Executive summary

Two prizes from AI in Taiwan's data centres
Taiwan · NT$ bn a year · build = 150-250 MW a year at NT$300-380 mn per MW, excluding IT equipment
Build spend on new halls, a year to 2030NT$45-95 bnAI inside operations, 2030NT$1.0-3.8 bnof which cooling and energyNT$0.5-2.3 bnBAR = MIDPOINT OF RANGE, NT$ BN A YEAR
Source: [Kira estimates] from [Taipei Times Aug 2026]; [TechNews May 2026]

The answer in three lines

AI is the tenant now. About 7,000 GB300 GPUs at one Guishan hall were nearly sold out before going live [TWM Mar 2026], and Foxconn's first B300 cluster is over 90% leased [Visionbay 2026].

AI as a tool is still a promise. Machine-learning control cut cooling energy by up to 40% at a Google site [DeepMind 2016]; we found no published result from a Taiwanese hall.

Power sets the pace. Of 4,523 MW requested, 303.5 MW is contracted at live sites [Taipei Times Aug 2026], so every saved kilowatt is worth more than the staff it frees.

Sell into the AI build now, and prove the in-hall gain in kilowatt-hours before pricing AI on headcount.
DeepMind 2016 · Taipei Times Aug 2026 · TechNews May 2026 · TWM Mar 2026 · Visionbay 2026 · Kira estimates = KIRA analyst triangulation · full citations in Methodology and sources

GPU clouds, cooling makers and AI-hall operators sit in the most-exposed corner; Taipower is exposed through demand, not through its own work

Each value-chain stage scored on how far AI changes what it sells and how far AI can change the way it works.

AI exposure by value-chain stage
Taiwan data centres · scores 1-5 · outside-in estimates · Oct 2026
Work changes, demand steadyMost exposedLeast exposedDemand changes, work steadyHow much AI changes demand for what they sell (1-5)How much AI can change their own work (1-5)GPU and AI cloudsCooling and power makersAI-hall operatorsHyperscale cloudsTaipowerBuilders and M&EFacility servicesLandowners and parks
Source: [Kira estimates] from the evidence in Chapters 1 and 2

Two kinds of exposure

Demand side: AI tenants buy dense, liquid-cooled halls, so GPU clouds, cooling makers and hall operators see their order books change first. Taipower feels it as load: 4,523 MW of requests [Taipei Times Aug 2026].

Work side: a hall is a sensor-rich system, and the first AI cooling models learned from data its sensors already logged [DeepMind 2016]. Facility services change most as maintenance moves from calendar to condition.

CaveatScores are our outside-in reading of public signals, not survey results [Kira estimates].
A supplier that sells into the build and also runs the controls has both kinds of exposure working for it.
DeepMind 2016 · Taipei Times Aug 2026 · Kira estimates = KIRA analyst triangulation · full citations in Methodology and sources