How much energy do data centers and artificial intelligence use?

The market mistake is to treat AI power demand as a broad electricity supercycle when the real bottleneck is node-level capacity. That shifts economic rents away from generic utilities and toward assets that can actually deliver firm MWs in constrained load pockets: merchant generation, nuclear, grid equipment, transmission, and data-center infrastructure in places like Northern Virginia. For hyperscalers such as GOOGL, the implication is constructive: lower perceived energy drag reduces the odds that AI monetization gets discounted for ESG or permitting risk.

The next 1-3 months are more about multiple discipline than fundamentals. If investors have been bidding up "AI power" beneficiaries on a simple shortage narrative, this data argues for a narrower winner set and lower enthusiasm for broad utility beta. That creates a cleaner relative-value setup: infrastructure names with backlog and order visibility can keep working, but regulated utilities with slow rate-case transmission will struggle to convert the story into EPS faster than the market already anticipates.

Contrarian takeaway: consensus is probably still overestimating how much AI load matters at the global level and underestimating how much it matters in a few congested grids. The thesis breaks if utility load forecasts, interconnection queues, and PPA pricing do not tighten over the next two reporting cycles; in that case, the power-themed re-rating should fade. Over 6-18 months, the key variable is not query energy intensity but chip efficiency and siting permits, which should cap the durability of any broad-based energy scare.