Data center growth hits power limits
Reports show US data‑center expansion is bumping into power‑grid constraints and local regulatory pushback, which is increasing buyer interest in more power‑efficient inference silicon reported reported. Energy limits are becoming a procurement decision variable, not just a facilities one.
Developers had about 241 GW in the U.S. data‑center project pipeline at the end of 2025([bloomberg.com)]. Q4 2025 additions slowed to roughly 25 GW, about half the Q3 pace, signaling a pullback in new announcements([publicpower.org)]. Wood Mackenzie reported utilities have publicly committed roughly 63 GW to large loads—a volume WoodMac equates to about a 12% increase in U.S. electricity demand([woodmac.com)]. The DOE‑backed Lawrence Berkeley analysis shows data‑center electricity use rose to ~176 TWh in 2023 and could reach 325–580 TWh by 2028 (about 6.7–12% of U.S. demand) under current trajectories([newscenter.lbl.gov)]. Hyperscalers are responding with custom silicon: Meta announced plans to design and deploy four new generations of MTIA inference/training chips over the next two years to lower per‑inference cost and power([about.fb.com)]. Industry procurement is updating scorecards—vendors are now evaluated on "tokens per watt" and inference‑per‑dollar metrics, according to industry tooling and design guidance([synopsys.com)]. Analysts also report buyer sourcing now formally includes energy‑stability and grid‑connection risk as supplier qualification criteria([randtech.com)]. Power strategies are shifting on site: Bloom Energy’s 2026 analysis and Data Center Frontier coverage document rising adoption of onsite generation, virtual power plants and gigawatt‑campus designs to shorten interconnection lead times([bloomenergy.com)]. Open silicon is gaining traction as an efficiency lever—SiFive’s NVLink Fusion initiative and RISC‑V International’s 2025 report show RISC‑V platforms and tighter CPU‑GPU interconnects are being positioned for data‑center AI deployments([sifive.com)].
