How to Rent H100 GPUs From a Compute Catalog
How to rent H100 GPUs from Pacific's catalog: pick type and size, choose on-demand or a reserved term, and confirm node config before you commit.
What you are renting with H100 SXM
H100 SXM is NVIDIA's Hopper generation GPU with 80 GB of HBM, listed in 8-GPU nodes. The node is the unit you rent, so plan your job around eight GPUs sharing one machine and one interconnect domain.
In Pacific's catalog, each H100 listing is a dedicated bare-metal Linux node with SSH access, NVIDIA drivers and CUDA installed, and an InfiniBand interconnect. Local NVMe is erased when the resource ends.
A training lead who has been queuing behind other teams on a shared cluster gets a machine they log into directly, with the stack already in place and the hardware dedicated to their work. Treat local disk as scratch. Push checkpoints and outputs to storage you control before the term ends.
When H100 is the right rental
H100 fits when your model and batch already fit in 80 GB per GPU. That is a narrow test, and you can run it against your current code and configs before you rent anything.
A team fine-tuning a mid-sized model on a stable recipe for a few weeks, with known memory use and a fixed eval set, is the classic H100 renter. The hardware is well understood, the software support is mature, and your existing images likely already target it. Start narrow.
If you keep hitting memory limits, sharding more than you want, or serving long contexts, H200's 141 GB per GPU may save engineering time on the same recipe. The H100 vs H200 comparison walks through that trade in plain terms, and the how to rent H200 GPUs guide covers the same catalog path for the larger card.
For long programs that need a dedicated cluster on a newer generation, read the reserved B300 section below before you sign a long H100 term.
Start in the Pacific Compute catalog
The catalog path has two steps: pick a GPU type and size, then review. Review is where the full commitment shows up before you rent or request.
Open Pacific Compute, choose H100 SXM, and pick how many 8-GPU nodes you need. If you are still sorting out which rental option fits, the GPU rental page sets out Pacific's rental options in one place.
Review revalidates capacity, the current quote and terms before you commit. An infrastructure lead sizing a four-node fine-tuning run should read that review screen the way they would read a purchase order, line by line, before anyone on the team clicks through.
Availability shown in the catalog is recorded by Pacific and confirmed when an order is accepted. It is not a live inventory feed. Plan as if the number can move until acceptance, and keep your internal start date tied to that confirmation rather than to the click that placed the order.
On-demand or a reserved term band
Choose the term by how long the work actually runs. On-demand suits short or uncertain jobs where stopping early has real value.
For H100, Pacific lists reserved terms of 7-30, 31-90 and 91-180 days, by request. A reserved term trades flexibility for a committed window, so pick the band your plan supports with evidence.
A good test: write down the job sequence, the expected run count and the restart buffer, then see which band covers it without padding. A platform lead planning a quarter of continuous fine-tuning and eval cycles across two teams has a different answer than a researcher validating one idea over a long weekend.
If the work clearly outgrows the longest band, read the guide to reserving GPU capacity before you request anything.
What to confirm before you commit
Confirm the node you will get matches the node your job needs. The listing describes the standard H100 setup, and your workload decides which details matter.
Check the node config: GPU count per node, CPU, system memory and local NVMe size. Check InfiniBand: ask whether your multi-node job will span nodes on the same fabric and what topology details Pacific can share for your NCCL setup.
Check the driver and CUDA versions against your container images before day one. A team that pins a CUDA version in its training image and finds a mismatch after the first node comes up loses a day rebuilding, so ask for the installed versions during review.
Ask where Pacific's support boundary sits, which layers Pacific handles and which stay with your team, and who you contact for each class of issue. Then confirm delivery. A paid order is not a ready resource, so ask how Pacific confirms the nodes are allocated and reachable over SSH, and how that confirmation shows up in writing.
How the pricing conversation works
Preview rates in the catalog are indicative until Pacific confirms the allocation. Treat them as a starting budget line and let review set the figure you can put in a purchase order.
Do not compare GPU hours in isolation. Compare cost per finished job: hours to complete the run, restarts, idle time between experiments and engineering time spent on setup. The guide to GPU cloud cost drivers breaks those inputs down.
A finance partner asking why the team chose one card or term over another will accept the answer faster when you show a single job's total cost, from first launch to final checkpoint, side by side across options. Keep the inputs written down. That record is your proof at renewal.
When reserved B300 is the better path
Reserved B300 fits long programs that need a dedicated cluster, beyond a few rented nodes. It is a project with its own review and a written offer.
Pacific's B300 reservation offer is a single-tenant bare-metal cluster, and the Pod 32 reference is 32 HGX B300 nodes with a dedicated, configurable fabric. Configuration and delivery window are confirmed through a project review and written offer, commercial terms are discussed privately on a qualified call, and the deployment timeline is a target confirmed site by site.
If your roadmap has a frontier-scale training run that would otherwise mean stitching together H100 rentals for months, compare the paths with this look at H100 versus reserved B300 capacity before you lock in a long H100 term. You can rent H100 for current work while a B300 review runs.
How to begin with Pacific
Start with one job you understand well and size the rental around it. That job becomes your proof for anything larger, because you can show hours, restarts and outcomes on hardware you already know.
Open the Pacific Compute catalog, select H100 SXM, pick the node count and term, then read the review screen against the checklist above. A lead who arrives with the job plan, the CUDA version pinned in their image and a clear answer on multi-node needs turns review into a quick confirmation.
If the conversation points toward a longer reservation or B300, ask Pacific for a project review and bring the same plan. Then rent the nodes you sized.