For most UK AI teams, renting a small GPU allocation first is the safer route, because published H100 rental rates run from about £1.48 to £4.45 per GPU hour and you pay nothing while you test your model. Buying an Nvidia DGX B300 makes sense later, once utilisation is steady and you have the power and cooling for a machine that draws up to 14.5kW. The sections below answer the questions UK buyers keep asking about price, regions, data residency and the rent or buy decision.
How much does it cost to rent an H100 per hour?
Published rates vary because the products differ, not because the same product is priced at random. On 10 November 2026, Runpod lists the H100 PCIe at $2.89 an hour on Secure Cloud and from $1.99 on Community Cloud, and the H100 SXM at $3.99 on Secure Cloud and from $2.69 on Community Cloud, all billed per second with no commitment. Civo lists a single H100 SXM instance at $2.99 an hour, including 184GB of RAM and 23 vCPUs.
AWS sells H100 capacity differently. Its EC2 Capacity Blocks let you reserve a set window of time, and the p5.48xlarge instance with eight H100 GPUs costs $47.757 an hour, which AWS shows as $5.970 per GPU.
I converted dollar prices to pounds at the European Central Bank's 9 November 2026 reference rates, which give $1 as roughly £0.745. Treat the pound figures as planning numbers, since each provider bills in dollars and card or tax charges come on top.
H100 rental price per GPU hour in pounds
Published H100 rates converted from dollars at the ECB rate of 9 November 2026, checked 10 November 2026
View the data
| Category | £ per GPU hour |
|---|---|
| Runpod PCIe Community | £1.48 |
| Runpod PCIe Secure | £2.15 |
| Civo SXM | £2.23 |
| Runpod SXM Secure | £2.97 |
| AWS Capacity Block | £4.45 |
| Offer | Published price | Approximate per GPU hour |
|---|---|---|
| Runpod H100 PCIe, Community Cloud | From $1.99 | From £1.48 |
| Runpod H100 PCIe, Secure Cloud | $2.89 | £2.15 |
| Civo H100 SXM, single GPU | $2.99 | £2.23 |
| Runpod H100 SXM, Secure Cloud | $3.99 | £2.97 |
| AWS P5 Capacity Block, per GPU | $5.970 | £4.45 |
What matters most when renting, interconnect, storage or price?
Match the bottleneck to the job before you compare prices. A cheap hour is poor value if the job waits for data or restarts. Memory fit comes first, then the bottleneck, then price.
Single-GPU inference or development needs model fit and sensible throughput, so measure latency and cost on your own model. Multi-GPU fine-tuning depends on how the GPUs are linked, so ask which ones share an NVLink domain. Multi-node training depends on the network between servers, so ask for the topology and test it. Dataset-heavy work depends on sustained reads from the storage you will actually use.
Storage can also keep billing after compute stops. Runpod lists volume storage at $0.10 per GB a month while running and $0.20 while idle, so set a deletion policy for checkpoints before the pilot starts.
Do you need an H100 at all?
Not always. Runpod lists the L40S at $1.09 an hour (about £0.81) and the A100 80GB at $1.79 (about £1.33) on Secure Cloud. If your model fits in 48GB of memory, an L40S may serve inference at a fraction of the H100 rate. Benchmark the smaller card first, because the useful comparison is often L40S against A100 rather than H100 against B200.
Which regions are cheapest, and what about UK data residency?
No single region wins everywhere. AWS publishes the same $47.757 P5 Capacity Block rate in London, Stockholm and several US and Asia Pacific regions, so keeping data in London carried no price penalty on that tariff. That is one product at one provider, not a rule for the whole market.
Residency is more than the location of the GPU. The ICO's restricted transfer guidance says a restricted transfer can include allowing an organisation outside the UK remote access to personal information held in the UK. It also says that using a UK-based cloud provider is not a restricted transfer, while using one based outside the UK very likely is.
So write the schedule around data, not just compute. Datasets, checkpoints, logs, backups, support access and deletion all belong in it. A label such as "Europe" is not a precise enough location promise for personal data.
What are the alternatives to buying a DGX system?
There are four, and they differ in commitment more than in hardware.
- Small on-demand allocation. Runpod or Civo instances suit pilots, and you can stop at any time.
- Scheduled capacity. AWS Capacity Blocks suit a planned training window, with an upfront reservation fee. AWS says prices update regularly and are next scheduled to change in January 2027.
- Dedicated or reserved capacity. Providers quote this by term and quantity, so ask for the exact allocation, region and fabric in writing.
- Owning a system. This gives local control but also capital cost, support and refresh risk.
When does buying an Nvidia DGX B300 make sense?
Nvidia's DGX B300 user guide lists eight B300 GPUs with 2.3TB of total GPU memory, a maximum power draw of 14.5kW and an inlet air range of 10 to 35 degrees Celsius. The system also carries eight 3.84TB NVMe cache drives. This is a facilities project, not a desktop purchase, and you need power, cooling and networking ready before it arrives.
I found no public Nvidia list price, so ask for a written hardware and support quote. Then compare annualised capital cost, support, power, staff and financing with your full rental bill, divided by useful work delivered. The sensible sequence is to define the data boundary, pilot a small rental, measure real utilisation, and only then negotiate reserved capacity or a purchase.
What changes next for GPU rental pricing?
Prices are not fixed. AWS reviews Capacity Block rates regularly and gives January 2027 as its next scheduled update, so a quote you take now may not hold for a later reservation. Re-check the provider's own pricing page on the day you sign, and ask whether the rate is on-demand, reserved or interruptible before you compare it with another quote.
Sources
- Runpod, GPU Cloud Pricing, checked 10 November 2026
- Civo, Prices, checked 10 November 2026
- Amazon Web Services, Amazon EC2 Capacity Blocks for ML Pricing, checked 10 November 2026
- Nvidia, Introduction to NVIDIA DGX B300 System, user guide
- Information Commissioner's Office, Quick reference FAQs on restricted transfers
- European Central Bank, Euro foreign exchange reference rates, 9 November 2026
