Ruth Porat didn't mince words. Google's President stepped onto the CERAWeek stage in Houston on Monday and dropped a line that cut through two years of industry hand-wringing: "We are concerned that we are not full throttle on energy." She wasn't talking philosophy. She was talking about the raw electrical power America needs to fuel its AI ambitions — power that, right now, simply isn't showing up fast enough.
The numbers are jarring. Lawrence Berkeley National Laboratory found data centers already chewed through 4.4% of U.S. electricity in 2024. By 2028, that figure could hit 12%. The IEA projects global data center demand more than doubling — from 415 TWh to 945 TWh — before 2030 arrives. Former Google CEO Eric Schmidt told Congress the U.S. needs 92 additional gigawatts of generation capacity just to sustain AI growth. That's the equivalent of building 92 nuclear power plants from scratch.
So what's actually broken here?
Most headlines frame this as an energy shortage. That framing misses the point entirely. America has abundant energy potential. What it lacks is synchronization — land, turbines, interconnection rights, substations, transformers, switchgear, gas supply, transmission authority, and local political buy-in all landing on the same timeline as a massive AI campus. Think of it like trying to throw a dinner party where the stove arrives Tuesday, the plates show up Friday, and the chef is booked until next March.
The dysfunction shows up everywhere you look. PJM Interconnection, managing roughly 180 GW across 13 states, has flagged potential supply shortfalls of up to 60 GW in coming decades. Its capacity prices exploded to $333 per megawatt-day in a recent auction — a twelvefold jump in two years. Meanwhile, ERCOT in Texas is sitting on 226 GW of large-load connection requests, nearly three times the entire current U.S. data center footprint. Switchgear lead times hit 44 weeks by mid-2025. Gas turbine slots stretch years out. Morgan Stanley projects a net U.S. power shortfall of 9 to 18 gigawatts through 2028.
Washington has started reading the room. The White House's March 2026 "Ratepayer Protection Pledge" essentially told Google, Microsoft, Meta, Amazon, and OpenAI to build, bring, or buy their own generation — stop leaning on public grids and burdening everyday consumers. Executive orders have targeted advanced nuclear deployment and directed FERC to fast-track AI data center interconnection requests. The message from D.C. is clear: public infrastructure won't carry this load alone.
Three popular trades are currently miscalibrated.
Nuclear looks sexier than it performs right now. Meta's January deals with TerraPower, Oklo, and Vistra for up to 6.6 GW of nuclear capacity confirm that corporate appetite exists — but near-term cash flows don't follow. Advanced reactors carry heavy execution risk and fuzzy timelines. Gas and electrical equipment monetize far sooner.
Utilities aren't a blanket buy either. Those with workable regulatory compacts and room to expand transmission in AI-friendly jurisdictions can act like toll roads — steady, durable, lucrative. Those stuck in congested or politically hostile markets may collect flashy load announcements without moving the earnings needle. Texas remains the most constructive U.S. market. PJM rewards owners of existing scarcity more than buyers chasing new load.
Data center pipeline figures have also become almost meaningless. Announced campus sizes and capex numbers sound impressive until you realize grid waits in primary markets stretch beyond four years. The genuinely scarce asset today is a credible, documented path to energization — confirmed interconnection, transformer procurement, fuel access, phased energization. Vague projects deserve serious skepticism.
Here's the deeper story Wall Street keeps sleeping on.
AI economics are quietly morphing into utility economics. When each new compute cluster demands bespoke behind-the-meter generation, long-lead equipment bought at scarcity prices, and custom transmission arrangements, the return on AI infrastructure capital stops depending on model performance or GPU costs. Delivered electrons govern everything.
The market keeps over-allocating narrative value to chip suppliers and under-allocating it to powered land, electrical equipment, interconnection rights, dispatchable generation, and grid-flexibility software. Porat's warning wasn't about demand. It was about deliverability. In this cycle, access to power has become a more decisive competitive differentiator than access to capital — and most of Wall Street hasn't fully priced that yet.
not investment advice