GB200 Price: What Moves the Cost of GB200 NVL72
GB200 price explained: why GB200 NVL72 is quoted per rack, what moves the cost of rental or reserved capacity, and how to compare quotes line by line.
Why GB200 has no list price
GB200 NVL72 is sold by the rack, so there is no single per-GPU price. Pacific's catalog lists it by request, with pricing by request, because the quote depends on scope you have not chosen yet.
The unit of purchase is one NVL72 system: 72 GPUs, 186 GB per GPU, in a dedicated rack. Sizes run 1, 2 or 4 racks, which means 72, 144 or 288 GPUs.
A finance lead who searches for a GB200 price and expects a per-GPU hourly figure will find a request path instead, because the number only exists once rack count, term and region are fixed.
Price follows scope. Fix the scope first.
The four inputs that size the quote
Rack count, term, region and tenancy set most of a GB200 quote. Each one is a choice you make before Pacific prices anything, and each one moves the total.
Start narrow on rack count. One rack is a single NVLink domain of 72 GPUs, while two or four racks add capacity and also add questions about how the racks connect and whether your job spans them.
Terms run 1, 2, 3 or 6 months, all by request. Longer terms change the commercial shape of the quote, so ask for the term your plan can actually fill.
Region options are Asia Pacific, North America and Europe. A platform team serving users in Europe while keeping training data in North America should price both regions, because data movement and residency rules can matter as much as the rack itself.
Tenancy is dedicated. You get SSH access to bare-metal Linux hosts with NVIDIA drivers and CUDA installed, and you pay for the whole rack whether it is busy or idle.
Interconnect and Arm software effort
The NVLink rack is the reason to buy GB200, and it carries extra scope. The catalog describes the interconnect as an NVLink rack confirmed with your order, so the exact fabric detail belongs in the quote you review.
Ask what sits outside the rack. Data has to reach the hosts, multi-rack jobs have to talk across racks, and checkpoints have to land somewhere, so confirm which of these are included and which are quoted separately.
GB200 pairs Blackwell GPUs with Grace CPUs, which use the Arm architecture. An ML engineer porting a training stack built on x86 hosts should budget time to rebuild container images, check Python wheels and custom kernels, and rerun a short job on Arm before the first full run starts using rack time.
That effort is part of the price. Ask Pacific what is preinstalled on the hosts and what your team has to bring.
Idle rack cost and utilization
A rack-scale system pays off only when your job uses the rack. Cost per finished job is the number that matters, and an idle NVLink domain raises it fast.
A team fine-tuning a model that fits comfortably on one 8-GPU node will pay for 72 GPUs while most of them wait, which makes the rack an expensive way to run a modest job.
A good test: list the jobs that need fast GPU-to-GPU bandwidth across more than one node's worth of GPUs. If that list is short, the rack will sit partly empty for much of the term.
If you plan to pack several smaller jobs onto one rack, confirm your scheduler and tooling can split it cleanly on these hosts. Shared use of a dedicated rack is your job to manage.
How to read a GB200 quote line by line
Read every GB200 quote as a per-rack commitment with named inclusions. Then check which lines are firm and which are confirmed only when the order is accepted.
Start with the unit: rack count, GPUs per rack, term and region. Then find what is included per rack, such as dedicated bare-metal hosts, SSH access, NVIDIA drivers and CUDA, and the NVLink rack interconnect.
Look for the lines marked as confirmed with your order. Interconnect detail and availability both sit there, and Pacific's review step revalidates capacity, the current quote and terms before you commit.
Plan the exit. Local NVMe is erased when the resource ends, so a buyer running a 3-month training run needs a copy-out step for checkpoints and logs scheduled before the term closes. For the general factors behind any GPU quote, see GPU cloud pricing cost drivers.
When an 8-GPU node is the better unit
A job that fits on 8-GPU nodes often costs less to finish there than on a 72-GPU rack. Compare cost per finished job; the per-GPU figure tells you little on its own.
B200 SXM nodes in Pacific's catalog show on-demand preview rates, with 1-node and 2-node sizes rentable directly, while B300 SXM nodes run by request in sizes of 1 to 8 nodes. A research team running inference, fine-tuning or training that scales well over InfiniBand can often start on a node, prove the workload, then decide whether a rack is justified. The B200 price and cost drivers guide covers that HGX-class path. Use the HGX B300 specs buyer checklist to confirm the B300 node configuration.
GB300 NVL72 is the adjacent rack option in the same catalog, listed the same way with 288 GB per GPU and pricing by request. If you are choosing between rack-scale GB200 and node-based B300 for a longer commitment, this GB200 vs B300 comparison for reserved capacity walks through the system choice. The GB300 vs B300 comparison for reserved capacity covers the Blackwell Ultra choice. For the B300 commitment, see what to put in writing when reserving B300 GPUs.
Compare the options side by side in the Pacific Compute catalog before you request a rack.
Questions to ask before you commit
Ask these questions before a GB200 rack goes into your budget. Each one closes a gap between the quote and the job you will actually run.
Ask how the NVLink rack interconnect is confirmed for your order, and how racks connect if you take two or four. Ask what storage, data ingress and networking sit around the rack, and which of them are priced separately.
Ask what software is on the hosts and what your team must rebuild for Arm. Ask how availability is confirmed, when the resource becomes ready after acceptance, and what happens to local data at the end of the term.
Do not compare a GB200 quote to an on-demand node rate without normalizing for term, rack count and utilization. A paid order is not a ready resource, so plan your schedule around confirmed readiness.
Get a current GB200 quote from Pacific
A current GB200 quote starts in the catalog and ends in review. The scope decisions take most of the time.
Open the Pacific Compute catalog, pick GB200 NVL72, then choose rack count, region and term. Submit the request, and Pacific confirms the allocation and returns the quote for review before anything is committed.
An infrastructure lead preparing a budget request should bring a short workload description, the target region, the term and a note on Arm readiness to that review, so the quote reflects the job and the rack it actually needs.
For the full step-by-step path, read how to rent GB200 GPUs. For node-based options and other GPU types, see GPU rental at Pacific.