The interesting artificial intelligence, for anyone who has to put skilled jobs and invested capital on a return, is the kind that does not get a keynote.
Washington made the point without meaning to. In the last days of September the Pentagon’s research chief cut the department’s critical-technology list from fourteen areas to six, and put the word applied in front of artificial intelligence. Beside it sat biomanufacturing, contested logistics, quantum and battlefield information dominance, directed energy and hypersonics. A long list used to be a brochure. A short list is a buying instruction: fund work that changes a decision, a process or a production line.
Britain supplied the other half of the picture in the same fortnight. A survey of metalforming and manufacturing firms, published on 25 September, found four in five already using or trialling AI, up from about half a year earlier, and nearly three in four engaged in or exploring defence work. The same firms named skills and cost as the constraints, and said growth is no longer funded from cash as it was in the spring. That day the Ministry of Defence opened a competition for up to twelve British companies to train models on Ukrainian battlefield imagery — some six million detections — on the condition the resulting systems be made in Britain.
None of this is a technology story. It is a location story that happens to involve software.
Economic development organisations have treated artificial intelligence as a sector that arrives, like a car plant, if the incentives are loud enough. A strategy. A campus. A gigafactory bid. Some of that compute will be built. Power, cooling and a sovereign-cloud argument will decide those sites. They will not decide where the skilled jobs in the industries a region already owns are kept or lost.
Applied AI is dull in the way that payroll is dull. It is inspection on a composites line that used to need a scarce technician. It is scheduling and yield in a bioprocess plant. It is design iteration on a space structure or a munitions casing that shortens the gap between a drawing and a qualified part. It is a model that keeps a small uncrewed system working when navigation is denied — and, more importantly for a region, the workshop that builds the airframe and the sensor around it. The model can be trained anywhere with a clearance and a dataset. The workshop locates where the adjacent industry and the people already are.
The mistake is to fund the model and call it a cluster. A place that announces an AI strategy and does not know which twenty firms could sell an inspected, qualified, dual-use part into the programmes now being written will get a press release and a vacancy for an AI lead. A place that already machines, forms, tests or formulates, and that can put those firms in front of buyers willing to pay for speed and domestic capacity, gets the conversion. The American defence authorisation signed at the start of October — some $901 billion, with fresh money for the industrial base, munitions and a biotechnology office inside the department — is a statement that production, including bioindustrial production, is the scarce thing. Europe’s gigafactory tenders will not staff a mid-tier supplier.
Two constraints decide whether the trial becomes a site. The first is skill. British manufacturers in that September survey still cited a shortage of qualified applicants as their main recruitment problem, even as AI use jumped. Automation does not dissolve that. It changes the job: fewer people watching a gauge, more people who can qualify a process, hold a clearance and talk to a prime. Regions that subsidise pilot projects and leave the technician pipeline alone will watch the work sit with the firm that already has the people. The second is cash and qualification. Firms that have stopped funding growth from reserves will not self-finance a defence line, a clean room or the paperwork that turns a clever inspection tool into a part a buyer will accept. An agency that can only offer a demonstration has nothing to sell them. An agency that can assemble a buyer, a site, a skills route and a route to capital has the only product that matters.
There is a quieter point about sovereignty. Applied AI inside a contested supply chain is an asset only if the data, the model and the factory sit in jurisdictions a buyer can live with. American policy this autumn treats AI, biotechnology, quantum and space as both opportunity and exposure: allied capital welcomed, adversary capital scrutinised. A region that hosts the workshop but not the trust will not host the contract. The question will be asked of life-science process plants as readily as of drone shops.
Stop buying AI as a destination. Buy it as a capability inside the sectors the place can already defend in a meeting with an investor: precision manufacturing, defence mid-tier, pharmaceutical process, space components, the odd quantum-adjacent instrumentation firm. Name the companies. Find the programmes that will actually order. Measure the work in jobs that did not exist and capital that arrived, not in pilots launched.
The factories are already trialling the tools. The buyers are already shortening their lists. The regions that treat this as an announcement will still be drafting strategies when the purchase orders land somewhere else.
LibLeo Global works with economic development organisations on the conversion — companies, buyers, sites — in space, defence, life sciences, AI and quantum. The argument above does not depend on that.