The IEA sees data center power use nearly doubling by 2030 — what could push that forecast higher or lower?

This explores what could make the IEA's forecast too high or too low. The baseline: global data center electricity use rises from 485 TWh in 2025 to about 950 TWh in 2030, and AI-focused facilities roughly triple How fast will data center electricity demand actually grow?. The IEA already names two brakes, supply chain bottlenecks and sensitivity to financing. The corpus suggests financing is the one to watch, because electricity is a surprisingly small part of the bill. Epoch AI estimates that a 1-gigawatt AI data center costs about $8.5B a year to own and run. Servers account for about $5B of that and energy for only about $0.6B What actually drives AI data center costs?. So the buildout depends less on whether power is cheap and more on whether investors keep paying for chips that have to earn back their cost before they become obsolete.

That makes investor patience a hidden variable. Brynjolfsson, Rock and Syverson's 'productivity J-curve' explains why AI's economic payoff can look weak in the early years. Companies spend heavily on things that don't show up in the statistics, such as retraining staff, redesigning processes and writing software, and that work pays off only later Why does AI productivity appear weak when investment is high?. If investors read the dip at the bottom of the curve as failure, capital could pull back before the payoff arrives, which would slow the buildout below the IEA path. If they wait it out, the forecast holds.

The strongest case for faster growth comes from demand. Emberson and Sevilla estimate that demand for AI output (measured in tokens) is growing about 10× a year, while the capacity to serve it grows only about 3.4× a year. They point to early signs of strain, such as Anthropic reducing usage quotas at peak hours Will AI inference demand outpace supply capacity soon?. A gap that size puts steady pressure on companies to build. Geopolitics adds to the push. When Amodei proposed deliberately slowing AI capability development, US and Chinese leaders rejected the idea within days Can AI safety pacing work without government cooperation?. Voluntary slowdowns look unlikely to limit construction.

The less obvious factor is efficiency research, and it could push either way. Routing each query to the best of several smaller models can match a frontier model at 27% lower cost Can routing beat building one better model?. Some researchers argue that better memory design now returns more than adding parameters Has memory architecture replaced parameter count as the scaling frontier?. Both point to more capability per watt. But a smaller model given more compute while answering can match a larger one on hard prompts Can inference compute replace scaling up model size?. That moves energy use from one-time training to the moment of every answer, which adds to the demand pressure above. Efficiency gains may change where the electricity goes more than they reduce the total.

A gap worth knowing about: the corpus says little about the physical constraints, such as grid connection queues, permitting, and turbine or transformer shortages. Those may decide the outcome as much as anything here. On the economics and demand side, the picture is clear: whether demand keeps outrunning supply, and whether investors stay patient, will likely matter more than the price of electricity.