The asset Britain is about to give away
- aiomniconversation
- Apr 23
- 4 min read
The most valuable AI resource in the United Kingdom is not a chip, a model, or a data centre. It is a database that took 75 years to build: cradle-to-grave health records on every person who has ever used the NHS. No other country in the world holds anything like it. And right now, Britain does not have a domestic AI company capable of using it responsibly. That is not a technology problem. It is a strategic emergency.
In a recent conversation on The AI Adoption Podcast, I spoke with Tom Parker, an independent financial journalist who has reported on AI's economic and policy dimensions for the Financial Times, the Centre for European Policy Studies, and a range of legal and professional services organisations. Parker brings a financial reporter's instinct to a subject that too often gets discussed in the language of innovation rather than investment, risk, and consequence.
The numbers do not yet add up
Before we reached sovereignty, we talked about the economics. The figures Parker cited are worth dwelling on. Amazon, Google, Microsoft and Meta are each expected to spend over $100 billion on infrastructure this year alone. That is equivalent to 50% of each company's prior year revenues. Meta earned $200 billion last year and still required debt financing to fund its build-out. These are not growth investments in the conventional sense. They are bets placed in advance of contracts being realised, infrastructure built ahead of the demand it is designed to serve.
Parker drew a careful comparison with the dot-com era. In 1999, the top seven technology stocks carried an average price-to-earnings ratio in the mid-60s. Nvidia, two years ago, reached 60 times earnings at its peak, though it has since settled around the mid-20s, in line with the major hyperscalers. The valuation environment is not yet at dot-com extremes. But the infrastructure commitments are larger, the interdependencies more complex, and the compute requirements far beyond anything the dot-com build-out demanded.
Parker also raised a question about OpenAI's long-term position that I have not heard put quite so directly before. Without the deep balance sheets of the big hyperscalers, OpenAI may face the same fate as the early search engines. Lycos and Ask Jeeves arrived first. Google won. Being the company that popularised a technology is not the same as being the company that ultimately controls it.
The energy constraint is not theoretical
The infrastructure story is inseparable from the energy story. OpenAI has stated publicly that it needs 30 gigawatts of compute capacity to meet its near-term requirements, equivalent to a third of peak US grid demand, and has committed to deploying approximately one gigawatt per week. Sam Altman has projected a need for 250 gigawatts by 2033. For context, 250 gigawatts is the current total energy consumption of India. Meeting that demand would require 30 million GPUs and, by Parker's account, would produce twice the CO2 output of ExxonMobil.
One gigawatt is the output of a small to medium-sized nuclear power station. Building a nuclear plant in the UK or across most of the West takes between 10 and 15 years. The energy infrastructure required to power the AI ambitions being announced today simply cannot be built in time using conventional sources and conventional planning timelines.
The cost of that gap will not be absorbed by the technology companies. Parker was direct on this point: when energy demand rises faster than supply, prices increase, and it is consumers who bear the difference. The energy implications of AI adoption are not an environmental footnote. They are a cost-of-living question.
Sovereignty is the question leaders are not asking
Britain produced DeepMind. It was among the first companies to demonstrate AI's potential in a meaningful domain, beating the world's best Go player and pioneering applications in healthcare. Google bought it. The sovereignty of that technology passed to America.
Parker's account of the current landscape is sobering. The US operates as a single funding environment from coast to coast. Europe has 27 jurisdictions, each with its own capital markets, its own regulatory frameworks, and its own political considerations. Mistral in France is the most visible European attempt to build a sovereign large language model, but the structural disadvantages are significant. Access to capital remains the central constraint.
For the UK, one risk is the NHS. In 2023, a £330 million contract was signed with Palantir, the US defence-linked AI firm, to build a system capable of aggregating NHS data. There is now active discussion about reversing that contract. The difficulty is that no credible domestic alternative currently exists to fill the gap.
The consequence Parker described is precise. If a foreign platform holds NHS data and develops the analytical infrastructure around it, the NHS becomes a subscriber rather than an owner. A health service already under severe budgetary pressure would pay an ongoing fee to access insights derived from data it originally generated. That is not a hypothetical risk. The commercial logic of the arrangement points directly towards it.
For any business leader operating in a sector that holds significant personal, financial or health data, the NHS example is a live case study in what happens when sovereignty questions are deferred until after the contracts are signed.
The decision in front of every organisation
The AI adoption conversation has, for too long, focused on use cases, implementation timelines, and productivity gains. Those conversations matter. But they rest on assumptions about infrastructure, energy, sovereignty, and trust that are far less stable than they appear.
Parker ended our conversation with a personal reflection. As a journalist and podcast host, he can already see the shape of the disruption heading towards his own profession. A junior marketing executive, he said, can now produce a podcast script in under half an hour using AI tools. Voice cloning at 80 to 90% accuracy now requires just three seconds of audio, down from three minutes two years ago. He is not in denial about this. His view is that a premium will remain on the human element, but the market for it will shrink.
Every leader I speak to is somewhere on the same spectrum: between acknowledgement and action. The question the evidence in this episode raises is whether the time available to act is longer or shorter than most organisations currently assume. My own view, shaped by conversations like this one, is that it is shorter.
Listen to the full conversation with Tom Parker on Spotify: open.spotify.com/show/296zibtjU3w4ANuUsPSu2D and Apple Podcasts: podcasts.apple.com/gb/podcast/the-ai-adoption-podcast/id1811897501
Comments