Electricity Is the Bottleneck
AMD warned at SEMICON China that the next big constraint for chip expansion may be power, not silicon—AI-driven demand is creating a Jevons-style efficiency paradox that will strain grids and utilities across Asia. That puts energy planning front-and-center for fab capacity and data-center growth as chip demand accelerates. (digitimes.com)
SEMICON China ran March 25–27, 2026 at the Shanghai New International Expo Centre and lists Mario Morales, AMD’s VP and head of corporate strategy and partnerships, among its featured speakers. (semiconchina.org) Morales has been advancing the same macro argument across industry briefings, including a November 5, 2025 GSA Executive Forum presentation titled “The $1 Trillion AI Semiconductor Tectonic Shift.” (gsaglobal.org) Taiwan’s TSMC has publicly signaled it will prioritize 3nm capacity allocation for cloud-AI customers and key partners, an example of fabs already setting capacity priorities as compute demand shifts. (digitimes.com) SEMICON China’s program explicitly included sessions such as “Compound Semiconductors Power AI Era,” showing organizers paired semiconductor roadmaps with power- and materials-focused tracks. (semiconchina.org) Industry reporting at the event referenced IEA-based estimates that data‑center electricity consumption could approach roughly 945 TWh by 2030, raising the scale of power planning required for AI growth. (domain-b.com) Trade coverage and analysts at the conference noted divergent strategies: some U.S. tech companies are pursuing private generation and on‑site energy, while China is investing in broader grid upgrades to move megawatts to compute hubs. (icis.com) Analysts cited at and around the show warn modern hyperscale AI sites can demand on the order of 100–500 MW per campus, a level that typically requires multi‑year utility upgrades and regulatory coordination in Asian industrial zones. (domain-b.com)
