What if Europe is copying the wrong AI strategy?

That may be true for the frontier labs. If you are OpenAI, Anthropic, Google or Meta, perhaps the only game in town is to keep scaling. More GPUs, more data, more power, more infrastructure. The logic is clear enough: if intelligence becomes the next platform layer, the companies that own the compute get to own the market.

The problem is that UK and European policymakers seem to have absorbed the same logic. They look at what the big American AI companies are doing and conclude that, to stay competitive, we need our own version of the same thing. More data centres. More sovereign compute. More frontier models. More public money trying to ensure we are not left permanently dependent on American infrastructure.

Some of that is understandable. Europe does need more technical capability. The UK cannot afford to be naïve about AI sovereignty. There are obvious risks in outsourcing too much of our intelligence layer to a small number of US companies, especially when the US government has already shown it is willing to restrict access to advanced chips and AI technologies when it suits its geopolitical interests. What looks like a sensible commercial dependency today can become a point of leverage tomorrow.

But I worry we are making a fairly basic strategic mistake.

We are not fast followers in this race. We are slow followers. And worse, we are slow followers with very different conditions.

The US has vast amounts of land, cheaper energy in many regions, deeper capital markets, the major cloud platforms, the frontier labs, the chip relationships and a handful of enormous technology companies willing to spend staggering amounts of money. The UK and Europe have a very different hand to play. We have more expensive energy, tighter planning constraints, less available land, shallower pools of growth capital and far fewer companies with the balance sheets of Microsoft, Amazon, Google or Meta.

Trying to beat America at America’s version of the AI race feels like a poor place to start.

So perhaps we should stop asking how we catch up, and start asking where we should deliberately take a different path.

That is why I’m becoming much more interested in open-weight models and edge computing. Not because they are a consolation prize, or a cheaper imitation of frontier AI, but because they point towards a strategy the UK and Europe could actually be good at: smaller models, running closer to the user, using less energy, protecting more private data, and doing more with less.

Doing more with less is not some romantic European virtue. It is a constraint we have had to get good at. We do not have the deep pockets of the US hyperscalers. We do not have endless land for server farms. We do not have abundant cheap energy. We do, however, have a long tradition of caring about privacy, institutions, regulation, standards, interoperability and the public consequences of technology.

Instead of treating those things as weaknesses in the AI race, perhaps we should build around them.

Most people are not trying to solve theoretical physics over breakfast. They are trying to summarise a document, search their notes, draft an email, organise a project, query a company knowledge base, write a bit of code, understand a contract, generate options, or turn a messy meeting transcript into something usable.

For that kind of work, “best model in the world” may matter less than “good enough, cheap enough, private enough and close enough to the user.”

Yet most of the current infrastructure debate assumes intelligence will keep being accessed remotely, in the same way we access storage, software, music, maps and almost everything else. You ask a question; the request disappears into somebody else’s infrastructure; the answer comes back a few seconds later.

There are good reasons for this. Training large models is expensive. The best systems still require serious compute. Most companies do not want to manage infrastructure. Consumers certainly do not. The cloud is convenient, and convenience usually wins.

But the current AI build-out also relies on a very particular assumption: that the future of intelligence will remain centralised enough to justify the cost.

That is not a given.

The railway analogy gets used a lot in discussions about AI investment. Yes, the argument goes, there may be a bubble. Yes, some investors may lose money. But the infrastructure will remain useful. The railway companies of the nineteenth century may have overbuilt, but the tracks still changed the world.

It’s a comforting analogy. I’m just not sure it works.

A railway line does not become dramatically less useful because somebody invents a better railway line three years later. The route, land, stations and physical network can hold value for decades. GPUs are different. They are extraordinary pieces of engineering, but they sit inside a brutally fast replacement cycle. Today’s frontier chip becomes tomorrow’s inference chip, then the day after tomorrow’s second-tier asset.

The land may hold value. The power connections may hold value. The cooling and networking may hold value. But the most expensive part of the build-out is not permanent in the way railway tracks are permanent. It is more like buying a warehouse full of Formula One cars and calling it transport infrastructure.

Maybe demand for AI compute will be so large that every generation of GPU finds profitable work. But we should be careful about treating compute as if it were a timeless public good. A data centre full of ageing chips is not the same thing as a bridge, a port or a railway line.

This becomes even more questionable if demand starts moving away from the centre.

Open-weight models have been improving quickly. They do not need to beat OpenAI, Anthropic or Google’s frontier models to change the economics of the market. They only need to become good enough for the fat middle of everyday work.

That is usually how technology markets get interesting. The technically superior product does not always capture the most value. Often the good-enough product wins because it is cheaper, more adaptable, easier to own, easier to modify, or simply closer to the context in which it is used.

Personal computers did not beat mainframes because they were more powerful. MP3s did not beat CDs because they sounded better. Digital cameras did not initially beat film because they produced richer images. They won because they changed the pattern of use.

The same could happen with AI.

If a law firm can run a capable model locally across its own documents, why send sensitive client material to a remote provider for every query? If a hospital can keep patient data inside its own infrastructure, why make external inference the default? If a bank wants auditability, control and predictable costs, why rely entirely on a black-box model accessed through an API? If a design team wants an assistant that understands its research archive, product decisions, brand guidelines and customer history, there is an obvious appeal in keeping that intelligence close to the work.

This is not just about cost, though cost matters. It is about privacy, latency, control, resilience and trust. It is also about fit. The more personal or organisational context an AI system needs, the stranger it feels to keep outsourcing that intelligence to a distant server.

That is where edge computing becomes interesting to me.

I’m not imagining a grand return to beige boxes and home networking projects. Most people do not want to manage infrastructure, and they certainly do not want to become part-time sysadmins just to use AI. If local intelligence is going to work, it has to disappear into the devices and environments people already use.

Your phone might run small models for quick, personal tasks. Your laptop might run larger ones for writing, analysis and creative work. Your company might run its own internal model across products, customers, policies and institutional memory. And yes, some homes may eventually have something closer to a personal intelligence server, but only if it feels more like a router or a smart speaker than a hobbyist project: quietly there, mostly invisible, useful because it coordinates the AI running across your devices.

You can imagine Apple doing something like this surprisingly well. Not another loud black box for enthusiasts, but something more like a Mac mini designed for local intelligence: a small, quiet device with a decent GPU, a private local model, and deep access to your own documents, messages, photos, calendar and notes. You would get something that feels closer to a really capable ChatGPT-style assistant, but with the context stored in your home rather than sprayed across a cloud service. For European users, that kind of model could be especially attractive. It gives you convenience without quite the same privacy trade-off. It feels like a very Apple-shaped answer to the AI problem.

That may sound slightly odd now, but so did having a media server once. The difference is that AI is more intimate than media. It wants context. It wants memory. It wants access to the messy private material that makes it useful. For some tasks, that makes the edge a more natural place for intelligence to live.

There will still be cloud models, of course. You might call out to a frontier model when the task is genuinely hard, when you need the best available reasoning, or when local models are not enough. But that does not mean every mundane act of AI-assisted work needs to be served from a hyperscale data centre.

A better pattern might be hybrid: local intelligence for the everyday layer, frontier intelligence for the exceptional layer.

This is where the policy conversation starts to feel strangely unimaginative.

If America is building centralised intelligence at massive scale, Europe could focus on distributed intelligence that is private, efficient, inspectable and close to the user.

That would mean taking open-weight models seriously as a strategic asset, not merely as a cheaper substitute for proprietary systems. It would mean investing in small and medium-sized models optimised for European languages, industries, public services and regulatory contexts. It would mean supporting inference at the edge: on devices, in offices, in hospitals, in factories, in local authorities, in schools and inside companies that cannot or should not send their most sensitive data to a remote model provider.

The prize is not national bragging rights on a benchmark table. The prize is useful AI capacity embedded throughout the economy.

There is a historical irony here. Britain had an enormous first-mover advantage in the steam age and the Industrial Revolution. We built early, scaled early, and for a time owned markets while everyone else was catching up. But early infrastructure can become a burden as well as an advantage. You inherit the old railways, the old factories, the old housing stock, the old energy systems. Newer countries can sometimes move faster precisely because they have less to rip out first.

AI infrastructure may follow a similar pattern, only on a much shorter timeline.

If the largest American companies want to spend hundreds of billions inventing the future, perhaps we should let them. Some of that work will be useful. Some of it will push the frontier forward. Some of it may create infrastructure the rest of the world can learn from. But it does not follow that the UK and Europe should rush to pour scarce capital into the same kind of infrastructure just as the underlying technology is changing so quickly.

The risk is not just that we arrive late. It is that we arrive late, spend too much, and inherit an ageing version of someone else’s architecture.

A more edge-focused strategy gives us more room to move. It does not require us to bet the farm on massive centralised compute before the economics are clear. It lets us benefit from open-weight models as they improve. It keeps more intelligence close to the data. It gives companies and public institutions a way to adopt AI without sending every sensitive query across the Atlantic. And if large parts of the current compute boom do turn out to be overbuilt, we will not have sunk quite so much national effort into infrastructure that is hard to unwind.

That feels like the more interesting bet.

Not a slightly weaker copy of the American compute race, but a more distributed, efficient and trusted model of AI adoption. One that treats privacy as a design constraint rather than a regulatory burden. One that makes open-weight models part of public and commercial infrastructure. One that helps useful intelligence live closer to the people, companies and institutions that need it.

The future of AI may still be built in giant data centres. But a surprising amount of it may end up much closer to home.

If that happens, the UK and Europe should not treat local, efficient, privacy-preserving AI as a second-best outcome. It may be the best strategy we have.