AI Data Centers Are Driving Up Power Demand — Nuclear PPAs Become Standard for the Giants
Training and inference turn data centers into a new grid load; Microsoft, Google, Amazon and Meta race to sign nuclear supply deals, and small modular reactors come into focus.
AI data-center energy use in 2026 has moved from an environmental talking point to an industrial constraint: training and inference want 24-hour baseload, while grid upgrades and transformer lead times run to years. Exact global shares depend on the accounting; this site will not write a single percentage. What is clear is that the load is rising fast, and power supply is starting to cap compute.
Microsoft, Google, Amazon and Meta have signed large-scale nuclear power agreements — Microsoft even restarted the Three Mile Island plant to power AI data centers. Small modular reactors (SMRs) have become tech giants' new favorite, with several startups raising billions of dollars.
Meanwhile, sparse computing, quantized inference and dedicated chips push energy per unit of compute down — and larger training and inference runs often eat the savings. Forecasts for AI's 2030 share of global electricity do not agree; the real bottleneck is dispatchable baseload, not any one projection.
From Talking Point to Hard Constraint: Power as the Ceiling on Compute
AI's energy problem has moved from an environmental talking point to a hard industrial constraint: data centers need 24/7 baseload power, while grid expansion and transformer lead times run years. That is the deeper reason Microsoft restarted Three Mile Island and everyone is racing to sign energy deals (see our Three Mile Island coverage) — behind Stargate's $500B and the Nvidia-OpenAI 10-gigawatt LOI, the real bottleneck has shifted from chip supply to power.
Efficiency and Scale Racing in Parallel
On one side, nuclear and SMRs expand supply; on the other, inference-efficiency gains cut demand: sparse computing, quantized inference and dedicated chips sharply lower energy per unit of compute. This shares DeepSeek's logic of engineering training cost down to a fraction (see our coverage) — when power becomes scarce, 'intelligence per watt' replaces 'how many GPUs' as the new competitive axis.
Our Take
Tracking tech companies' power purchase agreements (PPAs) has become the leading indicator of their compute roadmaps — power contracts get signed before chip orders. For investors, the second half of the AI race is extending from a 'chip arms race' to an 'energy arms race': whoever locks in stable, low-carbon, scalable power holds the ticket to the next stage of compute expansion.
This is an original analysis by the AI Tools Daily editorial team, based on publicly available information. Opinions are for reference only.
AI Tools Daily is a bilingual newsroom covering AI tool launches, product updates and industry trends. Editorial standards · Report a correction