AI Shrinks Moats

Britain sits on one of - possible the - greatest pile of data in the world. It's not as financially valuable as the vast patchwork of data that makes ad-tech work, and it's certainly not the easiest to use, but the clinical records of the National Health Service (NHS) have enormous potential to improve the human condition. Britain is unique among populous and ethnically diverse countries in having data like this.

I know some legacy (right-wing) media like to trash the NHS, and I know politicians like to sound off about its "lacking" data infrastructure when they want to look tech-savvy. But they know nothing. I've spent most of my career working with NHS data, and as I've progressed that's included many conversations with international peers. Take this to the bank - no health service holds the entire medical histories of almost 70 million patients on data infrastructure as advanced as ours.* None.

And the process of creating that data started a long time ago. Back in the nineties, most clinical information was on paper, either printed or handwritten on cards and stuffed into open-top envelopes called Lloyd George's. Everything about a patient was in these envelopes - young and healthy patients only had one, older or sicker patients would have several bound together with sellotape - and great racks of them took up space in doctor's offices.

Digitizing all of that information was a massive, nationwide data inputting exercise. Admin staff did it whenever they had a spare moment, and new staff were brought in, and teenagers were herded over the summer as low-wage labour - your author was one of them. It took years, but the digital data it created was a major factor in the NHS reigning as the world's best healthcare system for a long time (a position now lost - badly! - following massive under-funding post-Brexit).

You'd be amazed how many great advances in data and technology owe most of their existence to this kind of mass sweat equity. The Y2K bug wasn't averted by magic, for example, but through the Herculean efforts of the global software community. And every piece of web-scrapping, "content-borrowing" AI you've heard of is derived from decades of AI you haven't heard of, and that AI required massive amounts of manually labelled data - created, of course, by the world's largest collective of indentured workers: post-graduate students.

Over the years, I've met several tech companies who built their initial success on some kind of manual data asset. Normally, this is a library of items or mappings that a founder dedicated a year or more of evenings and weekends to creating. It's a particularly good moat because when an investor asks what you have that a competitor doesn't, you can point to that sweat equity and say, "everyone is at least that far behind."

Except the moats are shrinking. Last summer, I started offering clients advice on how they could incorporate GenAI into their software and workflows. (I'll write an article on this soon, but the key is to change the way you think about the "generative" part - almost everything you see and read about online is hype-crap). Today, I offer the same advisory service, only the advice is changing as fast as the sector is, and AI can now provide large amounts of the sweat equity that tech relies on.

This is a boon in an economic vacuum because more sweat means more innovation. But in a competitive economy, the equity it earns stops being worth much because "everyone is at least that far behind" is pretty meaningless when "that far" is an AI prompt.

Ironically, investors in GenAI have learned this the hard way. Recently, a Chinese company released a new AI model called DeepSeek-R1 which performs as well as the best American models. Only DeepSeek had a fraction of their production costs and was released free of charge under an open-source licence - meaning anyone can download it and do, largely, whatever they like with it. About a trillion American dollars were wiped out.

Then, in a stunning lack of self-awareness, OpenAI's CEO - himself facing huge criticism for "borrowing" other peoples' work to train ChatGPT - claimed that DeepSeek may have "stolen" his model's capabilities through a process known as distillation. If this is true (and I'm not aware of any evidence that it is), then it shows the best American models can be cheaply copied by rivals. If it's false, then it shows that America's models can be cheaply replicated. Either way, the AI moat has suddenly become a small puddle.

This opens up a tantalizing possibility. Many smaller advanced economies have struggled to assert themselves on the AI stage. This is partly because there was little incentive for governments to invest in a technology that requires America-sized budgets. Now that DeepSeek has slashed the cost, maybe countries like Britain can do meaningful things with AI, like leverage our NHS data for a common global good?

*Technically there are four NHS organizations - England's, Wales's, Scotland's and Northern Ireland's - but my point stands.