On 1 September 2026, Taiwan's Ministry of Digital Affairs (MODA) defended its Build-Own-Operate (BOO) project, launched earlier in 2026, to draw private investment into AI computing power centers. Minister Lin Yi-jing said the aim is to add capacity equal to more than 10,000 GPUs within a year, possibly far more in an optimistic case, according to Focus Taiwan. The framing matters. Lin was answering former minister Huang Yen-nun, who noted a day earlier that South Korea plans to buy 260,000 Nvidia GPUs by 2030, Japan has about 40,000, and Taiwan about 30,000.
The strongest case for state-bought compute
The argument for a Korean-style government purchase deserves a fair hearing. Frontier compute is scarce, and its supply is lumpy. A state buyer can place a very large order, secure allocation from Nvidia, and guarantee demand that no domestic firm could. Compute is also an input to research, public services and national security, so governments in many countries treat it as infrastructure, like ports or grids. If Taiwan's startups and universities cannot get GPUs at reasonable prices, they will rent them from foreign clouds, and that gap would widen.
Those are real concerns. They are an argument for ensuring access, though, not for the state owning every chip.
Why the BOO model fits better
Lin's counter was that having the government buy all computing capacity and give it free to industry could distort market mechanisms. He said technological progress should come from "free and fair competition" in the private sector, per Focus Taiwan. That is the right instinct, for three reasons.
- Hardware depreciates fast. A state that buys 260,000 accelerators at once locks in one generation of silicon on one procurement cycle. Private operators refresh on commercial timelines and carry the obsolescence risk themselves.
- Free compute misprices demand. Zero-price allocation invites hoarding and pushes the state into choosing winners among applicants. A priced market with targeted discounts lets scarce capacity flow to the best uses.
- Headline GPU counts are a poor yardstick. Comparing 30,000 with 260,000 measures budgets, not capability. What matters is how much usable compute reaches researchers and firms, and how fast.
The design in MODA's April 2026 announcement is also more specific than the September rhetoric suggests. Applicants must invest more than NT$300 million excluding land and must reach at least 15 PetaFLOPS at FP32 precision. They must use their own land or lease existing facilities. Applications closed on 14 May 2026. Operators are expected to provide computing power free or at a discount to government agencies and academic research. They are also encouraged to offer deals to local residents and disadvantaged groups and to help SMEs gain access at reasonable cost. Centers must also follow the Cyber Security Management Act, including off-site backup and business continuity plans.
That is a sensible hybrid. Private capital builds and owns the assets, and the state buys access for the public-interest users it cares about through the terms of the programme. Taiwan's government is not holding the stranded-asset risk.
Where the risks sit
Electricity. Lin said power supply "should not be a problem" on current assessments, and MODA will coordinate with the Ministry of Economic Affairs and Taiwan Power Co., according to Focus Taiwan. Compute centers are among the most power-hungry loads an island grid can add, and the programme's one-year clock leaves little room for grid upgrades. If connection approvals lag, the 10,000-GPU target slips regardless of how many bidders there are. The ministry should publish the capacity assessment behind Lin's assurance so investors can price the risk.
Soft obligations. The free or discounted access expected of operators is described as an expectation. If it becomes a de facto mandate with vague pricing, it could deter the very bidders the programme needs. Clear, published rate schedules and an audit trail would be better than discretionary pressure.
Measurement. MODA says 10,000 GPUs of added capacity. It should report capacity in delivered, usable compute, with utilisation rates, not just in chip counts, so the public can judge whether the programme works.
How this fits the AI Basic Act
The BOO project sits alongside Taiwan's AI Basic Act, which the Legislative Yuan passed on its third reading on 23 December 2025. As Baker McKenzie describes it, the Act imposes no immediate operational obligations on the private sector, and sector regulators will issue detailed duties based on a MODA-developed risk classification framework. The Act also envisages funding, incentives and an experimentation-friendly environment. That is a light-touch approach. Pairing it with compute supply that private operators finance is coherent: regulate by risk, and expand capacity through markets.
The risk is drift. If the risk framework hardens into heavy obligations while compute stays tight, Taiwan would burden its AI developers without equipping them. MODA holds both pens, which is an advantage if it keeps them aligned.
What to watch
Three things will show whether Lin's bet works: which firms win BOO status and at what scale, whether Taipower can connect them on schedule, and whether access terms stay predictable. If all three hold, Taiwan will have added meaningful capacity without turning the state into the world's largest GPU reseller. If not, the critics' point about the gap with Korea gets harder to dismiss. Either way, the right comparison is delivered compute per researcher and per startup, not who bought the most chips.