Every serious AI ambition eventually meets the same question: where will the compute come from? Graphics processing units, or GPUs, are the engines behind modern AI, and they are expensive, power-hungry and quick to date.
For most organisations, the answer is not to buy them. It is to rent them wisely.
The market has shifted in the buyer’s favour
Two years ago, GPUs were scarce and rental prices were at their peak. That has changed. According to industry analysis, the hourly cost of renting an NVIDIA H100 fell from around $8 at its late-2024 peak to roughly $3 by the end of 2025, as new capacity came online and hundreds of new providers entered the market.
Capacity is also arriving closer to home. Stargate UK, a partnership between OpenAI, NVIDIA and Nscale, is building sovereign AI infrastructure in the UK, aimed at uses where jurisdiction matters, such as financial services and public services.
Why renting usually wins
- No capital tied up. GPUs can cost tens of thousands of pounds each, before power, cooling and specialist staff. Renting turns that into an operating cost you can scale up or down.
- No stranded assets. New GPU generations arrive every year or two. Renting means you are not left holding yesterday’s hardware.
- Speed. Capacity can be available in days, not the months it takes to buy, install and commission hardware.
- Matched to real demand. Most AI workloads are bursty. Training, testing and seasonal peaks rarely justify hardware that sits idle the rest of the time.
When to rent, reserve or own
A useful rule of thumb from the same analysis is to look at how busy the GPUs would really be.
- Variable or experimental work: rent on demand. Pilots, proofs of concept and occasional workloads belong on flexible, pay-as-you-go capacity.
- Steady, predictable work: reserve. Committing to capacity for a fixed term lowers the price while keeping you off the balance sheet.
- Constant, heavy use: consider owning. Ownership only starts to pay back when GPUs are busy most of the time, and when you can run them well.
- Most organisations: blend. A reserved baseline with on-demand capacity for peaks gives control of cost without limiting ambition.
Five questions to ask before you rent
- Where will our data be processed? For regulated firms, the location and jurisdiction of the GPUs matter as much as their price.
- Who can access the environment? Treat rented compute as part of your estate, with the same security, identity and audit controls.
- How will we control spend? GPU hours add up quickly. Budgets, approvals and monitoring need to be in place from day one.
- What is the exit plan? Avoid lock-in by keeping models, data and workflows portable between providers.
- How does it connect to everything else? AI compute is only useful when it is joined up with your data, networks, cloud platforms and security.
Rent GPUs through Citycom
Through our partnerships with Microsoft and AWS, Citycom customers can rent GPU capacity on Microsoft Azure and Amazon Web Services, on demand or reserved, without buying a single piece of hardware.
We help you choose the right capacity and region for each workload, set it up securely, and manage it alongside your Cloud, Infrastructure, Security and Communications, on one contract with one support model. GPU capacity never becomes another silo, with its own supplier and its own bill.
Ark Intelligence brings every layer into one real-time view, so you can see what AI is costing and what it is delivering.
Tap into the expertise of multiple vendors. Deal with one.
If AI compute is on your board’s agenda, let’s talk. Start a conversation at citycom.co.uk or call 0330 912 1212.
