and liquid cooling serve the whole rack. Maintenance or a fault on any of them can affect every partition, so ask how rack-level maintenance is scheduled and announced to partial tenants.
• Who configures the partition. Partition membership is set in the rack's NVSwitch control plane, which the provider operates. Ask for the partition membership to be recorded, so you can confirm your trays match what you are paying for.

On GMI Cloud, all GPU resources are dedicated, single-tenant instances: the GPUs in your allocation are not shared with other users. How a partial GB300 NVL72 allocation is laid out at the switch, network, and rack level is part of the pre-order discussion and belongs in the quote.

What a smaller allocation can prove, and what it can't

A partial allocation is a strong test of anything that happens inside a tray or a partition. It is a weak test of anything that depends on all 72 GPUs working as one NVLink domain. Keeping that line clear stops a successful pilot from being read as proof the full rack will behave the same way.

What it can prove:

• Whether your model state fits in GB300 memory per GPU, with your real framework overhead
• Whether your software stack runs cleanly on Grace CPU hosts, which use the Arm architecture, including images, drivers, and libraries
• How parallel groups that fit inside the partition behave over NVLink
• Per-GPU throughput for inference replicas that each fit in the partition
• How the provider handles onboarding, monitoring, support requests, and a hardware incident

What it can't prove:

• Scaling behavior when a single job spans the full 72-GPU NVLink domain
• Collective communication performance at 72 GPUs, which is a main reason to choose NVL72 over eight-GPU servers
• Full-rack power and thermal behavior. NVIDIA's reference architecture states that a full GB300 NVL72 rack can require up to 142 kW, and a partition only exercises part of that load
• Failure rates and job restart behavior at full-rack scale

For inference fleets made of independent replicas, the gap between the two lists is small, and a partition can carry real production traffic. For training runs designed around the full NVLink domain, the partition proves the software and per-GPU numbers, and the full-domain result stays an assumption until you have the rack.

Scale-up terms to agree before you sign the smaller allocation

The time to negotiate the path to a full rack is before the small allocation starts, while the rest of the rack is still unassigned. These terms cover growth in size. How a short commitment converts into a longer one is a separate negotiation with its own terms.

1. A claim on the rest of the same rack. Whether the remaining trays are held for you, offered to you first, or sold freely. If they are held, for how long, and at what cost. If they are offered first, how much notice you get before they go to another tenant.
2. The price of added GPUs. Whether GPUs added to reach the full rack are priced at your existing per-GPU rate, at a rate agreed now, or at the rate in effect on the day you expand.
3. Expansion in place. Whether growing to the full rack is done by repartitioning the same hardware, so your existing trays, local data, and attached storage stay where they are. Networking and storage are often provisioned and billed separately from GPU compute, so make sure the storage attached to your partition carries over as well.
4. Repartitioning downtime. NVIDIA's Mission Control guide warns that altering NVLink partitions is disruptive to workloads and calls for a maintenance window with user workloads paused, followed by a GPU reset or compute tray reboot. Expansion will interrupt the jobs on your existing trays, so agree on when that window happens, how long it runs, how much notice you get, and whether expansion time is billed.
5. The fallback when the rack fills. If the remaining trays go to another customer, what the provider offers instead: a full rack in the same data center and network fabric, a rack elsewhere, or nothing. A fallback outside your current fabric means migrating data and revalidating, which is exactly what the smaller start was meant to avoid.

Providers and what they publish about allocation size

What a provider publishes about allocation size tells you who expects this conversation. Whatever a page says, ask each provider for the allocation unit in writing.

A provider that sells only full racks is still worth a conversation if its full-rack terms suit you, but it isn't the answer to this particular question. A provider that sells small slices without an expansion path suits a permanent small deployment better than a first step toward a rack.

Starting smaller with GMI Cloud

Our GPU page positions GB300 NVL72 for long-context and high-capacity model training, next-generation multi-node clusters, and large-scale reasoning and high-density AI systems. It is on pre-order and coming soon, so the allocation unit, including whether you can start below a full rack, is settled with our sales team and written into your quote before delivery, not capacity you can open today.

Our pricing is built so commitments can follow workload maturity. Usage-adaptive pricing lets pricing flex as workloads stabilize, avoiding forced lock-ins or premature long-term commitments, and commitment-based pricing reduces unit GPU costs through reserved capacity and sustained deployments. GPU pricing covers compute; networking and storage are provisioned separately based on your workload, so both belong in the same quote as the partition.

Bring the following to the conversation: the starting size you have in mind, the full-rack target you expect to reach, the parallel group sizes your jobs need, and the five scale-up terms you want written into the agreement. NVIDIA H100 and H200 are available now, from $2.60 and $3.20 per GPU-hour, for work that can't wait for Blackwell delivery.

FAQ

Is a single compute tray enough to evaluate GB300 NVL72?

For memory fit, software compatibility on Grace hosts, and single-replica inference throughput, four GPUs on one tray answer most questions. For anything that depends on NVLink communication across trays, a partition of several trays gives more useful results, and only the full rack answers questions about the 72-GPU domain.

Can a 36-GPU partition run the same training job we plan for the full rack?

Not without changes. Every NVLink-connected parallel group has to fit within 36 GPUs, but that alone isn't enough: the job's total model state and activations also have to fit in half the GPUs and half the memory. A job planned for 72 GPUs runs on 36 only with a different layout, so throughput and scaling numbers from the partition won't transfer directly to a full-rack layout.

Plan a smaller GB300 NVL72 start

Tell our team the starting size you have in mind and the full rack you expect to grow into. The allocation unit, isolation, and expansion terms are set in your GB300 NVL72 pre-order quote. Contact our sales team to start.