The UK is losing the Red Queen's Race

Britain's AI infrastructure buildout is tangled in contradictions

"Well, in our country," said Alice, still panting a little, "you'd generally get to somewhere else – if you run very fast for a long time, as we've been doing."

"A slow sort of country!" said the Queen. "Now, here, you see, it takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast as that!"

Lewis Carroll, Through the Looking-Glass

Introduction

Back in January, Keir Starmer pledged to “mainline AI into the veins of this enterprising nation”, as he unveiled the AI Opportunities Action Plan, a report authored by entrepreneur and investor Matt Clifford, who was then serving as the Prime Minister’s AI advisor.

The snappily-named Opportunities Action Plan set out a series of sensible measures across infrastructure, sovereignty, procurement, and public sector adoption that likely tested the limits of what the Treasury was prepared to fund.

But since the announcement of the plan, the ambition of AI announcements from the UK Government has steadily slowed, while it picks up elsewhere in the world.

In the last few weeks, the US Government has published its own AI Action Plan, which among other things, contains 90 different actions spanning compute, workforce, procurement, and security. In March, the EU announced six additional sites for its high-performance computing clusters (dubbed ‘AI factories’), a month after the French had unveiled a €109 billion investment plan in French AI infrastructure. Meanwhile, the US-Gulf cooperation on compute only appears to be deepening.

At first, the UK appeared to be powering ahead on AI policy. In 2023, the country was first to establish a dedicated state-backed institute focused on evaluating the safety of frontier models. The AI Security Institute is a rare example of a new government institution that is adequately funded and able to attract top talent.

AISI has a clearly focused mission, so it can get on with its work with relatively little involvement from the rest of the government. If you’re evaluating the capabilities of a frontier model pre-deployment, the pool of stakeholders is pretty narrow (you, the people building the model) and your levers are pretty defined (asking them to make changes following evaluation).

Unfortunately, the UK has now plucked all of the low-hanging AI policy fruit. Every other important intervention requires navigating a world of departments with conflicting missions, regulators with limited interest or expertise, private actors, and legal roadblocks. This manifests itself acutely across areas like energy, infrastructure, defence, healthcare, and education.

Thanks to a combination of fragmented authority through endless outsourcing to quangos and various esoteric acts of financial engineering, the British state is particularly poorly placed to handle these. And this matters. Considering the growth in both capabilities and adoption, standing still and basking in the glory of things you set up in 2023 means falling behind.

To illustrate this, I’m going to focus on one element of the UK’s AI ambitions - its Compute Roadmap and ambition to establish ‘AI Growth Zones’. I’m not looking to debate the rights or wrongs of the specific proposals or targets here (although I’m sure AI infrastructure is a theme I will return to another time), more the extent to which the government has the ability to meet its own stated policy goals.

The devil isn’t in the detail

A significant portion of the AI Opportunities Action Plan was devoted to the foundational infrastructure for AI. The two most important proposals for our purposes are instructing the government to expand the capacity of the AI Research Resource (a set of publicly-funded supercomputers available to researchers) 20x by 2030 and to establish ‘AI Growth Zones’. These zones would match up private AI infrastructure providers with land and access to clean power through streamlined planning. So far, DSIT is yet to publicly name a private partner for a growth zone, six months after the call for expressions of interest first opened.

The Department for Science, Innovation, and Technology (DSIT), which published the Action Plan, followed up with the UK Compute Roadmap. This set out the specifics of the proposal, including a goal of six gigawatts of AI-capable data centres, including one site of more than a gigawatt. It also includes a headline pledge to invest up to £2 billion in sovereign compute by 2030.

These six gigawatts presumably have to be powered by something and connected to the grid. They also translate into fifty more terawatt-hours per year – the sort of scale that requires binding power contracts and grid reinforcements.

Instead, the plan only commits to ‘exploring’ new power options like small modular reactors or onsite renewable generation, while applicants only need a ‘pathway’ to 500 megawatts of capacity. The roadmap is silent on unglamorous enablers like planning reform or grid connections.

Why is DSIT so quiet on these points? In short, because it has little power to do anything about most of them. It can spend up to £2 billion on compute itself, a relatively trivial amount, especially as £750 million of this is being allocated to restoring a single project the same department cancelled last year. But the rest of the plan amounts to a suggestion note to the rest of the government.

To demonstrate why, let’s walk through an example.

The will to power

A potential solution to the government’s missing six gigawatts could lie in bringing more nuclear onto the grid. Nuclear is clean, reliable, and could be cheap in theory. The government said earlier this year that it wanted “new nuclear technologies such as small and advanced reactors for the first time, providing flexibility to co-locate them with energy intensive industrial sites such as AI data centres”.

In June, Rolls Royce’s small modular reactor design won the competition to be selected as Great British Energy – Nuclear’s preferred bidder to build the country’s first small reactors.

Considering the government’s rhetoric about “ripping up the rules”, an outside observer would likely interpret this as meaning that Great British Energy, the government’s newly minted and publicly owned energy company, is going to buy some nuclear power stations. The government has found the reactor design it likes and is committed to streamlining the rules, so Rolls-Royce should be able to get building sharpish.

In the UK, the government does not buy nuclear power stations. It partners with a developer and creates a special ring-fenced project company, which has to win over equity backers and debt lenders. These will typically be a blend of government, pension funds, utilities companies, and large institutional investors. Raising the money can take years, investors often get cold feet, and negotiations with government over funding models have resulted in entire projects being abandoned.

Investing in a UK nuclear project is not an immediately attractive proposition. With the exception of countries that have outright banned new nuclear power, the UK has the world’s most conservative nuclear regulator, alongside stringent (and frequently duplicative) environmental consents. Developers also have to be gracious and fund up to three legal challenges against their own projects, thanks to a convention the UK should probably leave.

To derisk this proposition, the government has cycled through different models. For Hinkley Point C, currently under construction in Somerset, the government embraced the Contract for Difference. Here, the developer receives a fixed price for every megawatt-hour for 35 years, guaranteed by a government-owned counterparty. If the market price is below the guaranteed strike, consumers make up the difference through levies on bills.

For Sizewell C and future large builds, the government switched to the regulated asset base model. Under this system, nuclear developers begin receiving revenue from consumers during construction, years before the plant produces electricity. Consumers, through their bills, effectively provide an advance cashflow to lower the developer’s cost of capital. If the project overruns or faces delays, regulators adjust the “allowed revenue”. The result is that construction risk, traditionally borne by investors, is transferred to the public, while investors enjoy a regulated, inflation-linked return once the plant operates.

The Regulated Asset Base. Hope that clears things up. Source: Lexology

No successful nuclear rollout in world history has ever been based on this bizarre mix of project companies, government-owned settlement bodies, revenue contracts, and capped liability schemes. All of this financial plumbing exists to keep nuclear off the public balance sheet and to help preserve the fiction that the government is sharing the risk with the private sector, even though it is ultimately the lender and insurer of last resort.

In short, the government is responsible, but not in control.

Winning the SMR competition doesn’t guarantee Rolls-Royce a single order. Instead, it will find it easier to access money awarded to Great British Energy. When it comes to any physical reactor builds, the whole private fundraising process and hunt for investors will need to be re-run project by project. The competition itself took ten years of behind-the-scenes dithering by the government before it got off the ground.

Financing aside, the Rolls-Royce reactor design still has to make it through the gauntlet of UK nuclear licensing. This sits outside government in an independent regulator and can take four to six years, versus the 12 to 18 months you might expect in a country like South Korea. At the moment, ministers have limited power to speed up approvals.

It’s not for nothing that the interim report from the government’s taskforce reviewing UK nuclear regulation found that “the system is perceived to be unnecessarily slow, inefficient, and costly” and indicated the need for a “radical reset”.

Even if a Rolls-Royce SMR project found financing, the prospect of it being approved, constructed, and plugged in within the next decade is remote.

Missed connections

Let’s imagine all of the above was fixable. Congratulations, you’ve got regulatory approval for your power source and financing lined up for your data centre. You’re almost on your way to meeting the government’s compute and clean power targets.

Or are you?

There’s another obstacle that neither No 10 nor DSIT can help you with. This one involves Britain’s electricity system.

At the top of the system, sits the high-voltage transmission network. This moves power around the country. In England and Wales, this is owned and operated by National Grid Electricity Transmission, in Scotland, by two other companies. These firms are private utilities. Ministers do not have a guaranteed right to instruct them prioritise one project over another, any more than they can tell BT which broadband line to lay first.

These companies operate as regulated monopolities. Ofgem (the UK’s independent energy regulator) sets rules through multi-year price control frameworks currently called RIIO-2 (short for “Revenue = Incentives, Innovation, Outputs”). Every five to eight years, the network companies submit detailed business plans to Ofgem setting out what assets they want to build or upgrade, and why. Ofgem reviews these plans, consults, and then decides what investment is justified, performance targets will apply, and return on equity the company is allowed to earn.

If a project isn’t an Ofgem-approved plan, the company cannot simply build it because ministers or customers want it. If it does, it has no automatic right to charge consumers for it.

Sitting below the high-voltage transmission network are distribution networks, which carry electricity at lower voltages into towns, businesses, and homes. These are controlled by a mixture of privately-owned distribution network operators (DNOs) that act as regional monopolies, and newer independent operators (iDNOS) that build local distribution networks for specific developments.

The layers of the electricity network Source: National Grid

For a moderate-sized data centre in the tens of megawatts region, the DNO is the critical gatekeeper. Even at Growth Zone scale, once bulk power has been brought from the high-voltage transmission system down into a form that can be used locally, a DNO or iDNO is still needed to build the ‘last mile’ of networks. This means that they can act as bottlenecks. Some DNOs have already had to tell developers in congested regions, like West London, that they cannot reinforce their networks quickly enough. iDNOs can sometimes ease this by building infrastructure of their own, but they still depend on connection points into the DNO or transmission system.

But when you are talking about 500 megawatts or more, the primary bottleneck is the transmission interface. Only the monopoly transmission owner can build the high-voltage substations and pylons needed to connect loads of this size. Developers must join a national connection queue run by the National Energy System Operator (NESO), a government-run body which coordinates with the transmission companies. The queue for connections is so bloated with speculative projects that it runs into the 2030s. Historically, it operated on a first-come, first-served basis. NESO is now implementing a reform programme to prioritise serious and nationally significant projects.

Even if a project moves forward, building the physical connection is slow. A new transmission-level substation and associated reinforcements typically take seven to ten years.

This is why some developers propose co-locating generation – for instance, a gas turbine or one day a small nuclear reactor – directly with their data centres. The idea is to produce power on site rather than wait for the grid. In practice, even these ‘self-sufficient’ sites still need a grid connection, for two reasons.

First, the industry standard for a data centre is 99.999% uptime. In short, it cannot afford to go dark because an on-site plant or storage system fails. It needs backup from the national system.

Second, surplus electricity from the private generator has to go somewhere when the data centre is not consuming at full tilt.

One idea under discussion is to allow Growth Zones to build private high-voltage substations and connection equipment to link their generation and demand directly to the transmission system, bypassing the need to wait for the monopoly transmission company to build one.

The one snag is that this is currently illegal. Private actors cannot directly connect to the transmission network. There appear to be no official moves afoot at the time of writing to change the law or create a carve-out under the Electricity Act to make this happen.

None of this is even accounting for the UK’s famously efficient planning system…

Fight the power

This planning point isn’t trivial, because our extra six gigawatts runs into another challenge - namely, the UK’s Net Zero ambitions. NESO, Ofgem, and the Department for Energy Security and Net Zero (DESNZ) are pursuing a policy called Clean Power 2030. The goal is that by 2030, 95% of Britain’s electricity generation should come from clean sources.

NESO forecasts electricity demand rising by about 11% in 2030, mostly from electric vehicles and heat pumps. If NESO’s assumptions held, DSIT’s six gigawatts of baseload power from the AI data centres would be responsible for 18% of all UK electricity consumption (more than twice NESO’s forecast).¹

In essence, the government’s clean power and AI goals are flatly incompatible.

To attempt to square this circle, officials seem likely to decide that AI Growth Zone developers will not be able to use gas cells as an interim power source. This is good for clean power, but fatal for AI Growth Zones.

If the legal obstacles on private generation described above had been eased, then gas might have at least made the idea of an AI Growth Zone theoretically feasible, while we wait for the SMR revolution to sweep the land. The alternative is powering an AI cluster with solar panels, but given the UK’s entire installed battery storage capacity could only power a one gigawatt scale cluster for a few hours, this seems unwise given our weather.

To bring this tension into sharp relief, we are currently witnessing a face-off between DSIT and DESNZ, with a two gigawatt data centre proposal seemingly colliding with a planned hydrogen plant planned for a nearby plot of land. The data centre has received planning permission already, but the BP, the developers of the hydrogen plant are pushing Ed Miliband, the Secretary of State at DESNZ to sign the development consent order for their plant. This would include compulsorily purchasing the land. It remains unclear how these two projects will coexist, and Number 10 seems content to sit on the sidelines and let the inevitable legal battle resolve it.

The centre cannot hold

The UK has a government department with ‘science’ and ‘technology’ in its name, but it’s not clear why it exists. The UK’s future AI adoption or innovation will largely not be determined by questions by pure technology policy. Instead, it will be influenced by areas like fiscal policy, defence, education, and energy.

A strong central body capable of compelling coordinated activity across all of these would be one thing. But this is not what DSIT does. DSIT can publish strategies suggesting that the departments or regulators that oversee these areas take certain actions, but it has no power to make them do anything.

Alongside imploring British internet users not to download VPNs, DSIT’s primary activities seem to be shuffling inconsequential pots of money around and organising roundtables.

While there are many talented individuals working there, the current institutional setup is not a serious vehicle for delivering serious change.

Even where DSIT does have authority, it uses it ineffectively. DSIT theoretically has meaningful power over scientific research in the UK, but has opted to outsource this work to the UK’s research councils. These bodies, which wield an annual budget of nearly £9 billion, have presided over the UK’s declining standing in international research rankings. Angela McLean, who heads the Government Office for Science, which sits inside DSIT, described this decline as “actually a fantastic thing”, as it was proof that other countries were improving.

A cultural indifference to the UK’s relative decline is a poor foundation for a scientific renaissance.

Perhaps the most frustrating point is that none of these challenges are insoluble. An energetic prime minister, who concluded that there were a series of connected challenges across departments that were collectively serving as a bottleneck on critical capabilities, could push through them with sufficiently energetic leadership. Ministers could challenge officials who tell them action is too difficult or risky. The law could be changed to make regulators work together. Britain is, in many ways, a very centralised country. If the executive is determined to run the country, it can just do things.

But the overwhelming sense I get in my interactions with politicians and civil servants is that the country’s governing class has given up. Westminster and Whitehall have been consumed by a nihilistic passivity.

If you’re working to change anything, the consistent feedback you receive is that ministers are only interested in ‘quick wins’ and that you should avoid anything that involves primary legislation. The solution to every problem is a new taskforce or appointment. These taskforces individually may have a great deal of merit (the nuclear taskforce’s early work has been stellar), but this is not a sustainable or scalable way of governing them.

A government, one year into a five year term and with a majority of 150, that takes this approach is bound to fail.

Disclaimer: These are my views and my views only. They are not the views of my employer, regulators with conflicting mandates, or anyone else. I’m not an expert in anything, I get a lot of things wrong, and change my mind. Don’t say you weren’t warned.

1. 6 GW×8,760 h/yr= 52.56 TWh. With NESO’s 2030 forecast standing at 287 TWh, this comes out at about 18%.