How to Rent GB300 GPUs: NVL72 Racks by Request
How to rent GB300 NVL72 racks from Pacific: scope rack count, term and region, submit a request, and learn what review confirms before SSH access opens.
Who GB300 NVL72 fits
GB300 NVL72 fits work that needs one large NVLink domain kept busy. An NVL72 rack is a single NVLink domain of 72 GPUs, and Pacific lists GB300 with 288 GB per GPU in dedicated NVL72 racks.
GB300 pairs Blackwell Ultra GPUs with Arm-based Grace CPUs. That pairing matters most when your model or job gains from many GPUs sharing one rack-scale fabric without crossing a network between nodes.
A team serving a large mixture-of-experts model that spills past one node, and that loses throughput every time traffic crosses a network hop, is the buyer who gets the most from a full rack.
An 8-GPU node is often the better unit. If your training run shards cleanly across nodes, or your inference service fits in one node's memory, a B200 or B300 node on Pacific's compute catalog is easier to scope. Start narrow when you can.
The rental path, step by step
Renting GB300 runs through a request and review, with no instant checkout. The catalog does not show GB300 as available now, so plan for review before any access.
Open the Pacific compute catalog and select GB300 NVL72. Choose 1, 2 or 4 racks, which maps to 72, 144 or 288 GPUs, then pick a duration of 1, 2, 3 or 6 months and a region in Asia Pacific, North America or Europe. Every option is by request, and pricing is by request.
Submit the request. Review then revalidates capacity, the current quote and terms before you commit, and once the order is accepted and the resource is ready, you reach dedicated bare-metal Linux hosts over SSH with NVIDIA drivers and CUDA installed.
A paid order is not a ready resource. An infrastructure lead booking a training window should hold the launch date loosely until review comes back, since the catalog shows recorded availability and order acceptance is what makes it real. For the general flow across GPU types, see Pacific's GPU rental overview.
What to scope before you request
Write down five decisions before you submit a GB300 request. A clear request gives review something concrete to confirm, and it keeps your own budget honest.
Rack count and term come first. Pick the smallest rack count that holds your job, and pick a duration that matches the run plus setup and teardown time. Region follows from where your data lives and where your users or downstream systems sit.
Data movement and storage around the rack come next. Local NVMe is erased when the resource ends, so plan where datasets land, where checkpoints go and how results leave the rack before the term closes.
Software readiness on the Grace Arm CPU is the step teams skip. Check that your containers, Python wheels, custom kernels and any host-side tools build and run on aarch64, then confirm the job plan keeps all 72 GPUs in the NVLink domain busy with input ready.
What gets confirmed with the order
The catalog listing tells you the shape of the offer before any order. Capacity, the quote and terms are what review confirms.
The GB300 NVL72 listing states dedicated NVL72 racks, SSH access to dedicated bare-metal Linux hosts, NVIDIA drivers and CUDA installed, and local NVMe erased when the resource ends. The interconnect is listed as an NVLink rack, confirmed with your order.
Treat everything else as open until review closes it. A buyer who has already told finance a start date should send review the exact rack count, term and region, then wait for the confirmed quote and terms in writing before moving any production timeline.
For what moves a rack-scale quote, read the GB200 price and cost drivers guide. Rack count, term length and region shape a GB300 request in much the same way.
Choosing between GB300 and GB200 NVL72
GB300 and GB200 NVL72 share the same listing shape on Pacific's catalog. Both are dedicated NVL72 racks with 72 GPUs per system, by request, in 1, 2 or 4 racks across the same regions and durations.
The visible difference is memory per GPU: 288 GB on GB300 against 186 GB on GB200. Choose GB300 when your model, context length or batch plan needs that extra memory per GPU to avoid extra sharding.
A research team that already fits comfortably on GB200, with headroom for checkpoints and activations, gains little by waiting on GB300 and can follow the GB200 rental path instead.
If you are weighing rack-scale GB300 against 8-GPU B300 nodes for a longer reservation, this GB300 vs B300 comparison for reserved capacity buyers covers that system choice.
Questions to ask before you commit
Ask review specific questions that turn a request into a confirmed order. Vague questions get vague answers, and your budget pays for the gap.
Ask which region and rack count review has confirmed, and whether the interconnect matches what your order states. Ask how the term start is defined and what happens to data on local NVMe at the end.
Ask how access is handed over, which drivers and CUDA versions are installed, and whether your Arm builds can be tested early. An ML lead planning a multi-week training run should also ask what the restart plan looks like if a host needs attention mid-term.
Do not compare a GB300 quote against a node preview rate. They describe different units, so compare like with like.
When a reserved commitment fits better
Step up to a reserved commitment when a rental term stops matching the plan. A six-month rental is the longest GB300 term the catalog lists.
If your roadmap needs dedicated capacity past that window, or a cluster delivered with acceptance evidence you can audit, read how to reserve GPU capacity before you extend renting term by term.
Pacific's reserved B300 offer is a single-tenant bare-metal cluster built on a Pod 32 reference of 32 HGX B300 nodes with a dedicated, configurable fabric. Configuration and delivery window are confirmed through a project review and written offer, and the deployment timeline is confirmed site by site, never as a guaranteed date.
A good test: if you would renew the same racks twice, price the reservation.