Anthropic sounds the alarm on America's AI energy gap

Anthropic just shifted its focus from neural networks to the electric grid, and its message to Washington is simple: build more power, or lose the AI race.

The AI lab published an energy policy blueprint called "Build AI in America," arguing that without an unprecedented buildout of domestic power generation and transmission, frontier AI training in the United States could hit a hard physical wall.

The numbers in the report are eye-opening. Anthropic projects that training a single state-of-the-art model will require a 2-gigawatt data center in 2027, jumping to 5 gigawatts by 2028.

Factor in competing frontier labs, and Anthropic estimates U.S. frontier training alone will demand 20 to 25 gigawatts by 2028. That is roughly double the peak electricity consumption of New York City. Add the energy needed to run inference for those models every day, and the U.S. AI sector will need at least 50 gigawatts of capacity in just three years.

Cutting red tape to feed power-hungry clusters

The report sounds a clear alarm about the pace of infrastructure. While the U.S. added a few dozen gigawatts of capacity last year, China added over 400 gigawatts, creating a massive disparity in how fast new computing clusters can come online.

To bridge the gap, Anthropic outlined two strategic pillars. While giving credit to the Trump administration for moving on nuclear targets and streamlining environmental reviews under NEPA, the company wants the federal government to move much faster.

First, Anthropic urges officials to open up federal lands for mega-datacenter construction to bypass years of local zoning fights. It also suggests using federal national security authorities to fast-track grid interconnection for top-tier training facilities.

Second, the lab wants a broad overhaul of energy permitting, covering natural gas, geothermal, and next-generation nuclear power. It even recommends creating national strategic reserves of critical grid components and gas turbines, alongside nationwide permits from the Army Corps of Engineers to speed up environmental water reviews.

Why it matters

When a frontier AI lab starts lobbying for gas turbines and federal land use, the nature of the AI race has clearly changed from who writes better algorithms to who can plug into the grid.

For enterprise users and developers, this means the compute crunch is shifting upstream to utility companies. If energy supply cannot keep up with cluster sizes, access to tomorrow's frontier models could become significantly more expensive and tightly rationed.

Strategically, Anthropic is trying to turn power generation into an urgent national security priority by pointing directly at China's massive infrastructure advantage. We think this strategy is smart politically, because framing electricity as a geopolitical race helps cut through federal bureaucracy.

However, the real risk lies in local pushback. Asking to bypass regional zoning and use federal powers to jump the utility queue will look like Big Tech getting a taxpayer-backed shortcut. If surging AI demand starts hiking household electric bills, the political backlash against data center buildouts will only intensify.