Goldman flags 220% AI power growth
- Goldman Sachs says AI is turning data centers into a power-and-grid story, with electricity demand now rising fast enough to reshape build schedules. - The eye-catching number is 220%: Goldman’s April 2026 forecast says global data-center electricity use could hit 1,350 TWh by 2030. - That shifts the bottleneck from chips alone to substations, transformers, utility hookups, and generation that can actually arrive on time.
The AI build-out has a new choke point — electricity. Not chips, not even buildings, but the messy physical stuff between a server rack and the grid. Goldman Sachs’ latest call is that global data-center power demand could rise 220% from 2023 to 2030, hitting about 1,350 terawatt-hours. The bigger point is what that number does to the rest of the system: it turns utilities, transformers, and interconnection queues into the real schedule risk. (msn.com) ### What actually changed? Goldman had already been warning that AI would push power demand sharply higher. In April 2024, its U.S.-focused work said global data-center power demand would more than double by 2030 and that utilities would need about $50 billion of new generation investment to support the load. By January 20(msn.com)raming is more aggressive: 220% growth globally by 2030, with the U.S. taking roughly 60% of the increase. (powerit.com) ### Why does 220% matter so much? Because this is not a normal demand bump. Goldman’s own late-2025 public note still talked about data-center power demand rising 175% by 2030 from 2023 levels. Moving from that range to 220% means the bank now sees the grid stress as broader, faster, or both. At 1,350 TWh, data centers start to look less like a niche load pocket and more like a new industrial sector arriving all at once. (goldmansachs.com) ### Why aren’t GPUs the whole story anymore? A GPU is only useful if the campus around it is energized. That means land with transmission access, a utility willing and able to serve the site, substations, switchgear, transformers, backup generation, and cooling systems sized for very dense A(goldmansachs.com)wer to scale than compute demand. Basically, you can buy servers faster than you can buy the electrical path that makes them run. (goldmansachs.com) ### Where do delays show up first? At the utility interface. The Department of Energy says distribution transformers are still facing supply-chain constraints and long lead times. Separate grid-planning work says large-load interconnection queues are now a serious obstacle for data centers, (goldmansachs.com)e steps happen before the flashy part of the project. If they slip, everything behind them slips too. (energy.gov) ### Why are transformers suddenly a star? Because they are the adapter between ambition and reality. AI campuses can ask for hundreds of megawatts, sometimes more than a small city. Every one of those campuses needs electrical equipment that is specialized, heavy, and not easy to replace with a workaround. Industry tracking outside Goldman shows large transformer and turbine(energy.gov)s keep treating “power equipment” as its own AI trade. (powermag.com) ### Does this hit households too? Yes — that is the uncomfortable part. Goldman told clients in February that electricity prices rose 6.9% in 2025, more than double headline inflation, and argued data centers could account for 40% of electricity-demand growth through the end of the decade. If utilities have to build more generation and grid equipment quickly, some of that cost lands in customer bills. (cnbc.com) ### So what’s the real takeaway? AI infrastructure is now a power-development problem wearing a tech label. The winners are not just whoever gets the best chips. They are the developers, utilities, and equipment suppliers that secure power first — and secure it years ahead of need. (goldmansachs.com)
