Yujiro Masaki is a researcher at the External Affairs Institute and an entrepreneur specializing in artificial intelligence, digital policy, and international cooperation. He studied for an M.S. in European Studies at UCLouvain and previously served at Japan’s Ministry of Foreign Affairs (Middle East Bureau). He is the Founder and CEO of Jexen. Views are the author’s.
1. Introduction
Europe has spent the middle of this decade writing the world’s most ambitious artificial-intelligence rulebook and, more recently, an ambitious set of industrial plans. The AI Act entered into force on 1 August 2024. The AI Continent Action Plan of 9 April 2025 and the Apply AI Strategy of 8 October 2025 now speak of an “AI-first” public sector and a “buy European” preference. EuroHPC has selected nineteen AI Factories. InvestAI has been announced as a political target of €200 billion. The ReArm Europe / Readiness 2030 package of March 2025 sketched financial levers that could support up to €800 billion in additional defence expenditure.
None of this is yet the thing that historically turns a research technology into a strategic industry: a government that buys.
The commercial internet was not a spontaneous market. ARPANET, the Defense Advanced Research Projects Agency (DARPA), public standards, and government demand built the base; firms built the market. The same sequence produced the Global Positioning System, much of the semiconductor industry, and the high-assurance cloud. Europe still treats “intervention” as a distortion. That reading is wrong for foundation models, advanced chips, and energy-hungry data centres. These look less like smartphone applications and more like the early packet networks of the late 1960s: expensive, incomplete, dual-use, and not yet priced by ordinary consumers.
What Europe lacks is not another strategy paper. It lacks a DARPA-style habit — small teams, high discretion, a high tolerated failure rate — and, above all, a first serious customer. Without ministries, armed forces, and agencies buying European AI at operational scale, strategic autonomy will remain a vocabulary, not a capability. Refusing the risks that markets will not take on a useful timetable is not neutrality. It is a decision to buy, later and dearer, from someone else’s first customer.
2. The myth of the unassisted market
The popular history of the internet begins in a garage and ends on a stock exchange. Netscape, Amazon, Google, and later Facebook were products of private competition. That part of the story is true. It is also incomplete. It describes the shop, not the road.
The road was public. After Sputnik in 1957, the United States created the Advanced Research Projects Agency in 1958, later DARPA. The first host-to-host connection on ARPANET ran on 29 October 1969 between UCLA and the Stanford Research Institute. By the end of that year four university nodes were up. Demand came from defence and science, not from consumers. No firm in 1969 had a reason to lay a nationwide experimental network whose payback could not be stated.
The Department of Defense then forced the common language. On 1 January 1983 ARPANET cut over from the Network Control Program to TCP/IP. Hosts that did not migrate dropped off the network. Interconnection became cheap because a public authority made a standard mandatory. NSFNET, built from the mid-1980s to connect researchers to supercomputer centres and requiring TCP/IP, turned an experiment into daily university infrastructure. The backbone was decommissioned on 30 April 1995, after commercial networks were ready. Consumers arrived when the road was already there.
The sequence matters. A market did not invent interconnection and then invite the state to regulate it. The state financed a network no firm would have built, imposed a standard no vendor would have coordinated, and only then stepped back.
Mariana Mazzucato’s The Entrepreneurial State makes the same point about a later consumer object. The technologies that made the iPhone “smart” — the internet, GPS, the multi-touch lineage, the voice assistant that became Siri — were publicly funded long before they were privately assembled. Apple’s achievement was integration and design, not the invention of the stack. To credit only the integrator is to misread the production function, and then to write the wrong policy.
Silicon Valley’s origin is equally unromantic once the Cold War is restored. Early semiconductor demand came from Minuteman missiles, Apollo, and reconnaissance satellites. The Department of Defense and NASA paid for reliability and miniaturisation before calculators and personal computers existed as markets. Venture capital later accelerated firms whose prototypes were already visible. It did not, as a rule, fund the decade that made the prototype conceivable. Confusing those two risks is how a polity imports accelerators and wonders why the science does not follow.
GPS repeats the pattern. NAVSTAR was a Department of Defense programme, launched in 1973, with first satellites in 1978 and a full constellation operational in 1993. Civilian use became commercially decisive when Selective Availability was discontinued on 1 May 2000. The mapping industry did not conjure a satellite constellation. It inherited one. Europe’s Galileo programme shows that the Union already understands this logic in navigation. It has been slower to apply it to computation.
Cloud computing is often told as a purely commercial story, and Amazon Web Services, launched in 2006, did begin as a way to standardise Amazon’s own infrastructure. The story does not end there. In 2013 the CIA awarded AWS the Commercial Cloud Services contract, reported at the time as worth up to $600 million. Government demand did not invent the cloud. It made the cloud a national computer. In-Q-Tel, chartered in 1999 at the CIA’s initiative, is the same structure in another form: the state as customer and, sometimes, as early capital.
The competitive grammar is stubborn. The state takes risks that do not yet have a price. Universities translate those risks into protocols and people. Firms turn the result into something that can be bought twice. Transfer is not one-way. Once the commercial layer is large, the state returns as a customer, a regulator, and an export-control authority. Europe has become fluent in the last two roles. It is still hesitant in the first.
3. What DARPA actually did (and did not do)
DARPA is more often invoked than described. The useful description is institutional, not mythological.
The agency is small. Its fiscal year 2024 financial report records roughly 250 government personnel, including nearly 100 programme managers, overseeing about 300 projects on a budget of approximately $4.1 billion. It buys research from universities, firms, and other government laboratories rather than running giant intramural labs. Programme managers serve short tours, typically three to five years, so that a manager who must produce a technical surprise will not hoard a quiet programme for a career.
Selection is simple. Former director George H. Heilmeier left an eight-question catechism that DARPA still publishes. What are you trying to do, without jargon? How is it done today, and what are the limits? What is new, and why might it work? Who cares, and what changes if you succeed? What are the risks? How much will it cost? How long will it take? What are the mid-term and final exams? European subsidy applications are often an order of magnitude thicker and an order of magnitude less decisive. The catechism is a decision tool, not a communications product.
The Grand Challenge of 2004–2005 is the cleanest public illustration. On 13 March 2004, fifteen vehicles left Barstow, California, for a 142-mile unmanned run to Primm, Nevada. None finished. The furthest covered about twelve kilometres. DARPA announced a second race the next day. In October 2005 Stanford’s Stanley completed 132 miles in 6 hours and 53 minutes and took $2 million. Five teams finished. The 2007 Urban Challenge added traffic rules; Tartan Racing, led by Carnegie Mellon, won. The Challenge was a complement to ordinary procurement: a public, failed, then repeated demand signal. The state did not specify the winning vehicle. It specified an unsolved task and paid for attempts.
Two things DARPA did not do matter as much as what it did. It did not pick a single national champion and feed it until the budget expired. It ran parallel paths and killed the ones that failed the exams. Industrial-policy failure is not the act of choosing. It is the refusal to stop. Its programmes last a few years; its mission — to create and prevent technical surprise — lasts across administrations. Europe often inverts this: thirty-year visions, twelve-month work packages that cannot be stopped without a political inquest.
A DARPA-style habit is therefore not a logo. It is a small team with multi-year money, the authority to stop, an obligation to publish failure, and a customer waiting if the exams are passed. Without the customer, the rest is a research council with better stationery. The test is whether a programme manager can end a project in a week, and whether a ministry will buy the project that survives.
4. Government as first customer
Procurement moves industry faster than prohibition. Regulation writes what must not be done. A purchase order writes what must exist by a date.
The United States has used that instrument repeatedly. The Small Business Innovation Research programme, created in 1982, forces federal research agencies to set aside a share of extramural funds for small firms — a bridge across the stage that venture capital usually refuses, with a federal mission as the implied first buyer. Tesla’s January 2010 Advanced Technology Vehicles Manufacturing loan of $465 million financed the Model S and the Fremont plant. Tesla repaid the loan in May 2013, years ahead of schedule. Repayment is not the only test of a public risk. It is enough to retire the claim that public credit is always waste.
For AI, three domains will not be supplied by quarterly markets alone. The first is foundation research: mathematics, materials, semiconductor processes, cooling, and power electronics whose payoff is a decade away. Large American and Chinese firms can fund some of this from cash flow. Latecomers cannot wait for the winners to publish their training menu. The second is the capital stock: leading-edge fabrication, data centres, grid connections, and subsea cables. Assets that take more than ten years to recover will be built where a state guarantees offtake or writes predictable rules. The third is long research and development itself. Extreme ultraviolet lithography took more than twenty years from concept to volume. GPS and the internet took as long. Venture capital’s typical horizon is three to seven years. The actors who have historically held both long time and high technical risk are states, state-like procurers, and a few near-monopoly firms.
The correct division of labour is not “the market or the plan.” It is a government that buys knowledge whose social value survives failure, that places the first block of demand, and that sets standards; and a private sector that translates the result into a product someone will pay for twice. When the state designs the product and the firm waits to evaluate the science, both fail. When the state refuses the first purchase and calls the refusal “neutrality,” the purchase happens anyway — in another capital.
A first customer is not a protectionist warehouse. Buying an inferior system because it is local is how industrial policy becomes folklore. The specification should be operational: a model or secure cloud that improves on European data, under European legal constraints, on a timetable a ministry or a brigade can use. Origin is a means. Continuous improvement in the buyer’s mission is the good being purchased. A foreign foundation model may sit at the base for a time; the operational data, evaluation, fine-tuning, and the right to switch should not.
The same logic scales from a ministry to a supply chain. The CHIPS and Science Act of 2022 showed that the United States will treat fabrication as a state concern once it has given away too much of it. By 2026 there is no path back to a clean global market in frontier computation. There is only a choice of whose demand and export rules the market will be designed around. Calling every intervention a “distortion” presupposes an equilibrium that does not exist.
5. Europe’s missing buyer
Europe is not short of documents. It is short of invoices.
Public authorities in the Union spend around €2 trillion a year — on the order of 14 per cent of GDP — on works, goods, and services. More than 250,000 contracting authorities participate. In energy, transport, health, and education, the state is already the principal buyer. The European Court of Auditors noted in 2023 that innovative criteria are still rarely used in tenders. The Commission has said it will present a revision of the public procurement framework in 2026. Until that revision treats AI capability as something to be bought, not only as something to be lawful, the largest demand pool in the single market will continue to behave like a compliance machine.
The AI Act is a genuine achievement. It is also, by design, a regulator’s instrument. It classifies systems, prohibits some uses, and imposes obligations on general-purpose models. Trust is a public good. It is not a substitute for demand. A firm can comply with the Act and still have no European ministry as a reference customer. In that case the paying customer will be an American cloud, a Gulf fund, or a firm whose training cluster sits outside Union law.
The newer industrial plans begin to see the problem and then stop one step short. The AI Continent Action Plan and the Apply AI Strategy correctly note that only a minority of Union firms use AI — 13.5 per cent of enterprises with ten or more employees in 2024, rising to 20.0 per cent in 2025, according to Eurostat — and they now endorse a buy-European preference. Preference language in a communication is not a purchase. A contracting authority that prefers European AI in a recital and then awards on lowest price to a hyperscaler’s wrapper has not become a first customer.
Compute policy shows the same gap. EuroHPC’s AI Factories — selected in three rounds from December 2024 to October 2025, now nineteen sites plus thirteen Antennas — are a serious supply-side instrument. InvestAI, announced at the Paris AI Action Summit in February 2025, sets a political target of mobilising €200 billion, including support for a handful of AI Gigafactories. These are the right class of asset. They are not a market. A factory without an offtake agreement is a photograph. The American sequence was not “build a backbone and hope.” Defence and the National Science Foundation needed a network, so they paid for one, and later commercialised it.
Defence is the most plausible first customer Europe now has, and the most fragmented. The European Defence Fund has a 2021–2027 envelope of about €7.3 billion. That is not a replacement for national procurement. The ReArm Europe / Readiness 2030 package, the SAFE loan instrument of up to €150 billion, and the European Defence Industry Programme of €1.5 billion all push toward joint buying. The 40 per cent joint-procurement target for 2027 will matter only if the systems being bought include European AI for intelligence, command, logistics, and unmanned systems — and if losing bidders are allowed to die. Grants that produce demonstrators with no service customer reproduce the old pattern: Europe researches, others deploy.
Chips policy is learning the same lesson more slowly. The 2023 European Chips Act sought to mobilise more than €43 billion. The Commission’s Chips Act 2.0 proposal of 3 June 2026, still a proposal at the time of writing, adds demand-side tools: innovation procurement and security-of-supply criteria in public tenders. That is the correct correction. A fabrication plant without a buyer is a monument.
The political economy is familiar in Brussels. Intervention is easier to defend as a rule than as a purchase. The historical record runs the other way. The unassisted market in frontier computation is a myth. The assisted market is how the United States, and now China, actually built capacity. Europe’s fear of distortion has not preserved a clean market. It has preserved a market whose shape is set in Santa Clara, Seattle, and Shenzhen.
The missing buyer is not an abstraction. It is a ministry of defence that freezes software requirements on the day the tender is published; a health ministry that will not buy a model because no committee will own the first error; a tax administration that pilots twelve chatbots and scales none. Each already spends. Few spend in a way that teaches a European firm how to improve.
6. Recommendations
Six operational shifts would convert paper autonomy into a demand signal.
First, name the first customers. The Commission, the European Defence Agency, and national ministries of defence, interior, health, finance, and transport should each publish, within twelve months, a short list of AI capabilities they will buy — not study — by a dated milestone: classified document search; medical-image triage under a named hospital network; predictive maintenance on a named rail fleet; a brigade-level logistics twin; a customs risk engine. Heilmeier’s eight questions are a better template than a fifty-page impact assessment.
Second, rewrite procurement for software that learns. The 2026 revision of the Union public-procurement framework should create a fast path for iterative AI acquisition: short pilots, rapid payment, the right to kill, and the right to scale the winner across administrations without a new two-year contest. Award criteria should include measured improvement on the buyer’s own data, not a checkbox for “European headquarters.” Open-source and open-weight systems should be preferred where they meet the exam. A startup that waits nine months to be paid is not a first-customer success.
Third, attach offtake to compute. Access to EuroHPC AI Factories and to any future Gigafactory should be paired with public purchase commitments for models and services that pass independent evaluation. A startup that trains on a Union machine and then sells only to a non-European integrator is a leakage, not a success. The Union should buy the evaluation suites and the first production licences. Data centres that cannot secure grid connections on a known timetable will not host European training runs, whatever the industrial strategy says.
Fourth, use defence demand as the hard customer. The European Defence Fund, the European Defence Industry Programme, and SAFE should treat European AI software and secure cloud as capabilities, not as adjectives in a hardware programme. Joint procurement among at least three Member States should be the default for dual-use models in intelligence, simulation, and unmanned systems. National security exemptions should not become a way to buy American by default and European by communiqué.
Fifth, install DARPA habits without waiting for a DARPA brand. A handful of programme offices — in the Commission, in two or three willing Member States, and in a defence setting — should receive multi-year envelopes, fewer than two hundred staff each, programme managers on four-year tours with kill authority, and a legal duty to publish unclassified failures. Existing instruments should remain for the layers they already serve. The new offices should take only the bets those instruments reject for being too likely to fail. If, after three years, an office has closed no programmes, it should itself be closed. Accounting should treat technical failure as learning and fraud as fraud.
Sixth, buy improvement, not origin myths. A European-only mandate that locks in weaker models will discredit the strategy. The correct purchase is a stack whose data, evaluation, fine-tuning, and operations remain under European control, even when a foreign foundation model sits at the base. Over time, public demand should pull European base models up the stack where they are good enough. Joint buying with close partners on evaluation suites and secure compute is not a failure of autonomy. Fragmented national purchases of the same American service are.
None of these shifts require Europe to believe that officials out-predict entrepreneurs. They require Europe to accept that foundation research, first-of-a-kind plants, and the first production run of a dual-use model are not taken by markets on a useful timetable. The United States did not become an internet power by leaving the 1960s to venture capital. It became one by buying a network, forcing a standard, and then letting firms open shops on the road. Europe can still do the equivalent for AI. It cannot do it with regulation alone.
7. Conclusion
The internet was not made by firms. It was enlarged by them. The base — protocols, research networks, university talent, early demand — was public. The market — services, advertising, logistics, cloud — was private. GPS, semiconductors, and high-assurance cloud repeat the pattern. Foundation models, chips, and data centres are now in the phase that packet switching occupied in 1969: too expensive, too incomplete, and too strategic to be left to a consumer market that does not yet exist.
Europe has become the world’s most articulate regulator of AI and a serious builder of compute. It has not yet become a buyer. Strategy papers, AI Factories, and defence funds will not produce strategic autonomy if ministries, armies, and agencies write rules for systems they will not purchase. A DARPA logo would change little. A DARPA habit — small teams, high discretion, published failure, and a state that places the first serious order — would change the market that firms then build.
The Union does not need to choose between the market and the plan. It needs to choose to be the first customer of the systems it says it cannot live without. Until that choice is visible in tenders, deliveries, and killed programmes, “strategic autonomy” will describe an aspiration.
Selected references
Computer History Museum. “The First 50 Years of Living Online: ARPANET and Internet.” 29 October 2019.
DARPA. “The DARPA Grand Challenge: Ten Years Later.” 13 March 2014; and “The Heilmeier Catechism.”
DARPA. FY 2024 Agency Financial Report. November 2024.
European Commission. “AI Continent Action Plan.” COM(2025) 165, 9 April 2025; “Apply AI Strategy.” COM(2025) 723, 8 October 2025.
European Commission. “European Defence Fund”; “Public procurement”; “White Paper for European Defence – Readiness 2030,” 19 March 2025.
European Court of Auditors. Special Report 28/2023: Public Procurement in the EU.
EuroHPC Joint Undertaking. AI Factories selections, December 2024–October 2025.
Eurostat. ICT usage in enterprises: AI adoption 13.5% (2024) and 20.0% (2025).
GPS.gov. “Selective Availability.”
Mazzucato, Mariana. The Entrepreneurial State. London: Anthem Press, 2013.
National Science Foundation / Merit Network. NSFNET backbone decommissioned 30 April 1995.
Regulation (EU) 2021/697 establishing the European Defence Fund.
Thrun, Sebastian, et al. “Stanley: The Robot That Won the DARPA Grand Challenge.” Journal of Field Robotics 23, no. 9 (2006).
U.S. Department of Energy. Tesla ATVM loan: $465 million, January 2010, repaid May 2013.