GPU Cloud Pricing: What Drives the Hourly Rate and Total Cost
GPU cloud pricing depends on hardware, allocation shape, term and included services. Learn how to compare rental quotes on the same scope before you commit.
What a GPU cloud price actually covers
A GPU cloud price is a quote for a defined scope of work. The number only means something once you know which hardware, how much of it, for how long and with which services attached.
Pacific frames GPU rental pricing around four drivers: the hardware selected, the quantity and shape of the allocation, the term, and the services included. Networking, storage, support and deployment requirements can change the discussion after the first number is on the table.
Picture an infrastructure lead holding three quotes for the same training run, each with an hourly figure, and none of them describing the same nodes, the same term or the same storage. Those quotes cannot be ranked yet.
Settle the scope first. Price comes second.
The unit of purchase sets the commitment
Ask how the unit of purchase is defined before you compare any rate. A listed GPU, a full node and a dedicated cluster carry different commitments, even when each is described per GPU-hour.
A per-GPU figure is easy to read, and it can still hide the real purchase. On Pacific's GPU compute catalog, the H100 listing shows sizes from one node upward, and each node carries eight GPUs.
A price for a single GPU cannot be assumed to represent a whole cluster. When a team moves from one node for debugging to many nodes for a full training job, the fabric, the placement and the terms can all shift, so get the cluster quoted as a cluster.
A good test: ask each provider to restate the quote in the unit you will actually deploy.
Hardware, memory and allocation shape
Hardware selection is the first driver of the rate you see. Memory per GPU then decides whether a model fits, how you shard it and how many nodes you need.
Pacific's catalog lists H100 SXM at 80 GB, H200 SXM at 141 GB and B200 SXM at 180 GB, each eight per node, with preview rates per GPU-hour. B300 SXM at 288 GB, GB200 NVL72 and GB300 NVL72 are priced by request.
Quantity and shape matter as much as the model. If your fine-tuning job fits in one node on a larger-memory GPU but needs two nodes on a smaller one, the lower hourly card can produce the larger bill once interconnect and run time are counted.
Price the job you run.
Term: on-demand and reserved are different buying motions
On-demand rental and a reservation are separate buying motions with separate commitments. Treat them as two products when you build a comparison.
The H100 listing on Pacific shows on-demand availability plus reserved terms of 7 to 30, 31 to 90 and 91 to 180 days by request. The listing also shows a per-day figure before tax, which helps you read the term as a budget line.
A research team running a short evaluation sprint needs different flexibility than a team keeping a cluster busy for a quarter, and each will weigh the hourly number against how much idle time it is willing to pay for.
For a structured way to weigh the two, read this guide on evaluating reserved vs on-demand GPU instances. Then apply it to your own run pattern.
What is included in the node
Two nodes with the same GPU can carry very different value. Check the interconnect, storage, access level and data handling before you treat two rates as equal.
Pacific's H100 listing describes dedicated bare-metal Linux nodes with SSH access, NVIDIA drivers and CUDA installed, local NVMe and InfiniBand interconnect, along with node specs for CPU and memory. Local NVMe is erased when the resource ends.
If your team expects to attach shared storage, run multi-node jobs over a fast fabric and hand off clean data at the end of a contract, each of those needs to be written into the quote or confirmed as out of scope.
Missing services become later line items. Ask about them now.
How to compare GPU cloud quotes on the same scope
A short worksheet makes GPU cloud pricing comparable across providers. Fill it in once, then send the same scope to every vendor.
Write down workload size in nodes and GPU model, expected run length, the time between runs, and dependencies such as storage, networking, support and deployment requirements. Include run length and the gaps between runs when you estimate total cost, because idle reserved time and repeated setup time both land in the budget.
A team that trains for a few days and then spends weeks on evaluation and data work may find a reservation sitting idle between runs, while a team with back-to-back jobs may find repeated on-demand setup adds up.
Then confirm the moving parts. Pacific's preview rates are indicative until Pacific confirms the allocation, and they do not come from a live inventory feed. Review revalidates capacity, the current quote and terms before you commit. A paid order is not a ready resource, so ask every provider when capacity becomes usable.
When reserved or dedicated capacity changes the math
Dedicated capacity shifts the comparison from hourly rates toward delivery terms. At cluster scale, configuration, acceptance evidence and the delivery window matter alongside price.
Pacific's reserved B300 cluster offer is a single-tenant bare-metal cluster. The Pod 32 reference comprises 32 HGX B300 nodes with a dedicated, configurable fabric, backed by witnessed factory, site and integration acceptance evidence.
Configuration and the delivery window are confirmed through a project review and a written offer, and commercial terms are discussed privately on a qualified call. The targeted deployment timeline is confirmed site by site and is never a guaranteed date, so read Pacific's reserved capacity page with that dependency in mind.
If a permanent facility is delayed, Pacific offers bridge capacity for delayed facilities. A platform team waiting on a stalled build can price the bridge term against the cost of pausing training until the permanent site is ready, then decide with both numbers on the page.
How to start with Pacific
Start narrow with one workload and one quote you can check. Pick the GPU type and size that matches your worksheet, then review the full commitment before you rent or request.
Use the Pacific Compute catalog to compare preview rates for H100 SXM, H200 SXM and B200 SXM, and request pricing for B300 SXM, GB200 NVL72 and GB300 NVL72. Availability shown is recorded by Pacific and confirmed when an order is accepted.
If the job needs a reserved term or a dedicated cluster, follow the steps in how to reserve GPU capacity and bring your worksheet to the qualified call. The budget conversation moves faster when the scope is already settled.
Run one job, check the quote, then commit.