AMD warns electricity, not chips, will bottleneck AI
At Semicon China 2026, AMD warned that AI’s efficiency gains are driving more demand for compute—which paradoxically makes electricity the ultimate constraint for scaling AI systems. The Jevons‑paradox framing pushes energy optimization and edge/distributed inference to the front of architecture planning. (digitimes.com)
Mario Morales — listed as AMD’s VP and head of corporate strategy and partnerships — appeared on SEMICON China’s speaker roster for the March 25–27, 2026 conference in Shanghai. (semiconchina.org) AMD’s presentation cited a near‑term AI market valuation figure of about US$1.7 trillion that underpins current demand forecasts for datacenter and edge silicon. (digitimes.com) The International Energy Agency projects global electricity use from data centres will more than double to roughly 945 TWh per year by 2030, with the United States and China together driving about 80% of that growth. (iea.org) McKinsey’s infrastructure analysis estimates roughly US$6.7 trillion of data‑centre investment may be required by 2030, and global capacity demand could reach the 171–298 GW range depending on adoption scenarios. (datacentremagazine.com) Industry measurements show average rack power density climbed from roughly 8 kW to about 17 kW in two years, and analysts expect averages near 30 kW per rack by 2027 — a jump that motivates liquid cooling and redesigned power architectures. (networkworld.com) Hyperscalers are already locking multiyear energy supplies to address that gap: AWS expanded a PPA for up to 1.92 GW of nuclear power with Talen Energy (contracted through 2042) and providers are increasingly using large PPAs, microgrids and on‑site storage to secure firm capacity. (datacenterdynamics.com) Regulatory and market friction is rising as jurisdictions track who bears grid upgrade costs: recent reporting highlights local ratepayer pushback and spikes in capacity prices in U.S. regional markets where major datacenter clusters are expanding. (cnbc.com)
