B200 Rental: How to Rent B200 GPUs on Pacific
How to rent B200 GPUs from Pacific's catalog: pick type and size, choose on-demand or a reserved term, and confirm drivers and region before you commit.
What you are renting with B200 SXM
B200 SXM is NVIDIA's Blackwell generation GPU, listed with 180 GB of HBM and 8 per node. In Pacific's catalog you rent it as a dedicated bare-metal node, so the whole host is yours for the term.
This guide covers one path: how to pick, book and confirm B200 on Pacific Compute. If you are still weighing providers against each other, read this guide to evaluating B200 cloud rental providers first. Come back when you are ready to book.
A typical reader is an ML lead with a training run or fine-tuning backlog that has outgrown its current nodes, who needs to know what the listing promises and what still gets confirmed after an order. The rest of this guide follows that order.
When B200 is the right rental
Rent B200 when memory per GPU is the real constraint on your work. If your jobs already fit comfortably on H200, the extra memory may not lower what each finished job costs you.
B200 SXM lists 180 GB per GPU against 141 GB on H200 SXM. That gap matters for larger models, longer contexts and bigger batches that would otherwise need more sharding across GPUs.
A team running a recurring fine-tuning job on a mid-sized model may find H200 nodes finish the same work with less migration effort, while a team training a larger model gets more from the B200 headroom. If the choice is still open, work through H200 vs B200 for your workload before you book.
If H200 wins, the steps to rent H200 GPUs follow the same dedicated bare-metal pattern. If you need a fixed large cluster for many months, a rental may be the wrong shape. The reserved B300 section below covers that case.
Check software readiness for Blackwell first
Blackwell GPUs need recent NVIDIA driver and CUDA releases, so your stack must keep up. A good test is whether your framework build and container image already list Blackwell support.
The B200 listing says NVIDIA drivers and CUDA are installed on the host. Ask which versions, then check them against what your framework, compiled kernels and container base image expect.
Before you book, have one engineer pull your production image, note the CUDA version it was built against, and list any custom kernels or extensions that would need a rebuild for the new architecture. That list is your readiness check.
Do this before the clock starts. A rented node that sits idle while you rebuild images is budget spent on setup.
Start on Pacific Compute and pick a term
The catalog flow is short: pick a GPU type and size, then review the commitment. The review step does the real work, because it revalidates capacity, the current quote and terms before you commit.
Open the catalog, select B200 SXM, and read the full listing: dedicated nodes, multiple regions, SSH access to bare-metal Linux hosts, InfiniBand interconnect and local NVMe erased when the resource ends. For background on how Pacific handles dedicated GPU rental, read that page next to the listing.
On-demand suits short proof work and bursty jobs. If you are testing a new training recipe, start on-demand for the proof run, then move to a reserved band once you know how many nodes a full run needs and for how long.
Reserved terms of 7 to 30, 31 to 90 and 91 to 180 days are available by request. Larger sizes can move from rent to by request, so a bigger node count may route to a conversation before it books. The steps to reserve GPU capacity cover what that request involves.
What to confirm before you commit
Treat the listing as the starting record and confirm each line before you pay. A paid order is not a ready resource, and availability shown is confirmed only when Pacific accepts the order.
Check the node configuration on the listing: GPU type, count per node and dedicated host. Confirm InfiniBand covers the node count you plan to run across. Confirm driver and CUDA versions against your image.
Confirm the region. The listing shows multiple regions, so ask which one your allocation lands in and whether it suits your data location and access rules.
Ask where the support boundary sits: what Pacific handles on the host and what stays with your team. An infrastructure lead planning a long training run should record the delivery confirmation received at order acceptance and schedule the first job from that confirmation. Local NVMe is erased at the end, so plan checkpoint exports early.
How the pricing conversation works
Preview rates on the listing are indicative until Pacific confirms the allocation. They do not come from a live inventory feed, and review revalidates the current quote before commit.
Do not compare GPUs on the hourly figure alone. Compare cost per finished job: node hours per run, time lost to setup and idle nodes, and how often a job restarts.
A team moving a training run from H200 to B200 should estimate the full run on both, including image rebuilds and data transfer, before deciding which line in the budget actually goes down. That estimate is the number to bring to Pacific.
The GPU cloud pricing cost drivers guide walks through the inputs that move it. Term length, node count and region all belong in the conversation.
When reserved B300 is the better path
Reserved B300 fits when you need a fixed cluster for a long program of work. It follows a different process: a single-tenant bare-metal cluster scoped through a project review and a written offer.
Pacific's Pod 32 reference comprises 32 HGX B300 nodes with a dedicated, configurable fabric, backed by witnessed factory, site and integration acceptance evidence. In the catalog, B300 SXM lists 288 GB per GPU and is by request.
If your roadmap needs the same large cluster for many months and your team wants acceptance evidence before go-live, a reserved B300 cluster is worth a qualified call before you stack B200 rental terms. Configuration and delivery window are confirmed in the written offer.
Commercial terms are discussed privately. The deployment timeline is a target confirmed site by site, and it is never a guaranteed date.
How to begin with Pacific
Begin with one proof run on the smallest B200 size that answers your question. That run tells you whether your image works and how long a full job takes.
Open the catalog, select B200 SXM, read the listing, and use review to confirm capacity, quote and terms. Raise your questions on region, InfiniBand, driver versions and support before you pay.
When the proof run is done, bring the measured job length and node count back to Pacific so the next request, whether a reserved band or a larger by-request size, is scoped on evidence. Start narrow. Then scale on proof.