WASHINGTON — The Treasury Department on Thursday proposed federal rulemaking that would require hyperscaler operators of AI training clusters above defined size thresholds to file grid-impact disclosures and quarterly reporting through an interagency channel anchored at the Federal Energy Regulatory Commission, in what officials described as the most significant federal step yet toward putting infrastructure conditions on the largest tier of AI compute facilities.

The notice of proposed rulemaking, published in the Federal Register at 10 a.m. Eastern Thursday, would apply to training clusters with a nameplate accelerator population above 25,000 chips of identified leading-edge classifications, or with a steady-state grid draw above 200 megawatts. Facilities meeting either threshold would be required to file a pre-construction grid-impact assessment with FERC, a quarterly load-and-utilization report with the Department of Energy, and a curtailment-cooperation attestation with the relevant grid operator.

The proposal arrives less than a week after the five largest U.S. AI developers signed the May 8 voluntary disclosure framework that established a 60-day pre-deployment notification window and an industry-funded Frontier Model Assurance Council. Treasury officials framed Thursday’s proposal as “the infrastructure-side complement” to the voluntary commitments, addressing the physical buildout of training capacity rather than the disclosure of resulting model capabilities.

“What the May 8 framework did was create a discipline around model release,” said Treasury Under Secretary for Industrial Strategy Linda Chen in a Thursday-morning briefing at the department’s headquarters. “What this proposed rulemaking does is create a discipline around the buildout that produces the models. They are different questions, and they require different tools.”

The rulemaking is structured under existing Federal Power Act authorities, with cross-references to the Defense Production Act and to the Investment Company Act. Treasury and FERC, in a joint statement issued Thursday, said they had concluded that existing statutory authorities were sufficient to support the proposed disclosure regime and that additional congressional authorization would not be required for the rule to be finalized. Several legal analysts said Thursday afternoon that the statutory framing was likely to face legal challenge from at least one of the affected parties.

The proposal’s thresholds were set at levels intended to capture the next-generation hyperscaler training facilities while leaving conventional cloud-region buildout and most enterprise data-center operation outside its scope. The 25,000-chip threshold maps roughly to the cluster size used for the most capable currently deployed models. The 200-megawatt threshold is consistent with the threshold that California’s narrowed data-center moratorium bill adopted Wednesday in the state Senate energy committee. Several Treasury officials acknowledged in the briefing that the California threshold’s emergence had influenced the federal calibration.

The proposed quarterly load-and-utilization report would be the most detailed federal disclosure framework yet imposed on AI compute facilities. Required fields include: monthly peak load, monthly average load, equivalent capacity factor, water consumption (where applicable), curtailment events of more than fifteen minutes, contracted firm versus interruptible power volumes, and an attestation of cooperation with the relevant grid operator’s emergency-mode protocols. The information would be filed quarterly through a secure FERC submission channel, with public aggregates published with a six-month lag.

The proposal would also impose what Treasury characterized as “modest” cooperation obligations. Facility operators would be required to commit to curtailment of non-critical workload during grid-emergency events declared by the relevant balancing authority, with an exemption for inference traffic deemed time-sensitive by the facility’s principal customers. Treasury officials said the cooperation framework had been the subject of “extensive” pre-publication consultation with the affected parties.

Industry response has been mixed. Trade associations representing the hyperscaler community indicated through prepared statements that the proposal’s threshold and disclosure framework were “in the direction of workable” but expressed significant concerns about the proposed curtailment-cooperation provisions. The Information Technology Industry Council, in a statement issued Thursday morning, said the curtailment provisions would require “careful calibration to ensure that they do not impose disproportionate burdens on customer-facing services or on facilities serving national-security functions.”

A senior official at one of the four hyperscalers most directly affected, contacted Thursday afternoon, said the company was “supportive of the disclosure framework” but had specific concerns about the proposal’s treatment of multi-tenant facilities, in which a single physical building may host workloads from several different operators. The official said the company would engage in the formal comment process and was hopeful that final rule language would address the multi-tenant question.

Congressional reaction tracked predictable lines. Senators Maggie Hennessey, D-Colo., and Marsha Blackburn, R-Tenn., who introduced the bipartisan AI Transparency and Grid Impact Act last week, issued a joint statement Thursday afternoon welcoming the Treasury proposal as “a useful step that complements rather than substitutes for the statutory framework the Senate is preparing.” The Hennessey-Blackburn bill, which has gathered seventeen cosponsors since its introduction, would create a statutory disclosure framework with binding penalties for noncompliance.

The Republican leadership in the House has been more cautious. House Energy and Commerce Chairman Brett Guthrie, R-Ky., in a statement issued Thursday afternoon, said the committee would “carefully examine” the Treasury proposal and that the rulemaking should not “preempt or substitute for” the broader legislative work the committee had begun on AI infrastructure questions. Guthrie’s statement did not commit to specific legislative action.

State-level officials have responded with broad support. California Governor Gavin Newsom, in a statement issued Thursday afternoon, said the Treasury proposal “demonstrates that the questions California has been working on in Sacramento are recognized at the federal level” and indicated that the state’s own data-center moratorium bill would be “harmonized” with the federal framework if both reach the implementation phase. New York’s Department of Public Service issued a similar statement.

The formal comment period for the Treasury proposal opens Thursday and runs through July 28. Treasury officials said they expected to issue a final rule by the end of the fourth quarter, with the rule taking effect ninety days after publication in the Federal Register if not subject to a successful legal challenge.

A senior Treasury official, asked at the morning briefing about the prospect of legal challenge, said the department had “designed the rulemaking with a clear view of the statutory authorities being relied upon” and was prepared to defend the rule’s substantive provisions if necessary. The official declined to speculate about specific litigants.

The proposal’s appearance Thursday morning was the third of three federal-level actions on AI infrastructure in the past two weeks. The May 8 voluntary disclosure framework, the introduction of the Hennessey-Blackburn legislation, and Thursday’s Treasury rulemaking together constitute the most consolidated federal posture on AI infrastructure questions since the regulatory cycle that began with the original executive order on artificial intelligence in 2023.