H100 PCIe vs SXM pay-as-you-go rental for UK AI teams

H100 PCIe vs SXM pay-as-you-go rental for UK AI teams

7 min read

H100 PCIe and SXM GPUs can both be rented by the hour in the UK. We compare dated prices in pounds from Civo, Runpod, Lambda, OVHcloud and AWS, explain interconnect and storage, billing increments and UK data checks, and when a DGX B300 purchase starts to make sense.

CTC Staff
Written by CTC Staff House editorial team

Yes, both H100 PCIe and H100 SXM GPUs can be rented by the hour with no contract. On 17 November 2026 single-GPU list prices ran from about £1.51 to £3.71 per GPU hour, with London capacity from Civo and AWS. Pick PCIe when your model fits on one card, pick SXM when GPUs must talk to each other, and test storage and billing rules before you commit a production workload.

Can you rent an H100 PCIe or SXM on pay as you go?

You can rent both variants without buying an Nvidia DGX system. Civo's GPU price list shows a single H100 PCIe at $2.49 an hour and a single H100 SXM at $2.99 an hour on demand, both marked in stock. Civo's regions documentation lists GPU Compute and GPU Kubernetes as available in LON1 and unavailable in its other regions.

Runpod's pricing page lists the H100 PCIe at $2.89 an hour on Secure Cloud or from $1.99 on Community Cloud, and the H100 SXM at $4.39 on Secure Cloud or from $2.69 on Community Cloud. Runpod says these are billed per second with no commitment. Lambda's price list offers a single H100 PCIe at $3.79 and a single H100 SXM at $4.89 per GPU hour, plus applicable sales tax or VAT.

Two larger names sell H100 capacity in a different way. OVHcloud's UK price list shows its h100-380 instance, one H100 with 80GB, at £2.41 an hour excluding VAT. AWS sells H100 time in London through EC2 Capacity Blocks, which are reservations for a fixed window rather than open-ended pay as you go.

How much does an H100 cost per hour in pounds?

Most providers bill in dollars, so the pound figures below are planning numbers. I converted them at the European Central Bank reference rates for 16 November 2026, which give $1 as about £0.758. Card fees and VAT come on top unless the provider says otherwise.

OfferPublished priceApprox £ per GPU hourBilling model
Runpod H100 PCIe, Community Cloudfrom $1.99£1.51Per second, no commitment
Civo H100 PCIe$2.49£1.89Hourly, rounded up
Runpod H100 PCIe, Secure Cloud$2.89£2.19Per second, no commitment
Civo H100 SXM$2.99£2.27Hourly, rounded up
OVHcloud h100-380£2.41 ex VAT£2.41Hourly list price
Lambda H100 PCIe, 1 GPU$3.79£2.87One-minute increments
Runpod H100 SXM, Secure Cloud$4.39£3.33Per second, no commitment
Lambda H100 SXM, 1 GPU$4.89£3.71One-minute increments
AWS p5.4xlarge, Europe (London)$5.970£4.52Capacity Block reservation

H100 rental price per GPU hour in pounds

Published single-GPU H100 list prices checked 17 November 2026, dollar prices converted at the ECB rate of 16 November 2026

H100 rental price per GPU hour in poundsPublished single-GPU H100 list prices checked 17 November 2026, dollar prices converted at the ECB rate of 16 November 2026£0.00£1.00£2.00£3.00£4.00£5.00Runpod PCIe CommunityRunpod PCIe Com…Civo PCIeCivo PCIeRunpod PCIe SecureRunpod PCIe Sec…Civo SXMCivo SXMOVHcloud h100-380OVHcloud h100-3…Lambda PCIeLambda PCIeRunpod SXM SecureRunpod SXM Secu…Lambda SXMLambda SXMAWS London Capacity BlockAWS London Capa…Runpod PCIe Community, £ per GPU hour: £1.51Civo PCIe, £ per GPU hour: £1.89Runpod PCIe Secure, £ per GPU hour: £2.19Civo SXM, £ per GPU hour: £2.27OVHcloud h100-380, £ per GPU hour: £2.41Lambda PCIe, £ per GPU hour: £2.87Runpod SXM Secure, £ per GPU hour: £3.33Lambda SXM, £ per GPU hour: £3.71AWS London Capacity Block, £ per GPU hour: £4.52
View the data
H100 rental price per GPU hour in pounds
Category£ per GPU hour
Runpod PCIe Community£1.51
Civo PCIe£1.89
Runpod PCIe Secure£2.19
Civo SXM£2.27
OVHcloud h100-380£2.41
Lambda PCIe£2.87
Runpod SXM Secure£3.33
Lambda SXM£3.71
AWS London Capacity Block£4.52
Source: Runpod, Civo, OVHcloud, Lambda and AWS pricing pages, November 2026

The spread is not random. Each row bundles different CPU, memory, local disk and support terms, and the cheapest Community Cloud tier runs on different hosting from Runpod's Secure Cloud. AWS's Capacity Blocks pricing shows the eight-GPU p5.48xlarge in London at $47.757 an hour, or $5.970 per GPU, which is about £36.18 for the whole instance.

Commitment changes the sums again. Civo lists 36-month rates of $1.99 for PCIe and $2.49 for SXM, about £1.51 and £1.89, against its on-demand $2.49 and $2.99.

What is the real difference between H100 PCIe and SXM?

Both carry the same Hopper chip family, but they are built for different jobs. Nvidia's H100 specification lists the SXM part with 80GB of memory, 3.35TB/s of memory bandwidth, up to 700W of power and NVLink at 900GB/s between GPUs. The PCIe-based H100 NVL has 94GB, up to 400W and NVLink at 600GB/s through a bridge, while a plain PCIe Gen5 link runs at 128GB/s.

GPU-to-GPU bandwidth by H100 link type

Maximum interconnect bandwidth per GPU in GB/s from Nvidia's H100 product specification

GPU-to-GPU bandwidth by H100 link typeMaximum interconnect bandwidth per GPU in GB/s from Nvidia's H100 product specification02004006008001,000H100 SXM NVLinkH100 SXM NVLinkH100 NVL NVLink bridgeH100 NVL NVLink…PCIe Gen5PCIe Gen5H100 SXM NVLink, GB/s: 900H100 NVL NVLink bridge, GB/s: 600PCIe Gen5, GB/s: 128
View the data
GPU-to-GPU bandwidth by H100 link type
CategoryGB/s
H100 SXM NVLink900
H100 NVL NVLink bridge600
PCIe Gen5128
Source: Nvidia, H100 Tensor Core GPU specifications

In practice the gap matters most when one model is split across several GPUs. If your model and its context fit on one 80GB card, a PCIe instance does the same job for less money. If you shard a large model or run distributed training, GPU-to-GPU bandwidth decides how much of each hour goes on useful work. Ask the provider which links your instance really has, because "SXM" on a single-GPU instance gives you no peer to talk to.

Interconnect, storage or price per hour, which matters most?

Start with memory, then interconnect, then storage, and only then compare the hourly price. A cheaper GPU that waits on data or on its neighbours costs more per finished job than a dearer one that stays busy.

Storage specifications vary more than buyers expect. OVHcloud's h100-380 comes with 200GB of NVMe plus a 3.84TB NVMe passthrough drive and 8Gbit/s of network, while its four-GPU h100-1520 rises to 25Gbit/s. Lambda's eight-GPU HGX H100 instance includes 22TiB of SSD at $4.59 per GPU hour, about £3.48, which can suit teams that stage large datasets locally.

The fair comparison is cost per successful job. Add compute, storage you keep between runs, data you move and any retries, then divide by work completed. For inference, compare cost per token only at the same model, precision, context length and number of users.

How billing rules change what you pay

Billing increments differ more than headline prices. Civo's billing documentation says services are billed hourly, rounded up to the nearest full hour, and that a stopped instance is charged exactly like an active one because it still holds resources. Civo also states its prices are in US dollars and exclude VAT.

Lambda's billing guide says on-demand instances are billed in one-minute increments from the moment they pass health checks until you terminate them, busy or not. Its filesystems are billed for as long as they exist, even when not mounted. Runpod lists volume disk at $0.10 per GB per month while running and $0.20 while idle.

So "stop" does not mean "stop paying" everywhere. Build shutdown scripts around each provider's own rules, and delete or snapshot storage on purpose.

Keeping UK data in the UK when you rent GPUs

A London data centre helps, but it does not settle UK GDPR on its own. The ICO's guidance on restricted transfers asks you to check whether you can meet your aims without sending, or allowing access to, personal information to an organisation outside the UK. Remote support from abroad can count, not just where the server sits.

Civo says data on its GPU service stays within UK borders, and AWS lists London for H100 Capacity Blocks. For any provider, record the legal entity you contract with, its subprocessors, where support staff log in from and where backups go. The NCSC cloud security principles give a ready checklist for identity, administrative access, secure erasure and audit information.

When does buying a DGX B300 beat renting?

Buying makes sense once your workload is steady, your model choice is settled and you have the facilities. Nvidia's DGX B300 user guide lists eight B300 GPUs with 2.3TB of GPU memory in total, 10 rack units and 14.5kW of power consumption. That is a data centre purchase, and Nvidia sells it through quotes rather than a public price.

Renting the same generation first is possible. AWS lists the eight-GPU p6-b300.48xlarge in London at $129.168 an hour through Capacity Blocks, about £97.87, or $16.146 per GPU. Running your real workload on rented B300 or H100 capacity for a few weeks gives you the utilisation figures a purchase case needs.

The usual pattern for UK teams is to rent PCIe for experiments, move to SXM or multi-GPU instances when models outgrow one card, reserve capacity when demand is predictable, and only then ask for a DGX quote.

Sources

Frequently Asked Questions

Is H100 SXM available on pay as you go cloud services?

Yes. Civo, Runpod and Lambda all list single H100 SXM instances with no contract, at $2.99, from $2.69 and $4.89 per hour respectively on 17 November 2026. Lambda also rents eight-GPU HGX H100 instances at $4.59 per GPU hour.

Is H100 PCIe available on pay as you go?

Yes. Runpod lists it from $1.99 an hour on Community Cloud and $2.89 on Secure Cloud, Civo at $2.49 and Lambda at $3.79. PCIe suits inference and fine-tuning jobs that fit on one 80GB card.

How do I rent an H100 in the cloud?

Create an account, add a payment method, choose the GPU size and a CUDA image, then launch with SSH keys and a restricted firewall. Civo's documentation, for example, says to run nvidia-smi once the instance is active to confirm the GPU is visible. Test your real workload before moving data in, and check how the provider bills a stopped instance.

What are the alternatives to buying an Nvidia DGX system?

Rent single or multi-GPU H100 instances on pay as you go, reserve a fixed window through AWS Capacity Blocks, or take a longer commitment such as Civo's 36-month rates. Each lowers the risk of buying hardware before your usage is proven.

How can I price GPU compute by the second across GPU types and regions on one meter?

Record one usage metric, GPU-seconds, and attach GPU type and region to each event as pricing dimensions. Stripe's Metronome migration guide explains that compound group keys enable this dimensional pricing and cannot be changed after creation, so plan them first. Divide each hourly rate by 3,600 to get the per-second rate.