AI boom collides with power constraints
Analysts warn that surging AI compute demand is creating a new chokepoint in electricity supply, with Goldman Sachs forecasting large increases in data‑centre power demand and utilities flagging major grid investment needs. Some operators are shifting to behind‑the‑meter generation and microgrids to secure capacity, and companies that can guarantee on‑site power have seen market value move quickly. (benzinga.com) (cbsnews.com) (datacenterknowledge.com) (livemint.com)
Artificial intelligence data centers are running into a simpler limit: not enough electricity can be delivered fast enough. (goldmansachs.com) Goldman Sachs said on March 4 that it now expects data-center power demand to rise 220% by 2030 from 2023 levels, up from its earlier 175% forecast. The bank tied the increase to heavier hyperscaler spending on artificial intelligence infrastructure. (goldmansachs.com) Utilities are preparing for that load with record spending plans. A PowerLines review of 51 investor-owned utilities found at least $1.4 trillion in capital spending planned through 2030, and CBS News reported April 14 that data centers were a top driver cited in earnings calls. (powerlines.org) (cbsnews.com) The basic problem is timing. Data centers can be built in a few years, but new transmission lines, substations and generation often take longer, leaving what Data Center Knowledge called a “time-to-power” gap between developers and utilities. (datacenterknowledge.com) That gap is pushing operators to generate electricity at the site instead of waiting for a full grid hookup. Data Center Knowledge reported April 15 that behind-the-meter systems and microgrids are being used as bridge power and, in some cases, as a longer-term operating model. (datacenterknowledge.com) The Electric Power Research Institute put the scale in national terms last month: U.S. data centers could consume 9% to 17% of all electricity by 2030. Its state-level scenarios show Virginia already above a 20% share, with several other states potentially crossing that threshold by the end of the decade. (powering-intelligence.epri.com 1) (powering-intelligence.epri.com 2) The pressure is already moving markets. Bloom Energy shares rose 24% to a record $218.68 on April 14 after the company said it would supply Oracle with up to 2.8 gigawatts of fuel-cell systems for data centers, with an initial 1.2-gigawatt contract. (livemint.com) CNBC reported the same day that Oracle shares gained more than 4% as investors focused on the expanded Bloom deal and Oracle’s race to add capacity for artificial intelligence workloads. The agreement gives Oracle access to on-site power that can be deployed faster than some conventional alternatives. (cnbc.com) (livemint.com) Consumers are part of the story too. CBS said utilities often recover grid spending through regulated rates, and PowerLines estimated residential customers could end up paying for nearly half of the planned capital spending if current patterns hold. (cbsnews.com) (powerlines.org) The new bottleneck in the artificial intelligence buildout is no longer just chips or servers. It is whether developers can secure power, on the grid or beside it, before demand outruns the wires. (goldmansachs.com) (datacenterknowledge.com)
