An H100 price is never just an H100 price. The same three characters sit on a PCIe card in a marketplace host, an SXM module in an eight-GPU node, and an NVL pair with 188 GB of memory between two boards. If you compare their hours as if they were one product, the cheapest cell wins the spreadsheet and loses the job.
This is the short version I use before I open a listing. Everything below is also on the H100 page, next to the live catalog rows. For the price spread itself, see the H100 price per hour section and the GPU pricing chart.
The three H100 form factors in one list
- H100 SXM: 80 GB HBM3, about 3.35 TB/s of memory bandwidth, NVLink at 900 GB/s, up to 700 W. The module you find in HGX nodes.
- H100 PCIe: 80 GB, about 2 TB/s of bandwidth, a slower NVLink bridge option at 600 GB/s, around 350 W. The card you find in standard servers and many marketplace hosts.
- H100 NVL: 94 GB per GPU, about 3.9 TB/s of bandwidth, two bridged boards per pair, 350 to 400 W. Built for large-model inference.
These figures are the ones the catalog uses on every H100 listing. They are datasheet numbers, not what your trainer will measure. They are still enough to tell you the three are different products.
H100 SXM: the training reference
SXM is the one with the full memory bandwidth and the full NVLink fabric. If you train across eight GPUs and AllReduce is on the critical path, this is the card the benchmarks were written for. It is also the one that needs a real node: 700 W per GPU is not a sad PSU in a spare chassis.
I rent SXM when NVLink is the point. Multi-GPU training, large tensor-parallel serving, anything where the interconnect is the bottleneck. I do not rent it for a single-GPU job that fits 80 GB and does not care about the fabric.
H100 PCIe: the cheaper hour with a catch
PCIe is the lower-power, lower-bandwidth card. It has the same 80 GB envelope, and it often has the lowest published H100 hour because marketplaces sell it. For fine-tunes that fit one GPU, batch inference and experiments, that is a real bargain.
The catch is the fabric. Multi-GPU jobs go over PCIe unless the host has an NVLink bridge, and the memory is slower. A cheap hour that doubles your step time is not cheap. Check interconnect on the listing, then do the math with the hour cost calculator.
H100 NVL: when memory per GPU is the point
NVL is the odd one. Each GPU carries 94 GB and the boards come in bridged pairs. It targets LLM inference where a 70B-class model and its KV cache would otherwise spill an 80 GB card. It is not what you need for a generic H100 job, and it is rarer in on-demand catalogs.
If a listing says NVL, confirm you are renting the pair and not half of one. If it says H100 and gives no form factor, assume nothing and ask.
How to read an H100 listing
- Form factor: SXM, PCIe or NVL. If it is missing, treat the row as unverified.
- Interconnect: NVLink, InfiniBand or plain PCIe. This is the training question.
- GPU count: an 8x node price per GPU is not a 1x PCIe price.
- Billing: on-demand, spot or waitlist. Do not mix them in one comparison.
- Stock: a waitlist from-price is later, not cheaper.
I read pricing pages in that order, then the dollar. It is the same order the methodology uses. If you are still choosing between Hopper and the next generation, read H100 vs H200 and H100 vs B200 after you settle the form factor.
FAQ: H100 SXM vs PCIe vs NVL
Is H100 SXM faster than PCIe?
Yes, in memory bandwidth and interconnect, and at the higher power it runs at. For a single-GPU job that is not memory-bound, the difference is smaller than the price gap suggests. Benchmark your own workload.
Does H100 PCIe support NVLink?
Through a bridge, at a lower bandwidth than SXM, and only on hosts that install it. Most marketplace PCIe listings do not give you the fabric. Read the interconnect column.
What is H100 NVL for?
Large-model inference. More memory per GPU and higher bandwidth than the 80 GB cards, delivered as a bridged pair. It is not the default H100 for training.
Which H100 should I rent for fine-tuning?
If the job fits one 80 GB GPU, PCIe is often the cheaper hour. If you train across GPUs and the fabric matters, SXM. If a 70B-class model and its cache are the problem, look at NVL or at H200.
Some outbound links in these articles may be affiliate links. Rankings on GPUBeacon stay independent.