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.
Introduction
Mario Draghi’s 2024 report on European competitiveness stated the problem with unusual bluntness. The Union’s productivity gap with the United States opened in the mid-1990s because Europe failed to capitalise on the first digital revolution: it neither produced the platform firms nor diffused digital technology into the rest of the economy. Around 70 per cent of foundational AI models developed since 2017 have come from the United States. Three American hyperscalers account for more than 65 per cent of the global cloud market and of the European one. The largest European cloud operator holds about 2 per cent of the EU market.
Those figures are now familiar in Brussels. So are the replies: the world’s first comprehensive AI law, the vocabulary of data protection, the Brussels Effect. All of that is true. None of it answers Draghi. A jurisdiction can write the rules of a market and still fail to own the stack the market runs on.
Japan is the closest historical analogue, and not the one European officials usually reach for. Japan did not lose the internet for lack of engineers. It built hardware at world standard, over-adapted that hardware to a rich and protected home market, and never made a global software platform. By the late 2000s it had some of the fastest household broadband on earth and a mobile internet that predated the iPhone by eight years — and almost no say over search, cloud, or smartphone operating systems. It became an advanced user of systems whose defaults were set elsewhere.
Europe is closer to that Japan than it admits. The Artificial Intelligence Act, the General Data Protection Regulation, and the Brussels Effect are instruments of genuine external power. They are not a substitute for models, compute, or firms that set defaults. The lesson is not “deregulate.” Industrial policy and market design have to travel together. If Europe only polices AI, it becomes a well-governed customer of American and Chinese systems.
What Japan actually lost
The claim that Japan “missed” the internet is, to anyone who worked near the networks, almost offensive. In October 1984 Jun Murai and colleagues connected the Tokyo Institute of Technology, Keio University, and the University of Tokyo on JUNET. The WIDE Project began in 1988. Internet Initiative Japan started commercial ISP service in December 1992, within hailing distance of American commercialisation. After a late start, Japan led much of the world in household fibre: NTT introduced mass-market FTTH in 2001, and by the mid-2000s Japanese residential networks were among the fastest then in service.
Hardware was not the problem either. Japanese firms took the global DRAM market in the 1980s: overall DRAM share reached about 80 per cent by 1987, and Japan’s share of world semiconductor output peaked near half the market in 1988. The 1986 US–Japan Semiconductor Agreement measured how seriously Washington took that position.
What Japan lost was not connection. It lost authorship of the order: the standards the rest of the world would rely on, the squares where users and developers would gather, and the fees those squares would collect. Search, social networks, cloud, and smartphone operating systems were occupied by American firms. English was a disadvantage. The stronger constraint was the country’s own success. When a firm is winning in hardware, software looks like the control logic attached to a machine. An accessory does not invite the imagination of a platform.
NTT DoCoMo’s i-mode is the cleanest illustration, and European operators once knew it at first hand. The service launched on 22 February 1999 and made the mobile phone a daily internet terminal years before Western markets did the same. Content providers were paid through the phone bill. Domestic subscribers passed 40 million in 2003 and peaked at about 49 million in July 2010. On 31 March 2026, DoCoMo shut i-mode down with the rest of its third-generation network.
Calling i-mode a “Galápagos failure” is unfair. It was a domestic optimisation that worked. Compact HTML, carrier-specific specifications, and vertical coordination produced a finished product for a finished market. Japanese users experienced the mobile web earlier than most Europeans. That experience did not become a square a developer in Helsinki or Bangalore could enter with the same code. After 2007 the iPhone and the App Store converted a contest over handset finish into a contest over operating systems and distribution. Japanese manufacturers raised device quality for a home market of roughly 127 million people; Apple and Samsung opened an operating system onto a market of billions.
LINE is a near-exception; Rakuten and SoftBank are large at home. Exceptions disprove determinism; their scarcity is evidence of structure. The NEC PC-98 family dominated Japanese personal computing long after the IBM-compatible PC had become the world machine. Europe has its own versions — national payment schemes, national health-data silos — and should recognise the type.
What was actually lost was not the capacity to build. It was the capacity to organise a market that strangers would join. That requires a product incomplete enough that others can finish it. Japan’s quality culture is a genuine asset. In the early life of a platform, the same asset becomes a barrier. TCP/IP was crude. Crude is sometimes how a market is made.
Hardware without a market
The ability to make technology and the ability to make a market are not the same skill. Japanese firms demonstrated the first, repeatedly. The second is the organisation of demand that does not yet exist.
MITI’s Fifth Generation Computer Systems project, run through the Institute for New Generation Computer Technology from 1982 to 1992, is the public example. It spent on the order of ¥54 billion over about eleven years. It produced prototype machines and a generation of researchers. It did not rewrite the world’s computer market. American research fails too. One difference is what happens after failure: programmes funded by the Defence Advanced Research Projects Agency often dissolve into the next contract and into components a firm can buy. The Fifth Generation, as it perfected a dedicated machine and language, moved further from the general market.
i-mode is the successful version of the same diagram. The technology and the business model were elegant inside Japan. Their premises were Japanese regulation, billing custom, and handset distribution. When those premises could not be exported, the technology remained and the market did not. DoCoMo did try. From 2002 it licensed i-mode to European and other operators — E-Plus, KPN, BASE in Belgium, Bouygues, and others, some eighteen markets in all. Overseas carriers withdrew after a peak around 2005. In Europe the operator was an infrastructure provider, not a vertical gatekeeper, and Nokia had little reason to build to a proprietary Japanese specification.
Regulation, procurement, and standardisation freeze a successful adaptation. Until the 1985 liberalisation of Japanese telecommunications, the public network was in practice NTT’s domain. Japan later opened the market; fibre followed. The delay was not “being a regulated country.” It was the speed with which new subjects — researchers, small carriers, software firms — were recognised as first customers. Public procurement can be a first market. In the United States, defence and the research networks performed that function for the internet. Japanese procurement has preferred track record and scale. The preference reduces failure. It also reduces the learning that only an unproven technology can provide.
Risk capital is the other half of the same problem. A bank-centred system and the internal funds of large groups do not naturally sit with a money-losing software firm for years. Conditions improved after 2000. They did not produce a thick market of the Silicon Valley type. When private money is thin, public money can fill the front end. If procurement does not sit at the other end, public money ends as a report.
The industrial structure that won in automobiles and consumer electronics — a large firm at the centre, suppliers in series — fits software badly. A large information-systems department issues requirements to a systems integrator and accepts a system that meets them. Meeting the specification is quality. It is also the opposite of a platform practice, which updates the specification from user behaviour. Software was treated as a costed man-month and source code as something to rebuild per contract. Able people went to machines and to the shop floor. The error was to keep that allocation after the internet had moved the centre of value from objects to relations and data.
Europe will recognise the type. A continent strong in machinery, vehicles, medical devices, and aviation is strong in the layer Japan won. It is tempted, as Japan was, to treat models as the control logic of those machines, and to let procurement complete a domestic product and call that industrial strategy. The temptation is rational. It is also how the last success missed the layer that now sets defaults.
The Brussels Effect is not a stack
Anu Bradford’s “Brussels Effect” remains the most precise account of European regulatory power abroad. The Union has a large market, dense rules, and a corporate preference against splitting product lines. When those conditions hold, an internal rule becomes a de facto world standard.
The General Data Protection Regulation, applicable since 25 May 2018, is the textbook case. Consent, erasure, and administrative fines entered the language of jurisdictions that had no vote in the Council. That is real power, and a specific kind of power. GDPR exported a vocabulary. It did not produce a European platform. Eight years on, the firms that set defaults in consumer software are not European.
The Artificial Intelligence Act is the sequel, written as product-safety law. Regulation (EU) 2024/1689 was published in the Official Journal on 12 July 2024 and entered into force on 1 August. It prohibits a short list of practices, imposes conformity assessment on high-risk systems, and attaches duties to general-purpose AI models.
Application was always staged. Prohibitions applied from 2 February 2025. Obligations on general-purpose AI models applied from 2 August 2025. Most of the rest of the regulation, including the Article 50 transparency rules, applied from 2 August 2026. The Digital Omnibus on AI — Regulation (EU) 2026/1744, in force from 27 July 2026 — postponed the heavy high-risk duties: Annex III standalone systems to 2 December 2027, Annex I product-embedded systems to 2 August 2028.
Three legal families now sit on the same harms. The Union classifies by product risk. The United States, still without a horizontal federal statute, governs by actor, procurement, and export control. China regulates by content and setting, quickly and vertically. A unit of order is not a stack.
A stack is compute, data, models, distribution, and the firms that make those layers the default for everyone else. Draghi’s figures describe the present stack. Mistral AI, founded in Paris in April 2023, is the existence proof that Europe can still produce a laboratory that ships open-weight models. Microsoft’s partnership is also a fact. Neither is an ecosystem. One firm, still dependent for much of its capital and inference on extra-European cloud, does not cancel a 2 per cent share of the European cloud market.
Europe’s plausible third way, if there is one, does not run through consumer chat. It runs through the industries in which the Union already sets the safety conditions of the product: machinery, automobiles, medical devices, aviation, energy. Those are the Annex I and Annex III battlefields. A product that has once paid a heavy conformity duty is easier to buy in public procurement. That is a real industrial opening. If the duty is so heavy that the experiment is never run, the opening closes. The Omnibus bought time. Time is not a strategy.
The diplomatic temptation is to treat vocabulary as victory. Language from the Hiroshima AI Process and the Act’s own categories are already the terms in which other capitals speak. That is a Brussels Effect in motion, and a foreign-policy success. It is not industrial policy. DoCoMo’s domestic finish was a kind of standard. The standard did not travel. Travel of the rule is still not ownership of the default.
I studied European integration at UCLouvain and later served in Japan’s Ministry of Foreign Affairs. The two vantage points produce the same observation. In a foreign ministry a technical standard, once set outside, becomes the given of a negotiation. In a firm the foreign tool is rational, the data accumulate on someone else’s base, and adoption is recorded as a cost saving. Diplomats call the first vulnerability. Companies call the second vendor lock-in.
What Europe should copy — and refuse
Copy the refusal to treat engineering as a consumer story. Japan’s internet problem was never that households bought personal computers too late. Copy, too, the residual strength. Draghi noted that Europe still hosts around 22 per cent of worldwide activity in autonomous robotics and around 17 per cent in AI services. That is one of the few layers on which a third way can be built. Copy the later Japanese decision to open telecommunications as a warning about speed: the opening came after the platforms had already been assigned.
Copy, from the American side of the same history, the sequence that actually produced a stack: a public first customer, universities that circulate talent, and risk capital that can lose money for years on software. Mariana Mazzucato’s account of the entrepreneurial state is a correction of a myth, not a licence for every subsidy. Europe already possesses pieces of that sequence — EuroHPC, the AI Factories, the European Investment Bank. The pieces are not yet a sequence. A supercomputer that a start-up cannot book, and a start-up that a ministry will not buy from, are two halves of the Fifth Generation problem.
Refuse the completion of a domestic product as a substitute for an open base. The single market is larger than Japan’s home market, which is why the Brussels Effect works. Size does not cancel the reflex. A “sovereign” model trained only on data that twenty-seven legal services will certify can become a European PC-98: locally finished, globally unread. Refuse the treatment of software as an accessory to a machine, especially in the industries Europe still leads. A vehicle that runs on a foreign model and a foreign cloud is a European chassis, not a European product, in the layer that will matter in 2035.
Refuse procurement that treats track record as the only form of quality. The preference is defensible in a bridge. It is lethal in a platform. A ministry that will buy AI only from the integrator that already holds the contract is reproducing the Japanese information-systems department in European colours. Refuse also the aesthetic of complexity. Accumulation of compromise produces legitimacy. It also produces a compliance cost that a hundred-person firm cannot staff.
Refuse, finally, the story that the alternative to the Act is American laissez-faire. That is the poor binary Japan also reached for: deregulate and the market will arrive, or pick winners and the ministry will deliver. The United States talked markets and built a public backbone. Japan talked industrial policy and under-built the market-making half. Europe is at risk of talking rights and under-building the stack. The correction is not fewer rights. It is a government that designs a market: a narrow ban, a short conformity path, a first customer, and compute that a new firm can actually touch.
What must not be repeated is a specific combination: domestic finish ahead of world openness; the incumbent’s record ahead of a new subject’s learning; software as a bolt-on to hardware; the state designed only as a regulator. AI reproduces the combination easily, because the surface is convenient. A ministry that plugs a foreign general-purpose model into its documents tomorrow can report a successful deployment. Past the report sit compute, data, and the definition of safety — set in someone else’s institution. The import of convenience can be the export of sovereignty. That is not an argument for autarkic models. It is an argument against spending the next thirty years only as a user.
Recommendations
Six measures follow.
First, the EuroHPC Joint Undertaking, the Commission’s DG CONNECT, and the member states that host AI Factories should treat compute as a first market, not as a monument. The AI Continent Action Plan of April 2025 and the Factories already selected are the right instruments if access time does not accrue to laboratories that already know how to write the bid. The Governing Board should privilege start-ups and deploying public authorities, with decision times measured in weeks, and the Commission should report annually to the European Parliament on Factory hours that went to firms under ten years old.
Second, the EU AI Office and the national competent authorities should spend the Omnibus delay on a short path to conformity. Prohibitions (February 2025) and general-purpose-model duties (August 2025) should remain. The high-risk regime that now lands in December 2027 and August 2028 will bind European firms first if guidance and sandboxes are late. The Office should issue, by mid-2027, model technical files for the main Annex III use-cases and staff sandboxes to produce a reasoned yes or no. A rule that only a compliance department can survive is industrial policy for incumbents.
Third, the European Investment Bank, the InvestEU implementing partners, and national promotional banks should underwrite software platforms as they already underwrite fabrication plants. The missing money in Japan was not always the headline sum. It was capital that could stay with an unprofitable software product until a public buyer existed at the far end. An AI chip plant without a European model, and a European model without a ministry that will buy the service, are two unconnected subsidies. The EIB’s digital window should be instructed by the Commission and the Council to take platform and model risk on a ten-year clock.
Fourth, defence, health, and interior ministries in France, Germany, Italy, Poland, and the Nordic and Baltic states, working with the European Defence Agency, should run first-buyer programmes in the regulated industries Europe still owns. American research networks and defence demand did not pick the commercial winners of the 1990s. They paid for the first unreliable version. Europe’s equivalent is a contracted, evaluated deployment of European or openly licensed models in military logistics, hospital administration, the grid, and customs — with the evaluation published.
Fifth, the Commission and the Council should write the forthcoming Cloud and AI Development Act as market design rather than as a second AI Act. The objective Draghi named — a European hyperscale option, and a security policy that can use American capacity without surrendering encryption and ring-fencing — is an industrial objective. It fails if it becomes a labelling exercise. Interoperability, portability of fine-tunes, and public-data pools that firms will actually train on are the market-making half. Data-protection supervisors should be at that table so that the GDPR is a design constraint, not a late veto.
Sixth, the European External Action Service, the Commissioner for Trade, and member-state foreign ministries should stop counting exported vocabulary as stack power. The Hiroshima AI Process and the Act’s categories are useful diplomatic goods. They should not be briefed as evidence that Europe is “winning AI.” Japan’s experience is that a country can be an advanced user and a sophisticated regulator, and still arrive at the next negotiation as a price-taker. Partners will accept European terms of trust where those terms are enforceable. They will not move their training runs to a jurisdiction that offers only terms.
Conclusion
Japan did not lose the internet because it lacked engineers, fibre, or industrial policy. It lost the layer that organises users and collects the fee. It built the machine, completed it for a home market that rewarded completion, and watched a cruder, more open product become the world. i-mode, which closed in March 2026, is a warning that a finished domestic success can be the mechanism of a strategic loss.
Europe’s position in August 2026 is better advertised and not as different as it looks. The Union has a comprehensive law and a working Brussels Effect. It does not have the models, the compute, or the firms that set defaults. The Omnibus postponed the hardest duties; the AI Factories are being stood up; Mistral and a few peers exist. None of that is yet a stack. Draghi was right that a window remains, and right that the last one was missed.
The mistake to refuse is the one Japan made in a different vocabulary: to treat the completion of the internal rule as the industrial strategy. The Act is not the error. The error would be to stop at the Act. Industrial policy that does not open a market, and regulation that does not leave a short path for a new firm, will produce a familiar result. Europe will be a well-governed customer. The defaults will be written in California and Shenzhen. The papers will be in order.
A third way is possible only if it is a compute way as well as a rights way, and only if the state that writes the prohibition also buys the first imperfect system from a firm that does not yet have a track record. Japan still has the factory skill. Europe still has the legal skill. Neither, alone, writes the next thirty years. The countries that will are the ones that keep those skills attached to a market strangers can join.
Selected references
- Bradford, Anu. The Brussels Effect: How the European Union Rules the World. Oxford University Press, 2020.
- Draghi, Mario. The Future of European Competitiveness. European Commission, September 2024. https://commission.europa.eu/document/download/97e481fd-2dc3-412d-be4c-f152a8232961_en
- European Commission. “AI Act enters into force.” 1 August 2024. https://commission.europa.eu/news-and-media/news/ai-act-enters-force-2024-08-01_en
- European Commission. “AI Omnibus enters into force.” 27 July 2026. https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force
- European Commission. AI Continent Action Plan. COM(2025) 165 final, 9 April 2025.
- Regulation (EU) 2016/679 (GDPR). Applicable 25 May 2018.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act). OJ L, 2024/1689, 12 July 2024.
- Mazzucato, Mariana. The Entrepreneurial State. Anthem Press, 2013.
- NTT DoCoMo. “i-mode Subscribers Surpass 40 million.” 30 October 2003.
- “NTT Docomo to discontinue decades-old i-mode.” The Japan Times, 30 October 2019.
- Information Processing Society of Japan, Computer Museum. “Start of the 5th Generation Computer Project.”
- Japan Network Information Center. “The Internet Timeline.”
- Semiconductor History Museum of Japan. “80s Trends in the Semiconductor Industry.”