\n\n
Key developments:
- Grid operators are already overwhelmed. Oncor reported a backlog of 200 GW of interconnection requests in 2025, with 186 GW coming from AI‑focused data centres, and the queue grew to 271 GW by early 2026.
- Individual AI tasks still consume measurable power. An 8‑billion‑parameter Llama model needs about 57 J per text query, while the largest Llama (400 B parameters) uses roughly 3,400 J. Video generation is far more demanding – a five‑second clip costs about 3.4 MJ (≈1 kWh).
- Overall demand is exploding. AI token processing has risen ten‑fold annually since 2024, now topping 300 trillion tokens a day, driving a 50 % jump in AI‑focused data‑center electricity use in 2025 and a projected 9.5‑15.3 % share of U.S. electricity by 2030.
- Industry is responding with efficiency breakthroughs. Microsoft’s upcoming Maia‑300 system aims to deliver 40 % more tokens per megawatt from 2027 onward, while Nvidia and partners are standardising 800‑VDC power architectures to cut conversion losses.
- Geopolitics matter: China is expanding data‑center capacity in renewable‑rich western provinces, effectively pairing surplus wind‑solar output with AI workloads.
Why it matters:
- Power shortages could throttle AI progress, as megawatt‑scale requests outpace grid upgrades, risking delays, higher costs, and potential regulatory push‑back.
- Energy‑intensive tasks like video generation amplify the carbon footprint, even as per‑task efficiency improves.
- New high‑efficiency hardware (Maia‑300, 800 VDC) and renewable‑focused siting (China’s west, U.S. solar‑rich sites) could mitigate grid strain and curb AI’s climate impact.
What to watch next:
- Grid‑capacity expansions and policy decisions on who pays for upgrades, especially in the U.S. where utilities are beginning to scrutinise megawatt requests.
- Adoption rates of 800‑VDC power in upcoming AI factories and the performance data from Microsoft’s Maia‑300 launch.
- Further IEA and Berkeley Lab updates on AI‑related electricity shares, which will signal whether efficiency gains can keep pace with token‑level demand growth.
- International moves, such as China’s renewable‑linked AI hubs, that could set a model for energy‑intensive AI deployment. [[1, 2]]