Taipower proposes 'fiber for power' siting strategy for AI data centers
At the 2026 Taiwan International Smart Energy Week, Taiwan Power Company (Taipower) presented a "fiber for power" strategy under its "integrated grid" theme, arguing that AI data centers (AIDCs) should be sited near generation and send data over long-distance fiber instead of transmitting large amounts of electricity over long distances, reducing transmission losses and pressure to build new power infrastructure. Taipower splits AI-era demand into "making AI" — wafer fabrication, memory, advanced packaging and high-end servers, with semiconductor-industry electricity use projected by Taipower to rise from about 52 billion kWh in 2026 to 110 billion kWh in 2035 — and "running AI," meaning AIDCs supporting model training and inference. Taipower data put approved AIDC supply at 1,441MW, including roughly 750MW added in 2026, more than half the cumulative total, and likened AIDCs to a shot put for their dense, concentrated load on regional grids.
The report states that in the generative-AI era computing has become infrastructure and electricity a key constraint on AI expansion, shifting data-center siting from proximity to users toward power availability and grid capacity.
AI運算需求快速成長，資料中心的選址邏輯也正在改變。過去資料中心多半靠近使用者端，以降低網路延遲與傳輸距離，但在AI時代，GPU叢集與大型AI資料中心（AIDC）的高密度用電需求，「電力從哪裡來」成為新考量。
台電在這周舉行的「2026台灣國際智慧能源週」，以「積體電網」為主題參展，並提出「以光代電」策略，鼓勵AI資料中心靠近電源端設置，透過長距離光纖傳輸資料，取代長距離輸送大量電力的模式。
值得注意的是，台電將AI時代的用電需求分為「製造AI」與「運行AI」兩類。其中「製造AI」方面，包括晶圓製造、記憶體、先進封裝及高階伺服器等AI供應鏈，台電預估我國半導體產業用電將從2026年約520億度，增加到2035年的1,100億度。
在製造AI中，台電已涵蓋AI供應鏈的用電需求，包括未來10年AI帶動GPU、HBM記憶體、先進封裝等的用電需求增加，不只是半導體，還包括伺服器。
#Taiwan Power Company (Taipower)#Taiwan International Smart Energy Week