H200 Price Factors: What Moves a Reserved Capacity Quote

Picture a platform lead at a growing AI startup who types "h200 price" into a search bar on Monday, collects three quotes by Friday, and still can't tell which one is cheaper. One quote is per GPU-hour, one is per node per month, and one rolls storage and support into a single term figure. The numbers sit side by side in a spreadsheet, yet they describe three different things.

This post is for founders, ML leads and procurement owners who need H200 capacity on a rental or reserved basis and want to understand what moves the quote. You won't find price numbers here. You'll find the cost drivers, a worked example using Pacific's catalog, and checklists for reading and comparing quotes. If you want background on how dedicated rental and reserved terms work in general, start with Pacific's guide to dedicated GPU rental and reserved terms.

1. Pin down the unit of the quote

Before you compare anything, normalize the unit. H200 capacity is usually quoted in one of three ways: per GPU-hour, per node-hour or node-day, or as a total for a fixed term. All three can be fair. The trouble starts when you compare a per-GPU figure from one provider against a per-node or per-term figure from another without converting.

The second question is what the unit includes. A GPU-hour figure from one seller might cover only the accelerator, while another seller's node figure covers CPUs, system memory, local storage, interconnect and support. Write down what is bundled for each quote before you do any math.

• Unit: per GPU-hour, per node-hour, per node-day, or per term.
• Minimum commitment: shortest duration and smallest node count you can buy.
• Bundled hardware: CPU, system memory, local NVMe, interconnect.
• Bundled services: access method, drivers and CUDA, support scope.
• Extras billed separately: anything outside the node, such as shared storage or data transfer.

2. Know when 141 GB per GPU changes the math

The H200's headline feature is 141 GB of memory per GPU. That memory only matters to your budget if your workload is limited by memory. When it is, the effect on cost can be large, because it changes how many GPUs and nodes the job needs.

Larger models, longer context windows and bigger KV caches all eat memory. If a model plus its working set fits on fewer H200 GPUs than it would on a smaller-memory part, you may run with a lower parallelism degree, fewer nodes and less cross-node traffic. Fewer nodes for the same job can outweigh a higher per-GPU rate. On the other hand, if your model fits comfortably in less memory and the job is compute-bound, the extra memory may sit idle and you are paying for headroom you don't use.

• Measure peak memory per GPU on your real model, batch size and context length.
• Check whether the larger memory lets you drop a node or reduce tensor or pipeline parallelism.
• Estimate the KV cache needed for your longest production context.
• If memory never gets close to the limit, weigh the job on throughput and rate instead.

3. Read the node shape, using Pacific's catalog as the example

H200 capacity is normally sold as 8-GPU nodes, and what surrounds those GPUs affects both price and usefulness. In Pacific's GPU compute catalog of dedicated bare-metal GPU nodes, the H200 SXM listing (checked in November 2026) describes each node this way:

• 8 H200 SXM GPUs with 141 GB each.
• 192 vCPU and 2 TB of system memory per node.
• About 30 TB of local NVMe per node.
• 3.2 Tb/s InfiniBand interconnect.
• Dedicated nodes with SSH access to bare-metal Linux hosts, NVIDIA drivers and CUDA installed.
• Local NVMe erased when the resource ends.

When you compare this to another quote, check each line. A node with less system memory or less local storage may look cheaper, then cost you time in data loading or checkpointing. A node without a fast interconnect may be fine for single-node inference and painful for multi-node training. Dedicated bare metal also means no noisy neighbors, which matters for predictable job times.

4. Term band, duration and node count

The term band is often the biggest lever you control. In Pacific's catalog, H200 shows an on-demand preview rate, plus reserved terms of 7 to 30 days, 31 to 90 days and 91 to 180 days by request. Each longer reserved band shows a lower preview rate than the band before it, and all reserved bands show a lower preview rate than on-demand. Durations run from 1 day to 6 months.

The trade-off is simple: a longer band lowers the rate and raises your commitment. Match the band to how confident you are in the workload. If you know a training run or an inference deployment will last for months, a longer band usually makes sense. If you're still testing, shorter terms protect you from paying for capacity you stop using.

Node count matters too, both for price and for how you buy. Pacific lists 1, 2, 4 and 8 node H200 sizes as rentable directly. A 16-node size, which is 128 GPUs, is by request. Larger sizes usually go through a review because the provider has to confirm contiguous capacity on the same fabric.

• Pick the shortest band that covers the work you're confident in.
• Ask whether you can extend into a longer band later, and on what terms.
• Check whether your node count is directly rentable or needs a request.
• Remember that preview rates are indicative until the allocation is confirmed.

5. Region, support scope and what gets confirmed at order

Region affects price, latency to your users and data residency. Pacific's H200 listing offers Asia Pacific, North America and Europe. If your data has to stay in a jurisdiction, settle that first, then compare price within the allowed regions.

Support scope is easy to skip and expensive to discover later. A bare-metal node with drivers and CUDA installed gives you a working base, and you own everything above it. Ask who handles hardware faults, how a failed node gets replaced, and whether replacement time is covered in writing.

Also be clear on what is confirmed and when. In Pacific's flow, you pick a GPU type and size, then review the full commitment before renting or requesting. That review revalidates capacity, the current quote and terms before you commit, and availability shown in the catalog is confirmed when an order is accepted. A paid order is not yet a ready resource, so plan your start date around confirmation. Pacific's step-by-step guide to renting H200 GPUs walks through that sequence.

• Confirm which regions are acceptable before you request quotes.
• Get support scope and replacement handling in writing.
• Know the point at which capacity and price become firm.
• Plan data export before the term ends, since local NVMe is erased.

6. Compare H200, H100 and B300 on cost per finished job

A lower rate per GPU-hour is only useful if the job finishes for less. The number that matters is cost per finished job: rate times GPU count times hours to completion, plus the time lost to setup, failed runs and data movement.

H100 has 80 GB per GPU. For models that fit easily in that memory, an H100 node can finish the job at a lower total, and our post on H100 price factors for reserved capacity covers that side of the comparison. For memory-heavy models or long context, H200 may need fewer nodes for the same work, which can flip the result.

B300 sits further up with 288 GB per GPU. In Pacific's catalog it is by request with pricing by request, durations of 1, 2, 3 or 6 months, and sizes from 1 to 8 nodes. That makes it a different buying process as well as a different chip. For the trade-offs between the two, see H200 vs B300: which to reserve.

• Run the same benchmark job on each candidate, if you can get test time.
• Record GPUs used, wall-clock time to finish, and setup overhead.
• Multiply by the quoted rate for the term band you'd actually buy.
• Compare totals per job, and per month of steady-state use.

7. Read an H200 quote line by line

Once you have a quote in hand, go through it slowly. Every line should answer a question you'd otherwise have to ask later.

• GPU model and form factor: H200 SXM, with memory per GPU stated.
• Node count and total GPUs: and whether that size was rentable directly or by request.
• Node specification: vCPU, system memory, local NVMe, interconnect.
• Tenancy: dedicated bare metal or shared.
• Term band and duration: start date, end date, and how the band was chosen.
• Rate basis: per GPU-hour, per node, or per term, and whether it's indicative or confirmed.
• Region: where the nodes sit.
• Access and software: SSH, drivers and CUDA status at handover.
• Data handling: what happens to local storage when the term ends.
• Support: scope, fault handling and replacement.
• Confirmation step: when capacity, quote and terms are revalidated.

If a line is missing, ask for it in writing. A quote that leaves out node specification or term details is hard to compare and easy to misread.

8. Questions to ask before you commit

The last step is a short conversation with the provider. These questions catch most of the surprises buyers run into after signing.

• Is this rate indicative or confirmed, and when does it become firm?
• Are all nodes on the same interconnect fabric?
• What happens if a node fails mid-term?
• Can I extend the term or add nodes, and is pricing for that agreed now?
• Can I move to a longer reserved band later if the workload proves out?
• What is included at handover, and what do I set up myself?
• How and when should I export data before local storage is erased?
• What does the review step check before my order is accepted?

If you'd like a second opinion on an H200 quote, or you need a size or term that sits in the by-request range, Pacific Intelligent Technologies, Inc. can walk through node shape, term band and region with you. Book a 30-minute capacity call to talk it through.

FAQ

Why is there no single H200 price?

Because the price depends on the unit, the node shape, the term band, the duration, the node count, the region and the support scope. Two quotes for the same GPU can differ because they bundle different things or cover different commitments. Normalize those factors first, then compare.

Is H200 always cheaper per job than H100?

No. H200 tends to win when its 141 GB of memory lets a job run on fewer GPUs or nodes, such as large models or long context. For smaller models that fit easily in 80 GB, H100 can finish the same job for less. Test with your real workload and compare cost per finished job.

Can I rent 16 H200 nodes directly from Pacific?

In Pacific's catalog, 1, 2, 4 and 8 node H200 sizes are rentable directly, and the 16-node size of 128 GPUs is by request. Requests go through a review that revalidates capacity, the current quote and terms before you commit.