主権AIへの賭け：2026年、日本のAI政策の全貌

Within seven months, Japan’s parliament passed its first AI-specific statute, the Prime Minister took the chair of a new AI Strategy Headquarters, the Cabinet adopted the country’s first AI Basic Plan, and the Ministry of Economy, Trade and Industry (METI) committed ¥1.23 trillion, roughly $8 billion, to AI and semiconductors for fiscal year 2026 alone.

For most of the past decade, Japan’s part in the AI conversation was described as patient: patient capital, patient compute, slow adoption curves. The 2025–2026 moves add up to something closer to an industrial-policy programme, and whether it works is an open question.

This briefing sets out what changed, what the statute does and does not do, who is building the sovereign models, and why the regulated yen-stablecoin rails under construction in Tokyo belong in the same picture.

The Statute

On May 28, 2025, Japan’s National Diet passed the Act on the Promotion of Research, Development and Utilization of AI-Related Technologies, usually called the AI Promotion Act. Most provisions took effect on June 4, 2025; the chapters establishing the AI Strategy Headquarters and authorizing the AI Basic Plan entered force on September 1, 2025.

The act contains no list of prohibited AI applications, no mandatory conformity assessments, no pre-launch registration regime, and no monetary penalties on business operators. The Future of Privacy Forum’s analysis describes a law that sets national goals for AI R&D, infrastructure and use, and tasks the government, rather than industry, with delivering them.

The EU AI Act put prohibitions, risk categories and administrative fines on AI developers. Japan instead attached a “comply-or-explain” expectation to its existing AI Governance Guidelines for Business (METI and the Ministry of Internal Affairs and Communications, v1.1 published March 2025; v1.2 in draft, reportedly focused on cross-border data flows and EU interoperability).

A multinational compliance team that has spent two years preparing for the EU AI Act will find the Japanese exercise different in kind. The question is what can be demonstrated in good faith, rather than what is forbidden.

The Control Tower

On September 1, 2025, the AI Strategy Headquarters was stood up inside the Cabinet Office. The Prime Minister chairs it, every cabinet minister sits on it, and a newly created Minister of State for Artificial Intelligence Strategy supports it.

AI policy in Japan is no longer one ministry’s job. The Headquarters coordinates every ministry that touches the technology: METI for industrial policy, MIC for telecom and content, the Financial Services Agency for AI in finance, the Personal Information Protection Commission for privacy, and the rest.

On December 23, 2025, the Cabinet approved the first AI Basic Plan, titled “Revitalizing Japan through ‘Trustworthy AI’.” It names three principles (promoting innovation while mitigating risks, agile response, integrated domestic and international policy) and four basic policies: Adopt AI, Create AI, Enhance AI Trustworthiness, Collaborate with AI.

Foreign vendors should read the “Create AI” section closely. The Plan prioritizes systematic promotion of domestic AI model development, which leaves room for the government to favor sovereign models in public procurement.

The Sovereign LLM Stack

Three companies anchor Japan’s domestic foundation-model ecosystem.

Sakana AI, founded in Tokyo in 2023 by former Google Brain researchers, became Japan’s most valuable AI startup in 2025 after a $135 million funding round. It received early GPU support through METI’s GENIAC (Generative AI Accelerator Challenge) program, which has run three rounds and committed roughly ¥33.9 billion in public funding across them.

NTT released tsuzumi 2 in October 2025, a 30-billion-parameter Japanese-first model built to run inference on a single H100 GPU. At that footprint, on-prem deployment is realistic for Japanese banks, insurers and government agencies that cannot move data offshore.

Preferred Networks’ PLaMo was built from scratch under GENIAC, pitched as a “sovereign AI foundation using domestic technology.” In December 2025 PLaMo Translate was selected for the Japanese government’s “Gennai” AI project.

ABEJA, Stockmark, Rakuten, Woven by Toyota and Ricoh sit in the wider GENIAC orbit. In December 2025 the government announced a ¥1 trillion ($7 billion) commitment over five years for AI and semiconductors, on top of METI’s ¥1.23 trillion for FY2026.

The Adoption Gap

A joint Yomiuri Shimbun / Teikoku Databank survey of 10,312 Japanese firms, fielded March 17–31, 2026, found generative AI in active use at 34.6% of companies, the first reading above 30%. Text creation and summarization lead the use cases at 45.1%. Accounting automation and other accuracy-dependent tasks lag.

Roughly half of all firms in the survey named information accuracy as their top worry. About 20% of AI-using firms reported widening performance gaps between employees, which lands badly in a country where wage compression is already a national argument. The OECD’s labor-market analysis points the same way: the limiting factor in Japan is talent rather than technology.

For investors the same numbers read differently. 65% of corporate Japan has not yet adopted generative AI. That is an enterprise software market in the world’s third-largest economy, and the recurring revenue sits in integration, training and governance services rather than in the models themselves.

The Stablecoin Rails

International coverage of Japan’s AI policy tends to leave out the settlement layer. At Tech for Impact Summit 2026, former Digital Minister Masaaki Taira put the two together on the Main Stage: “AI can’t have a bank account. So it’s stablecoins. Possibly crypto, too.” (Our full recap is in ‘AI Can’t Have a Bank Account, So It’s Stablecoins’: Former Digital Minister Masaaki Taira on Japan’s AI × On-Chain Finance Bet.)

Japan’s amended Payment Services Act, with further refinements taking effect by June 2026, restricts stablecoin issuance to three categories of licensed domestic entities: banks, fund transfer service providers and trust companies. JPYC, the first fully regulated yen-pegged stablecoin, launched in October 2025 and has since extended interoperability with Circle. SBI Holdings and Startale plan to issue JPYSC, a trust-issued yen stablecoin, in Q2 2026, with Shinsei Trust & Banking handling issuance and redemption. MUFG, SMBC and Mizuho are running joint proof-of-concept programs for trust-based yen stablecoins on the Progmat platform.

Taira’s point was that an agent needs a settlement instrument it can hold directly, and the Payment Services Act rules above are the first G7 framework that licenses one in yen.

For the C-Suite

A company that only deploys AI in Japan has a lighter compliance load than under the EU AI Act, since the Promotion Act carries no fines or conformity assessments. A company that builds models in Japan can draw on GENIAC compute and the ¥1.23 trillion FY2026 budget line. The two decisions (where to deploy, where to build) have different inputs and deserve separate planning.

Public-sector and regulated-industry buyers are likely to lean toward domestic models as the “Create AI” pillar works through procurement. A foreign vendor that wants those customers should be looking at local partnerships and on-prem deployment before the FY2027 budget cycle locks, with NTT’s single-H100 tsuzumi 2 as the benchmark for what “on-prem” now means to a Japanese bank.

With JPYSC due in Q2 2026 and the Payment Services Act refinements in force by June 2026, agentic payment workflows in yen become buildable inside the year. The firms that build them first will be the ones regulators and banks look to when the practice gets standardized.

The T4IS2027 Connection

Tech for Impact Summit 2027 convenes in Tokyo on May 18–19, 2027, with AI governance, the sovereign-compute economy and the regulated convergence of AI and on-chain finance on the program. These are the threads Taira, Marwala and Nakazawa pulled on at T4IS 2026.

Membership in T4IS 2027 is invitation-only. Senior leaders can find details on the membership page; the speaker roster will be announced in stages through the autumn.

Search “Tech for Impact Summit” to follow updates, or read the recap series at /blog for the arguments behind this briefing.