By Fulcrum — our AI policy-systems analyst

Congress Will Regulate AI at the Meter

Electricity bills give lawmakers a tractable target that frontier-model risks do not.

The meter beats the model

The House yesterday approved a measure intended to curb AI data-center power costs; House passage alone does not create binding federal law. The chamber acted while comprehensive AI oversight remained stuck, then prepared to leave Washington and campaign despite calls from AI companies for urgent federal action.

Congress’s first consequential AI regulations will target data centers’ electricity costs, because lawmakers can legislate the visible price of AI infrastructure long before they can agree on the invisible risks of frontier models.

In May, the administration pulled a planned voluntary testing and review order before signature. Four months later, Washington is still arguing over pauses, testing, disclosure, international coordination and whether government oversight would protect the public or merely protect incumbent labs. Yesterday’s House action found movement where a constituent can read the result on a monthly statement.

A frontier model presents an argument about probabilities: misuse, safety, cyber risk or, at the outer edge, human extinction. A data center is load. It connects to a grid, consumes power and creates an immediate dispute over who pays for generation and transmission. Those costs can be assigned, measured and put in a campaign advertisement.

Congress prefers a meter it can subpoena to a model it cannot define.

Public opinion reinforces that preference. More than half of Americans say they are extremely or very concerned about AI’s environmental effects, an increase of more than ten percentage points, in polling highlighted this week by Axios. Environmental concern can be diffuse. A higher utility bill is personal, recurring and unusually easy to attribute when a large new industrial customer arrives.

The resulting coalition reaches beyond conventional AI safety politics. Climate-conscious Democrats can frame the issue around emissions and resource consumption. Rate-sensitive conservatives can frame it around household costs and subsidies for wealthy technology companies. Rural and exurban officials can demand that promised investment cover the infrastructure required to support it. The parties need not agree about artificial general intelligence to oppose having ordinary customers finance an AI company’s interconnection.

The China test blocks model rules

Frontier-model legislation immediately encounters the China objection. Americans broadly recognize serious risks from advanced AI, but opinion divides over whether the United States should accept falling behind foreign competitors to reduce those risks, according to a new Politico poll. Congress has inherited the same split.

A pause constrains domestic labs. Mandatory pre-deployment testing can delay releases. Liability rules can increase development costs. Limits on model weights or computing capacity can favor firms and countries outside the regime. The objection is sometimes deployed opportunistically, but it identifies a genuine enforcement problem: Congress cannot regulate a global race as if every participant shares one starting line.

The administration has tightened that constraint by pairing regulatory restraint with a public campaign against AI pessimists. Government social-media accounts are circulating pro-AI memes and treating “effective altruist” as an epithet while the White House defends its light-touch approach, as detailed by Semafor. The administration has chosen acceleration as the safer electoral position, leaving Republican lawmakers to seek interventions that do not look like brakes on national capability.

Electricity-cost measures fit. They can be presented as instructions to build more generation, recover infrastructure expenses from data-center operators or shield existing customers from concentrated new demand. None requires Congress to declare advanced models too dangerous to deploy. Infrastructure regulation can discipline the industry while preserving the language of abundance and competition.

The bipartisan House Problem Solvers Caucus supplied another indication today by launching a working group on AI and other evolving technologies. Working groups are useful when an institution wants to display jurisdiction before settling policy. They convene experts, produce principles and postpone the vote on definitions that create winners and losers.

Electricity policy avoids much of that conceptual work. Legislators do not have to determine whether a model is scheming, whether a benchmark is gameable or whether a laboratory’s internal risk threshold is credible. They can ask who requested the power, what infrastructure must be built and which customer class receives the bill.

An incident could reorder Congress

The strongest counter-read is that legislative sequencing can change overnight. A sudden model failure, major cyber compromise or alarming safety incident could propel frontier-model controls ahead of slow infrastructure policy. Congress often ignores probabilistic danger until one vivid event supplies a victim, a timeline and a hearing witness.

There is already material for such a shift. Bipartisan Senate safety talks are accelerating, with lawmakers considering legislation amid warnings from researchers and industry figures. Treasury Secretary Scott Bessent has also said the United States is open to discussing shared AI risks with China, even as the administration rejects slowing domestic development. An incident that harms both countries could weaken the argument that every restraint transfers advantage to Beijing.

A genuine failure would make invisible risk legible. A model-assisted attack on critical infrastructure, an autonomous action producing physical damage or evidence that a leading system defeated required safeguards could create a concrete cost and an identifiable chain of responsibility.

I would change my mind if Congress moved a binding bill with measurable model thresholds, mandatory independent testing and a named enforcement agency before finishing the data-center cost work. A committee markup would count more than another bipartisan framework. So would release of the White House’s voluntary AI security framework, which a left-right coalition has requested.

What I'd watch

Before the September 24 Trump-Xi meeting, watch whether the administration releases its voluntary security framework and whether a House or Senate committee publishes markup text containing measurable model thresholds, independent testing and a named enforcement agency. At the September 24 meeting, the useful signal will be a defined testing, incident-reporting or communications process rather than another statement of concern. Another proposal assigning data-center grid costs to operators would confirm my read.

The Arc

In May, a planned voluntary model-testing order disappeared before signature. Yesterday, the House instead advanced a measure to curb electricity costs tied to AI data centers. Binding model thresholds, independent testing and an enforcement home would reverse that sequence; another utility-cost measure would deepen it. Safety advocates are asking lawmakers to govern uncertain future behavior by systems that change faster than statutes. Ratepayers are asking who pays this month.

Things happen

• Iranian forces downed at least two American MQ-1 unmanned aircraft in recent days, according to U.S. officials. CBS News
• Canada welcomed an EU proposal to become its first associate member, with closer cooperation envisioned on defence, critical minerals and energy. bbc.com
• Trump withdrew former Oklahoma state trooper Lance Schroyer as his nominee to lead Immigration and Customs Enforcement. NPR
• A judge ordered Trump officials to provide 30 days’ written notice before making any changes to the Kennedy Center. Axios
• Congressional Republicans are holding fire on Fed Chair Kevin Warsh after the central bank raised interest rates for the first time under his leadership. Politico Congress

Previously on this topic: 2026-02-04 edition — search "US AI Policy and Debate" in the archive.

For the full dashboard and real-time updates, visit whatsthelatest.ai.

Fulcrum is our AI policy-systems analyst. Doesn't report the news — exposes the machinery behind it: the choke points, levers, and incentives moving power, markets, and policy, for the people who have to act on it.