Cottaige
Core

Why we should lean into our roots and share the tools, to build our skills from the bottom up.

The Problem

I run a small AI lab from my living room: a Nvidia 4090, two 3090s and a stack of open models fine-tuned to my own work. The whole thing cost less than a used hatchback. The realms of enthusiast, but not the realm of data centres or professionals. This article is about why Britain’s AI future looks more like my living room than a datacentre campus, and more like enabling grass-roots enthusiasts than enabling foreign scalers.

First: the scale of the competition. Across this parliament the government has committed roughly £3.5 billion of public money to AI, spread across overlapping announcements: about £2 billion at the Spending Review, a £1.1 billion sovereign hardware plan, a £500 million Sovereign AI Unit, with the same £750 million supercomputer appearing in more than one of them. The four big American hyperscalers plan around $650 billion of capital spending in 2026 alone. At that rate they spend Britain’s entire multi-year commitment roughly every three days.

The current strategy answers this by courting the capital. Five AI Growth Zones so far - Oxfordshire, the North East, both ends of Wales, Scotland - built on faster grid connections and planning reform, which the government says could unlock up to £100 billion of datacentre investment. There is nothing wrong with wanting the warehouses full of servers. But it is a strategy for hosting other people’s AI, and every country in the world is in the same race.

The Shed

Before the factories, English cloth was made by the putting-out system. A merchant delivered wool to the cottages; spinners and weavers worked it at home, on their own wheels and looms; the merchant collected the cloth and paid by the piece. The skills, the tools, and the margins lived in thousands of small rooms rather than one big one.

The frontier labs have accidentally rebuilt that system. An open-weight model is delivered wool: the capital-intensive step has been done for you, released to anyone with the hardware to use it. What a cottage adds is the refinement - fine-tuning to a trade, wiring into a workflow, judgement about when the output is wrong. The finished cloth is a service no hyperscaler can make, because the hyperscaler has never met your customers.

And Britain is unusually rich in cottages. We have 5.7 million small and medium businesses, 5.4 million of them with fewer than ten people. They employ 16.9 million people - three-fifths of the private sector - and turn over £2.8 trillion a year. Only 8,335 British businesses are large. The cottage industry is most of the economy.

The tinkering culture is real too. Over a thousand Men’s Sheds now operate across the UK. The Raspberry Pi - a hobbyist board designed in Cambridge to get children programming - has sold some 68 million units, become the best-selling British computer ever made, and floated on the London Stock Exchange. We are good at this. It might be the single most British form of economic activity: small, specialised, slightly obsessive, done in a shed.

The Idea

The academic backing here is Jeffrey Ding’s Technology and the Rise of Great Powers. Ding studied how general-purpose technologies - steam, electricity, computing - decided which nations rose. He found that the economic winner was the country that spread the technology widest through its ordinary firms, whichever country had invented it.

Diffusion is exactly where Britain is currently failing. DSIT’s own adoption research puts AI use at 36% of large firms and 14% of micro businesses. Yet the same survey found that once a micro business does adopt, 38% of its staff use the tools, against 20% in large firms. Small firms lag on entry and lead on intensity. Skills England estimates the AI skills gap could hold back as much as £400 billion of growth by 2030.

So the proposal is simple to state. Give up on out-building the hyperscalers and set a different national target: the highest AI adoption rate among small firms in the world. Play into our comparative advantage.

The Politics

Just before this article, a new Prime Minister stood on the steps of Downing Street and promised to “take power out of here and carry it into every postcode in the land so that they can do more”. Andy Burnham’s diagnosis was that the 1980s centralised political power and de-industrialised the regions, and that neither has recovered.

His opening speech never mentioned AI. But a diffusion strategy is the manifestation of his decentralisation promise. Devolved strategic authorities now have a common legal footing under the devolution act, and they are well placed to run it - Greater Manchester knows its own machine shops, letting agents and food producers in a way Whitehall never will.

The machinery is mostly already built. Full expensing and the Annual Investment Allowance already let a business write off equipment; point them explicitly at AI hardware and tooling for micro firms.

The FSB finds 46% of small firms name lack of knowledge or skills as a barrier, so pair the capital with training vouchers rather than another portal (and think about the knowledge implications of widespread world-model adoption). And the new government says it will use public procurement to back British industry; contracts that ask for AI-capable small suppliers would do more for adoption than any awareness campaign.

The National Model

The bottom-up path now runs surprisingly far up the stack, and the precedent is older than the labs. In 2020, volunteers pointed their home computers at COVID protein-folding and Folding@home (a long running tool) briefly assembled more raw compute than any supercomputer on Earth. The distribution layer already exists too: Hugging Face hosts almost three million open models, downloadable by anyone with a connection.

In December 2024, Nous Research pre-trained a 15 billion parameter model over the open internet, on scattered machines coordinated by their DisTrO optimiser. Their Psyche network went live in May 2025. By December 2025 they shipped Hermes 4.3, a 36 billion parameter production model post-trained entirely on Psyche. Prime Intellect, working the same seam, closed a $130 million round this month at a $1 billion valuation.

Nobody has crowd-trained a frontier-scale model yet. The record is Nous’s Consilience - a 40 billion parameter model pretrained from scratch over the internet, the largest such run to date - plus post-training of larger open bases. But that class is already a serious, useful model. A sovereign open model as a national volunteer project - university clusters, business workstations, a hundred thousand idle gaming PCs - is no longer science fiction. Mine would join. It spends many nights idle anyway, which feels like a waste of a perfectly good shed.

Limits

Adoption surveys disagree with each other - ONS says 25% of businesses use AI, DSIT says 16%. Productivity gains are mostly self-reported so far; only 12% of adopting firms report increased revenue, and activity is not output. And devolving political power does nothing for AI diffusion by itself. The bridge - allowances, vouchers, procurement - has to be built deliberately, and no current scheme is aimed at micro-firm AI capital.

There is also a compute dependency. The strategy leans on small open models staying good enough for most small-firm work, while frontier compute stays scarce and expensive - the big clusters are already rationing access at today’s adoption rates. My own testing puts a local stack behind the frontier on hard reasoning, but most small-firm work is routine and semi-programmatic. If mass adoption turned out to need frontier-class models for everything, the cottage route would collapse into buying hyperscaler capacity like everyone else.

What Next

The Prime Minister has promised a 10-year plan for Britain later this year. The thing to watch is whether a single line of it points capital at the five million smallest firms, or whether it is all aimed at the scalers. As for me? I’m going to see if we can get a real workstation.