Two days after the federal moratorium on frontier artificial-intelligence training died in the House Ways and Means Committee, the battle has migrated to state capitals, with bills advancing in Albany and Sacramento that would impose compute caps and energy-disclosure requirements on the largest AI developers operating within their borders.

The New York State Assembly’s Science and Technology Committee voted 14-9 on Thursday to send the Empire State AI Accountability Act to the floor, while a parallel measure in California’s Senate Appropriations Committee cleared a key procedural hurdle the same afternoon. Both bills would require any company training a model above a set compute threshold to register the run with state regulators, submit independent energy-use audits and accept a 12-month pause on training cycles exceeding the cap.

“The House made a choice on Tuesday. New York is making a different one,” said state Sen. Marisol Reyna, the Brooklyn Democrat who is the lead sponsor of the New York bill, during a press conference in the Capitol. “If Washington can’t decide whether to put guardrails on a technology that is rewriting our electricity grid and our labor market, the states will. We are not waiting.”

The federal bill introduced by Sen. Bernie Sanders, I-Vt., and Rep. Alexandria Ocasio-Cortez, D-N.Y., cleared the Senate 52-48 on April 7 before being killed in House Ways and Means by a 24-21 vote on Tuesday, with a bloc of moderate Democrats joining Republicans to defeat it. The committee’s chair argued the measure would push frontier research offshore. Ocasio-Cortez told reporters after the vote that “the fight does not end here,” and the rapid state-level mobilization appears designed to make good on that pledge.

Industry response has been swift and largely hostile. The Alliance for Responsible Compute, a trade group whose members include the largest U.S. hyperscalers, said in a statement Thursday evening that a patchwork of state rules would “fracture the American AI ecosystem and hand strategic advantage to rivals in Beijing and Abu Dhabi.” Daniel Korver, the alliance’s general counsel, told reporters the group is preparing constitutional challenges on dormant Commerce Clause grounds should either bill become law.

Smaller AI firms have been more divided. Halia Okafor, chief executive of Provenance AI, a mid-sized San Francisco research lab, said her company would back the California bill if its compute thresholds were calibrated to apply only to the largest training runs. “We have spent two years competing against companies that can spend a billion dollars on a single training cycle,” Okafor said in an interview. “Some sort of public accounting of those runs would not hurt us. It might even level the field.”

The bills differ in important particulars. The New York version sets its threshold at training compute equivalent to 10 to the 26th power floating-point operations, the same figure used in the failed federal measure. The California bill uses a sliding scale tied to the average power draw of the data center hosting the run, an approach designed to capture the grid impact rather than the raw computational output. Both define the regulated activity narrowly enough to exclude inference, fine-tuning of smaller models and academic research below specified thresholds.

Energy concerns have been a recurring theme in the debate. Data-center load growth in Virginia, Texas and the Pacific Northwest has strained regional grids, and utility regulators in three states issued warnings during the first quarter that interconnection queues for new AI facilities had grown longer than the planning horizon for new generation. A report released Wednesday by the Rocky Mountain Institute estimated that frontier-model training runs accounted for roughly 1.8 percent of total U.S. electricity consumption in the first quarter of 2026, up from 0.6 percent a year earlier.

“The energy story is what’s giving these state bills oxygen,” said Priya Venkataraman, a technology-policy analyst at the Center for American Progress. “It is much harder for industry to say ‘trust us’ when the local utility is telling ratepayers their bills are going up because a hyperscaler moved in down the road.”

The White House has so far declined to take a position on the state efforts. A spokesperson for the National Economic Council said Thursday that the administration “respects the right of states to legislate within their authority” but added that the president “continues to believe that federal leadership on AI policy is preferable to a fragmented landscape.” The remark was widely read in tech circles as a signal that the administration would not actively oppose the state bills but would not champion them either.

Governors in both states have signaled openness. New York Gov. Catherine Hwang told reporters in Albany that she would review the bill carefully but described its energy-disclosure provisions as “common-sense oversight.” California Gov. Adam Reyes has not committed publicly, though aides have said he is unlikely to veto a measure that clears both chambers with a comfortable margin.

Sanders, asked Thursday about the state-level revival, said it vindicated the original bill’s premise. “When the public sees what is at stake — the grid, the jobs, the basic question of whether a handful of companies should decide the pace of this transformation — they will demand action,” he said. “Tuesday was a setback. It was not the end.”

Floor votes in Albany and Sacramento are expected before the end of May. Several other states, including Washington, Illinois and Massachusetts, are reportedly drafting similar measures, with lawmakers there citing the New York and California bills as templates.