Hyperscaler CEOs flood Capitol Hill as AI moratorium fight enters final week
4 min read, word count: 916The chief executives of the four largest US cloud providers met privately with members of the House Ways and Means Committee on Wednesday, intensifying a lobbying blitz aimed at killing a Senate-passed moratorium on new frontier-scale AI training runs before a committee vote expected as early as Apr. 22.
The closed-door sessions, held in three separate suites in the Longworth and Rayburn buildings, drew the chiefs of Microsoft, Alphabet, Amazon Web Services and Meta into back-to-back meetings with a rotating group of fourteen members, according to two industry officials and a House staffer familiar with the schedule. Anthropic chief executive Dario Amodei and OpenAI’s Sam Altman appeared on Capitol Hill in a separate track of meetings the same day.
The legislation, passed by the Senate 52-48 on Apr. 7, would impose an 18-month pause on training runs above a defined compute threshold and require federal licensing for any datacenter buildout exceeding 500 megawatts of contracted load. Sponsors Sen. Bernie Sanders, I-Vt., and Rep. Alexandria Ocasio-Cortez, D-N.Y., have framed the measure as a response to surging power demand, soaring residential electric rates in datacenter corridors, and what they describe as an unaccountable race toward systems whose behavior even their builders cannot fully predict.
Industry executives argue the bill would freeze the United States in place while Chinese and Emirati developers continue scaling unimpeded. “A pause is not a pause. It is a handover,” said Marcus Lindgren, president of the American Compute Alliance, a trade group formed last fall to coordinate hyperscaler advocacy. “Eighteen months in this field is the difference between leading and being a footnote.”
Lindgren’s group has spent more than $42 million on advertising in 27 congressional districts since the Senate vote, according to disclosures reviewed by AdImpact, with the heaviest concentrations in northern Virginia, central Ohio, suburban Phoenix and the Texas Triangle — regions where datacenter construction has become a major source of construction jobs and local tax revenue.
The bill’s path through Ways and Means is uncertain. The committee splits 24-21 in favor of the majority, but at least four Democrats on the panel have publicly questioned the moratorium’s structure, and two Republicans have voiced support for portions of its grid-protection language. Chairman Drew Halverson, R-Tex., told reporters Wednesday afternoon that he expected a “vigorous markup” but declined to predict an outcome.
Sanders, in a statement issued from his Senate office, accused the cloud providers of “lobbying their way around the will of a democratic majority” and pointed to a study released this week by the Energy Information Administration projecting that datacenter electricity consumption will reach 14.2 percent of total US generation by 2028 if current buildout continues. Average residential bills in five datacenter-heavy states rose between 18 and 31 percent over the past year.
Ocasio-Cortez, appearing on a podcast Tuesday evening, said the House had “a narrow window to draw a line” before the next generation of training runs locks in compute footprints that “no future Congress will be able to claw back.”
Opposition has not been monolithic on the industry side. Smaller AI firms, including several open-weights developers, have endorsed a narrower compromise that would preserve the licensing regime for new gigawatt-scale datacenters but drop the training-run moratorium entirely. “The compute threshold is the part that scares us,” said Priya Venkatesan, chief executive of Helix Labs, a San Francisco-based startup that trains specialized models for biomedical research. “We don’t operate at that scale. The grid provisions we can live with. The cap, we cannot.”
The White House has remained publicly noncommittal. Press Secretary Karoline Leavitt told reporters Tuesday that President Donald Trump was “watching closely” and would “make a decision based on what is best for American workers and American technological leadership.” Two senior administration officials, speaking on condition of anonymity to discuss internal deliberations, said the West Wing was divided, with economic advisers warning against signing a measure they believe would invite retaliation from major employers and national-security aides arguing the licensing provisions would give the federal government useful leverage over critical infrastructure.
The fight has overlapped uncomfortably with the war in the Persian Gulf, which entered a ceasefire at midnight GMT. Several lawmakers said the cessation of hostilities had reduced the political appetite for what one called “another big, divisive vote” and could push moderate Democrats toward a no.
“Energy security and AI security are the same conversation now,” said Hannah Pruitt, an analyst at the Aspen Strategy Group. “Members who were ready to vote for the moratorium when oil was at $125 a barrel are getting cold feet at $98. The macro story has flipped under their feet in two weeks.”
State capitals are not waiting for Washington. The New York Assembly advanced its own moratorium bill out of committee on Monday, and the California Public Utilities Commission opened a formal proceeding Tuesday on whether to impose interconnection caps on datacenter loads above 250 megawatts. Texas legislators have introduced a contrary measure that would preempt local restrictions and offer a ten-year tax abatement for new datacenter sites.
Markup is scheduled to begin Monday morning. House aides said the chairman had not yet decided whether to hold a single combined vote or to break the bill into severable provisions, a procedural choice that several members on both sides described as potentially decisive.
Sanders, asked Wednesday evening whether he believed the bill would survive, paused before answering. “We will know in a week,” he said. Advocates on both sides said they expected the lobbying tempo to accelerate through the weekend.
Note: This article was partially constructed using data from LLM.