SemiAnalysis projects AI data center power demand will triple to 300 GW by 2030, up from 100 GW in 2025. The firm's model, shared via @SemiAnalysis_, highlights grid bottlenecks and a pivot to on-site generation.
Key facts
- 300 GW: Projected AI data center power demand by 2030
- 100 GW: Estimated demand in 2025, per SemiAnalysis
- 3x: Growth factor over five years
- Grid queues: Identified as primary bottleneck
- On-site nuclear/gas: Shift to bypass grid delays
The number: 300 GW by 2030. SemiAnalysis, the AI infrastructure research firm, projects global AI data center power demand will triple from roughly 100 GW in 2025 to 300 GW by 2030. The forecast, shared on X according to @SemiAnalysis_, covers GPU clusters, cooling, and associated infrastructure.
What's driving the growth. The model attributes the increase to GPU cluster scaling — more chips per site, higher rack densities — and rising cooling loads. Efficiency gains in newer chips (e.g., 3nm vs. 5nm) are factored in but do not offset the demand curve. The firm assumes no major breakthroughs in chip efficiency or cooling technology over the period.
The bottleneck is the grid, not the chips. SemiAnalysis identifies grid interconnection queues and transformer lead times as the primary constraints. This aligns with public reporting: the U.S. interconnection queue backlog grew 30% between 2022 and 2024 [per Berkeley Lab's annual report]. The firm notes a structural shift toward on-site generation — nuclear and gas — to bypass grid delays, echoing recent hyperscaler deals like Microsoft's Three Mile Island restart and Amazon's nuclear SMR investments.
Why this matters more than the headline. The 300 GW figure is a 3x jump, but the real signal is the split between grid-supplied and on-site power. If on-site generation dominates the marginal GW, it changes the economics of AI infrastructure — capex shifts from land and substations to reactors and turbines. It also pressures utility regulators to fast-track grid upgrades or face stranded assets.
What the model doesn't say. SemiAnalysis did not disclose regional breakdowns or the share of on-site vs. grid power in the forecast. The firm's X post is a summary, not the full model. Expect a detailed report in the coming weeks, likely behind its subscription.
Key Takeaways
- SemiAnalysis projects AI data center power demand triples to 300 GW by 2030, from 100 GW in 2025.
- Grid bottlenecks and on-site generation shift are key.
What to watch
Watch for SemiAnalysis's full report with regional splits and the on-site vs. grid power ratio. Also track hyperscaler nuclear deals in Q3 2026 — if Microsoft or Amazon announce additional SMR contracts, the 300 GW figure may trend higher.
[Updated 14 Aug via dcd_news]
Form Energy has raised $750 million to expand its West Virginia iron-air battery plant, explicitly citing surging data center demand. The company has signed supply deals with Google and Crusoe, positioning long-duration storage as a grid-bottleneck workaround—complementing the on-site nuclear/gas shift SemiAnalysis highlights. [per DataCenterDynamics]
