âš¡ AI's Insatiable Energy Appetite
The International Energy Agency's May 2026 World Energy Outlook special report on AI projects that electricity consumption by AI-focused datacenters will triple to 1,200 terawatt-hours by 2028, roughly equivalent to Japan's total annual electricity consumption. Each Blackwell B200 GPU draws 1,200 watts under full AI training load, and a single 100,000-GPU training cluster requires 150-180 megawatts of power delivery—comparable to a small city.
The IEA notes that while per-token inference efficiency has improved 10x since 2023 (thanks to model quantization, MoE architectures, and faster hardware), the Jevons paradox is in full effect: cheaper inference drives higher usage, which more than offsets efficiency gains in aggregate electricity demand.
Grid operators in key datacenter regions are raising alarms. Dominion Energy (Virginia, home to 70% of global internet traffic) projects that datacenters will consume 85% of its new generation capacity through 2035. The Electric Reliability Council of Texas (ERCOT) has added AI datacenter load as a specific category in its long-term planning for the first time, forecasting 40 GW of new AI load by 2030.
In Ireland, datacenters already consume 21% of national electricity, and the Commission for Regulation of Utilities has implemented a de facto moratorium on new datacenter connections in the Dublin region.
💡 Nuclear Renaissance and Geothermal Innovation
In response to energy demand and carbon commitments, major tech companies have become the world's largest energy buyers and developers. Microsoft signed a 10.5-gigawatt nuclear power purchase agreement with Constellation Energy in September 2025, the largest corporate clean energy deal in history, and is financing the restart of Three Mile Island Unit 1 (shuttered since 2019) for 835 megawatts of dedicated datacenter power.
Google has committed $8 billion to advanced geothermal projects with Fervo Energy and has signed PPAs with Kairos Power for small modular reactors (SMRs) starting delivery in 2030.
Amazon's AWS has committed $12 billion to a combination of nuclear (SMRs with X-energy), solar-plus-storage in Arizona and Virginia, and is the anchor customer for a 960 MW wind farm in Oklahoma dedicated entirely to datacenter operations. Meta has taken a different approach, co-locating its largest training clusters near existing hydropower in the Pacific Northwest and investing $3.5 billion in transmission infrastructure to bring stranded renewable energy to new datacenter sites in the Midwest.
📋 Fusion and Long-Shot Bets
Sam Altman-backed Helion Energy, which has raised $2.2 billion including $375 million from Altman personally, is targeting a 2028 net-energy-gain fusion demonstration at its Everett, Washington facility. Microsoft holds a first-refusal PPA with Helion for 50 megawatts of fusion power by 2029, though energy analysts consider this timeline aggressive given that no fusion experiment has yet achieved sustained net energy gain.
Commonwealth Fusion Systems (backed by Bill Gates and Google) is targeting 2032 for a commercial fusion plant in Virginia. The fusion timeline is unlikely to impact the 2026-2030 energy gap, but tech industry investments are accelerating fusion R&D beyond what government programs alone would support.
Water consumption is an increasingly contentious dimension of AI energy growth. AI datacenters are projected to consume 6.6 billion cubic meters of water annually by 2027, primarily for cooling, according to a UC Riverside study. In water-stressed regions like Arizona, datacenter water use has become a political flashpoint, with Google and Microsoft pledging to be "water-positive" by 2030 through watershed restoration investments that offset consumption.