“Taiwan's AI strategy has a public data problem” — Taipei Times
I wrote a feature for the Taipei Times titled “Taiwan’s AI strategy has a public data problem.”
The problem
Taiwan is preparing its grid for AI: four new combined-cycle units at Hsinta and Taichung, about 5.2GW combined, are due to start supplying or testing by year end, within a plan adding roughly 26GW of gas-fired capacity by 2035, predominantly imported LNG.
On the demand side, the numbers don’t line up. In June, officials cited about 1.2GW of identified AI data-center demand within the 2035 horizon, plus a 2GW-to-3GW allowance for faster growth. Taipower separately told the Taipei Times it had received 75 transmission-level requests totalling 4,523MW, agreeing to 47 for 3,331MW — with about 40% of those cases expiring before a formal application was filed. Investigative outlet The Reporter published earlier Taipower data in April, with different totals and an expiry rate above 70%. A National Chengchi University researcher could not reconcile either series with the only figure tied to an actual operating load: 303.5MW physically connected.
What’s missing
Neither dataset is necessarily wrong — they have different cutoff dates and may treat expired cases differently. But no published note explains how they fit together. Taipower produced a detailed AI demand estimate in 2024; the Taipei Times could not locate a subsequent update in the same format. National demand forecasts, meanwhile, fold AI into a flat 2.5%-a-year growth rate through 2035 without isolating its share.
That opacity has a real cost. The Ma-anshan nuclear plant has been offline since May 2025; if data-center and semiconductor demand exceeds the national forecast, Taiwan must burn more gas, accelerate other generation, or cut demand elsewhere. Taiwan has no international gas pipeline — all its LNG arrives by sea. Data centers concentrate in the north while generation sits largely in the south, and individual siting decisions remain largely internal to Taipower.
What I propose
Taipower already holds most of the data needed. An anonymized connection queue, by municipality and substation, separating requested, agreed, expired, contracted and energized capacity, would let preliminary requests be distinguished from firm commitments without naming customers. The Energy Administration could publish a current range for AI demand, its assumptions, and the effects on gas consumption and grid construction.
Taiwan may decide that greater reliance on imported gas is an acceptable price for its AI and semiconductor ambitions. Publishing the underlying data would not settle that choice — but it would let lawmakers, municipalities and host communities test the same assumptions as the ministries making them.