Rapidus, the Japan-government-backed foundry venture, has begun risk production of 2-nanometer logic chips at its Chitose fab in Hokkaido, a decisive milestone for Tokyo's decade-long effort to rebuild a domestic leading-edge logic capability. The first wafers entered the gate-all-around transistor module in late April, and process partner IBM has confirmed that early electrical test data is broadly aligned with the company's New York reference line.
The State-Industry Bet
Rapidus is the most expensive industrial gamble Japan has placed since the heyday of MITI in the 1980s. METI has committed roughly 920 billion yen of direct subsidy across multiple tranches, with additional support pledged through 2027 if performance milestones are met. The customer roster behind the venture spans Toyota, Sony, NTT, Denso, NEC, SoftBank, Kioxia, and MUFG, an unusually broad domestic coalition that reflects how seriously Japan now takes the logic-foundry gap with Taiwan and Korea.
Chief executive Atsuyoshi Koike, a former Western Digital and Sandisk executive, has said publicly that Rapidus's commercial-volume target is 2027 and that the Hokkaido site will eventually require a second module to reach meaningful scale. Risk production is, by definition, low-volume and aimed at validating yield, defect density, and design-rule maturity rather than at filling customer orders.
The Process and the Tools
The Rapidus 2nm node uses gate-all-around nanosheet transistors, the same architectural family as TSMC's N2 and Samsung's SF2. The process draws heavily on IBM Albany research, including the high-NA EUV roadmap and a backside power delivery scheme that Rapidus expects to pull forward to its 1.4nm successor. The Chitose fab houses two ASML EUV scanners delivered in 2024, with a third arriving in the third quarter and the company's first high-NA tool slated for early 2027.
Reported Early Wafer Metrics
- Defect density tracking within 1.4x of comparable IBM Albany reference
- SRAM bit-cell yield reportedly above 60 percent on test reticles
- First customer test chip taped out by a Japanese SoC vendor
- Cycle time on early lots running roughly 3.2x mature-node norms
The Customer Question
The hardest part of the Rapidus story is not making 2nm chips. It is selling them. TSMC's N2 has effectively absorbed all anticipated leading-edge demand from Apple, AMD, MediaTek, and Nvidia through 2027. Samsung Foundry remains the obvious second source. For Rapidus to find paying customers, it needs either a Japanese champion willing to anchor volume, such as Sony imaging or Toyota's automotive AI, or a US fabless customer interested in geographic diversification.
Industry sources in Tokyo say Rapidus has signed framework agreements with at least two US-based AI accelerator startups for low-volume risk lots in 2026 and 2027. Volume will likely depend on whether at least one anchor automotive or imaging program commits before the end of 2026.
Comparisons With TSMC and Samsung
TSMC's N2 ramp at Hsinchu is already producing wafers at meaningful volume, and the company's Kumamoto JASM site continues to scale on more mature nodes. Samsung's SF2 has had a more difficult ramp, with yield reportedly improving but customer wins still concentrated in internal Exynos and Galaxy designs. Rapidus enters this market two years late, with a smaller customer base and a smaller fab footprint, but with a structurally different proposition: a fully transparent, single-customer-per-shuttle service model aimed at design houses that want intimate access to the line.
What Could Go Wrong
Three risks loom. First, defect density may stall, as it did for Samsung's first GAA node. Second, the customer pipeline may not materialize fast enough to justify the second module. Third, METI's political appetite for additional subsidy could waver if a new government takes a harder line on industrial policy.
What to Watch in the Next Six Months
Three milestones matter most. The completion of the qualification flow for the first external customer, expected by the third quarter. The decision on whether to commit to a second EUV module at Chitose, which would push committed capex above 5 trillion yen. And, perhaps most tellingly, whether any non-Japanese fabless customer commits to a production tape-out at the node. Without that signal, Rapidus remains a national-security project. With it, it becomes a real foundry.