How to Rent GB200 GPUs From Pacific's Catalog
How to rent GB200 GPUs from Pacific's catalog: when GB200 NVL72 fits, what to confirm before commit, and how by-request pricing and review work.
What GB200 NVL72 is, in one line
GB200 NVL72 is a rack-scale Grace Blackwell system with 72 GPUs in one liquid-cooled rack. Arm-based Grace CPUs host the Blackwell GPUs, so the host side differs from the x86 nodes most teams run today.
In Pacific's catalog, GB200 NVL72 is listed by request, with 186 GB per GPU, 72 GPUs per system and pricing by request. That listing tells you the shape of the system. It does not tell you what is free to allocate right now.
If you are an infrastructure lead planning a large training run, treat the whole rack as the unit you rent, because the value of GB200 comes from its 72 GPUs working as one domain. One rack. One domain.
When renting GB200 makes sense
Rent GB200 when your workload needs many GPUs sharing one fast memory domain. A model that fits comfortably on an 8-GPU node rarely justifies the extra integration work.
H200 and B200 are listed in the catalog as 8 GPUs per node with on-demand preview rates, which makes them the easier starting point for fine-tuning, inference serving and most experiments. If you are weighing those two, read H200 vs B200 for training and inference first.
A good test: if your team spends its days fighting cross-node communication on 8-GPU clusters for large mixture-of-experts or long-context training, a rack-scale system may shorten each step enough to change the cost of the finished job.
Start narrow. If the job fits on nodes, rent nodes through Pacific's GPU rental options.
Starting on Pacific Compute
Begin in the Pacific Compute catalog by picking a GPU type and size. For GB200 NVL72, that choice opens a by-request path instead of an instant order.
Review the full commitment before you request anything. Pacific's review step revalidates capacity, the current quote and terms before commit, so the request turns into a conversation about a specific allocation.
In practice, an ML lead selects GB200 NVL72, notes the system count and term the training plan needs, and raises it on project review with the workload, the target window and any facility constraints attached.
A paid order is not a ready resource. Treat confirmation from Pacific as the moment capacity becomes real for your plan.
What to confirm before you commit
Get every rack-scale assumption written down before you sign anything. The offer or review should state the configuration baseline you are actually renting.
Ask how the 72 GPUs connect inside the rack and how each rack connects to storage, to other racks and to your network. Interconnect expectations decide whether a multi-rack plan works at all.
Confirm cooling and facility fit. GB200 NVL72 is liquid-cooled at rack scale, so ask where the system sits, who owns the power and cooling interfaces and what has been verified versus assumed.
Before commit, have your platform engineer check that your container images, drivers and libraries build for Arm Grace hosts and support Blackwell GPUs, because an x86-only dependency can stall the first days on a rented rack. Then confirm acceptance criteria and the delivery window as stated in the written offer or review.
How GB200 pricing conversations work
GB200 NVL72 pricing is by request, so there is no catalog number to compare. Preview rates shown for other GPU types are indicative until Pacific confirms the allocation.
Do not compare an hourly figure for a rack against an hourly figure for a node. Compare cost per finished job: total spend to reach a trained checkpoint or a served workload, including idle time, data movement, storage, engineering hours and failed runs.
A useful budget exercise is to price the same training run on 8-GPU nodes and on a GB200 rack, then ask Pacific which assumptions in the rack quote are confirmed and which still depend on review. For the full set of factors, see GPU cloud pricing and cost drivers.
When reserved B300 is the better path
Look at reserved B300 when you need a dedicated cluster for a long program. Renting from the catalog suits a defined job, and a reservation suits a roadmap.
Pacific's reserved B300 clusters are single-tenant bare metal, with a Pod 32 reference of 32 HGX B300 nodes on a dedicated, configurable fabric. Configuration and delivery window are confirmed through a project review and written offer, backed by witnessed factory, site and integration acceptance evidence.
If you are choosing between rack-scale Grace Blackwell and an 8-GPU B300 form factor for a reserved order, read GB200 vs B300 for reserved capacity buyers. This guide stays on the rental path.
The sibling guide for H200 rental
If your workload fits on 8-GPU nodes, the H200 path is simpler to start. H200 SXM is listed with 141 GB per GPU and 8 GPUs per node, shown with on-demand preview rates.
The steps in how to rent H200 GPUs mirror this guide: pick type and size, review the commitment, confirm the allocation, then plan onboarding.
Running a pilot on H200 while a GB200 request goes through review gives your team a working baseline, a real cost per finished job and a tested software stack to compare against the rack quote.
How to begin with Pacific
Open the catalog, pick GB200 NVL72 and raise it on project review with your workload attached. Bring the training plan, the term you need and your hosting assumptions.
Expect Pacific to revalidate capacity, the current quote and terms in review before commit. Until then, plan around confirmed facts and keep assumptions labeled as assumptions.
If a facility delay is the reason you need rented capacity, say so early, because Pacific also offers bridge capacity and colocation for owned hardware, and the right path depends on your situation. Get the offer in writing.