Grid delays force AI builders off American power grids
- Texas halted data center connections after power requests hit 474 gigawatts, five times peak grid capacity.
- Grid wait times reached four years, forcing AI builders to strip private jet engines for off-grid power.
- Capital is fleeing choked U.S. power grids for behind-the-meter solar and energy-rich foreign nations.
Texas hit a wall. When Governor Greg Abbott halted new data center connections, ERCOT was tracking 474 gigawatts in requests - more than five times the state's peak record.
The freeze in Texas exposed a structural breaking point. On This Week in AI, analysis showed that data centers account for 90 percent of all new power requests in the state. AI labs can double model parameters in months, but power utilities operate on decades-long planning horizons. Tax GPT CEO Kash Ali noted on the show that if Texas shuts its doors, developers have few domestic alternatives with comparable land and policy.
The bottleneck rapidly expanded beyond local regulators to global hardware supply chains. Discussing Anthropic's revenue targets on All-In with Chamath, Jason, Sacks & Friedberg the next day, Gavin Baker described severe backlogs for industrial power hardware. Turbine blade manufacturers running 24-hour shifts with 40-ton presses cannot keep up, forcing desperate developers to strip turbine engines off retired private jets for off-grid electricity.
"Old-world industrial supply chains now dictate the pace of artificial intelligence."
- Gavin Baker, All-In with Chamath, Jason, Sacks & Friedberg
Physical limits are pushing Silicon Valley into unprecedented financial engineering. To maintain momentum despite grid delays, Wall Street constructed a $500 billion private credit framework backed by Nvidia, BlackRock, and Goldman Sachs. Private lenders now securitize graphics processors like commercial real estate, funding compute capacity through future token revenue rather than traditional bank loans or equity dilution.
By Aug 15, 2026, the story shifted from regional grid strain to a systemic power deficit across the country. Clean energy analyst Ramez Naam reported on Moonshots with Peter Diamandis that connection wait times jumped from 15 months to nearly four years. While chipmakers will supply 200 gigawatts of compute by 2030, U.S. utilities will deliver barely 100 gigawatts.
"Whichever nation eliminates grid friction first will host the next generation of artificial intelligence."
- Ramez Naam, Moonshots with Peter Diamandis
Because monopoly utilities lack incentives to build quickly, hyperscalers are abandoning traditional power grids entirely. Developers are turning to behind-the-meter natural gas turbines and massive solar-battery arrays. A facility in Dubai pairs five gigawatts of solar panels with 19 gigawatt-hours of battery storage, generating continuous baseload power at six dollars per watt - less than half the capital cost of recent U.S. nuclear builds.
This energy gridlock is redrawing the global tech map. Unlike human populations, AI workloads are geographically mobile. Compute capital is now migrating toward energy-rich jurisdictions in Australia, Mexico, and the Middle East. Whichever nations modernize land-use and energy regulations first will host the physical backbone of frontier artificial intelligence.
Money can buy chips and train algorithms, but it cannot instantly forge steel turbines or lay high-voltage transmission lines. The software revolution has officially run out of electricity.