The UK’s AI strategy is missing one crucial piece: compute

Britain is starting to treat AI as a matter of national security as well as economic growth. Last week at the UN General Assembly, Prime Minister Andy Burnham announced a new AI partnership between the UK and US, designed to protect critical national infrastructure from sea lanes to airspace.

It is an important development, particularly as AI moves from chatbots and productivity tools into defense and other systems where reliability really matters. But underneath it sits a rather basic question that receives far less attention: where will the computing power come from?

We have become accustomed to talking about AI in terms of models. Britain has strong universities, a healthy startup ecosystem and a growing collection of companies building products around the technology.

Yet their work all depends on the same thing: reliable access to compute. Models have to be trained somewhere, inference has to happen somewhere and the GPUs doing that work have to be available when companies need them.

Why smaller companies are squeezed out

The largest cloud providers are committing enormous amounts of money to AI infrastructure, and the biggest AI companies are securing GPU capacity years ahead. From a commercial perspective, this makes sense. A hyperscaler wants predictable demand, and a large AI company wants to know that the computing power it needs will be there when its next training run begins. The problem comes further down the market, where a growing number of companies are competing for whatever capacity is left.

And it is not only companies. An Oxford University study found that just 32 countries host specialized AI data centers at all, with US and Chinese firms operating more than 90% of them. American tech giants run 87 major AI hubs worldwide, while European companies run just six. Entire nations are effectively standing in the same queue as the startups, renting computing power on someone else’s terms.

For a British AI company that has raised a few million pounds and needs 500 GPUs for eight weeks, the economics look very different from those of a company capable of signing a multibillion-pound infrastructure deal. It does not need to own a data center. It does not necessarily want to make a five-year commitment. It needs access to compute for a particular project, at a particular moment, and it needs to know what that access will cost.

At the moment, that is exactly what it struggles to get. Rental prices for Nvidia‘s latest chips rose by nearly half in just two months earlier this year, and CoreWeave, one of the largest specialist providers, raised prices and stretched its minimum contracts from one year to three. Andreessen Horowitz has described one of its startups signing a contract for GPUs, only to be told at the last minute it would have to wait another three months because a bigger customer had offered three times the price.

The appetite is clear: when the government’s Sovereign AI Fund opened up compute for startups, it received over 200 applications requesting more than 67 million GPU hours in just three months. Yet much of the market is still poorly designed for that kind of customer.

The paradox of scarce and idle capacity

There is an odd paradox developing in AI infrastructure. We hear constantly about a shortage of GPUs, yet much of the capacity companies already pay for sits unused. One analysis of around 23,000 cloud clusters found average GPU utilization of 5%, meaning organizations are paying for roughly 20 times more capacity than their workloads actually use.

Worried about being caught short, companies buy for their busiest possible day, lock in capacity before they know exactly what they will use it for and hold on to it in case they need it later, so for much of the time, expensive hardware simply sits idle. The result is a market where capacity can be scarce for one company while sitting idle somewhere else.

So on one side there are companies searching desperately for compute, and on the other, capacity that is mostly going unused. Connecting the two seems an obvious step. Bringing more of the GPU capacity spread across data centers, infrastructure providers and enterprise environments into a common marketplace would give British AI companies another way to get the computing power they need, particularly when they are too small to negotiate directly with the largest providers or do not want to lock themselves into a single cloud.

It also goes to the heart of the sovereignty question Burnham keeps returning to. Plenty of computing power already sits on British soil, spread across university clusters, company server rooms, regional data centers and specialist providers, but in pieces too small to matter on their own.

Pooled together, that capacity would give British companies a source of compute under UK jurisdiction, rather than one dependent on the expansion plans of a few American cloud giants. Because it is spread across the country rather than concentrated in a handful of sites, it would carry the benefits of AI well beyond London and the South East, exactly the kind of regional growth Burnham built his reputation on.

What government and policy should do

Britain wants to create AI companies capable of competing globally. Those companies need researchers, capital and customers, but they also need somewhere to run their workloads. If access to compute becomes dependent on the size of a company’s balance sheet or its relationship with a handful of cloud providers, Britain risks creating an AI economy where the infrastructure is effectively reserved for the largest players.

There is a role for the government here. It has committed over £2 billion to public compute, including expanding the AI Research Resource twentyfold by 2030, and that investment is welcome. But much of it is going into a small number of national supercomputers, while demand from startups already far outstrips what is on offer.

That points to a simpler idea. Alongside building new supercomputers, the government should invest in making the most of the compute that already exists. New facilities take years to plan, connect to the grid and fill with chips, and each one adds capacity a single site at a time. The technology to link GPUs from many different locations into one usable network already exists, and it can put idle capacity to work without waiting for a single new building to go up.

Public funding could back that kind of pooling directly, for example by buying time on it for British startups. Government procurement could also give smaller AI companies a route into compute without requiring them to make the sort of infrastructure commitments expected of much larger businesses.

If companies can move workloads between providers without rebuilding their entire technology stack, they have more choice, and providers have to compete harder on price and service. That is good for the companies buying compute and ultimately good for the UK economy.

AI policy has spent a lot of time asking what Britain should build. The more immediate question may be whether British companies will be able to get enough compute to build it. That is the problem worth solving.

The best cloud storage: tested, reviewed and rated by experts

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit