U.S. House Passes Bill to Curb AI Power Cost Shifting, Data Center "Front-of-Meter" Incremental Costs Gradually Implemented

The U.S. House of Representatives passed legislation aimed at protecting Americans from bearing the electricity cost of data centers by an overwhelming vote of 417 to 3. The bipartisan Ratepayer Protection Act (H.R. 9340) cleared the chamber under Republican leadership's "suspension of the rules" fast-track proceduretypically reserved for noncontroversial legislation and requiring a two-thirds majorityand has now been sent to the Senate. The vote comes as lawmakers face mounting pressure: the data center construction boom driven by the AI race has been cited in part as a cause of surging energy costs. The bill was reportedly introduced by Colorado Republican Representative Gabe Evans, who is himself in a competitive race, with several Republican lawmakers facing reelection pressure signing on as co-sponsors in recent weeks. What the bill says, and where it stands Its reach is far narrower than the vote tally suggests. The bill amends the Public Utility Regulatory Policies Act of 1978 (PURPA) to provide state regulators with a set of federal reference standards: it encourages states to establish rules ensuring that "large load" users, including large data centers, pay for the cost of new generation rather than shifting those costs onto consumers. According to Evans' office and the House Energy and Commerce Committee, the trigger threshold is a single user with electricity demand of 100 megawatts or more, covering all incremental costs of generation, transmission, and distribution upgrades; large-load users must also provide financial guarantees before utility investments are made, to prevent costs from falling on existing ratepayers if a project shrinks or relocates. The key is in the wordingthe bill requires states to "consider" this standard, does not mandate adoption, and only requires states to hold hearings on it within two years of enactment. It neither prohibits nor restricts the construction of any data center. The legislative process moved quickly: introduced with bipartisan support in June, passed unanimously out of the Energy and Commerce Committee 52-0 on July 21, and cleared under suspension of the rules 417-3 on September 16. It is the first data center-related bill passed by the 119th Congress, and likely one of the last pieces of legislation considered before the November 3 midterm elections. The real hurdle is the Senate. The companion bill S.5028, introduced by Ohio Republican Senator Jon Husted, has not even been scheduled for a committee hearing and has only one co-sponsor, Alabama Republican Senator Tommy Tuberville. Senate Majority Leader John Thune has reportedly said the bill could pass before the midterms, but only via the "unanimous consent" procedurethe Senate's calendar this month is packed, and a single senator's objection would collapse that timeline. Resistance has already emerged: New Mexico Democratic Senator Martin Heinrich called it a "fig leaf," while Senate Minority Leader Chuck Schumer made no commitment, saying only that "nothing gets done unless it's done in a bipartisan way." Impact on tech giants: the real constraint isn't the bill, it's time What does this bill actually mean for Microsoft, Amazon, Google, and Meta? The key conclusion: what it changes is not mainly electricity prices, but the certainty of cost allocation and the timeline for getting powered up. The first layer is "voluntary" becoming a "standard." According to information released by the White House and the EPA, on March 4, 2026, seven AI and hyperscale cloud providersAmazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAIsigned the Ratepayer Protection Pledge at the White House, committing to five obligations: build, bring in, or purchase all the power data centers need and bear its cost; pay for new power delivery infrastructure upgrades; pay agreed rates regardless of whether electricity is actually used; invest in local hiring and training; and cooperate with grid operators to enhance resilience. According to the pledge's official page, signatories have grown to 317 organizations (207 electric cooperatives, 71 utilities, 40 data center developers), with 23 governors also joining, covering 80% of U.S. power supply and involving 263 million Americans. In other words, the top seven had already voluntarily committed to bearing infrastructure costs themselves, and the bill's actual effect is to turn "voluntary" into a "federal reference standard." For those who have already signed, the marginal added cost is limited; the real change falls on third partiescompanies that have not signed the pledge, or that rely on third-party data center capacity leases, will face a rate environment where they could previously negotiate but now have federal text backing the other side. Media have previously questioned that the pledge contains no fines, penalties, or enforcement mechanismretail electricity rates are ultimately decided by state regulators in rate cases; this bill is meant to fill exactly that "teeth" gap. The second layer is more concrete: financial guarantees and "pay whether you use it or not" capacity charges front-load capital expenditure and raise the barrier for latecomers. The bill requires large-load users to provide financial assurance before utility investments, meaning cash outflows come earlier; combined with the pledge's "pay agreed rates regardless of whether electricity is used" clause, a campus not yet fully loaded must keep paying for already-reserved generation and transmission capacity. For cash-rich hyperscalers, this is an affordable certainty cost; for more highly leveraged third-party data center developers and compute leasers, it is a real increase in the barrier to entry. The third layer is decisive: reliable power access is replacing GPUs as the real bottleneck for AI expansion. Goldman Sachs Research estimates U.S. data center power demand will rise from 31 gigawatts in 2025 to 66 gigawatts in 2027, while only about 60% of new capacity may come online on timethe bottleneck is not chips, but interconnection queues and the physical pace of grid expansion. According to PJM (the largest regional grid in the U.S., covering 13 states and about 65 million people) capacity auction data and estimates by the independent market monitor, its capacity clearing price rose about 11-fold in two years to $329.17 per megawatt-day, with about 63% of the increase caused by data center load, equivalent to roughly $9.3 billion. And according to Politico, a memo from the Republican Senate campaign arm in August warned that data centers are "an anchor around Husted's neck." Giants have already acted: from "buying power" to "becoming a grid resource" Rather than waiting for legislation, tech giants' responses began long ago, and the direction is clearturning electricity from an external procurement item into an asset they control. On the very day the bill passed, one deal provided the most direct footnote. According to an 8-K filing Generac submitted to the U.S. Securities and Exchange Commission on September 16, the company signed a transaction agreement with Amazon to issue warrants to an Amazon wholly owned subsidiary, exercisable at $200.9266 per share for up to 1,693,745 shares of common stock (worth about $340 million), with vesting conditions tied to the total amount Amazon pays for backup generatorsup to $8 billion; the two sides also signed a long-term supply agreement, with initial backup generator deliveries expected to total $2.4 billion in 2027 and 2028. Generac shares surged more than 28% in premarket trading Thursday. Notably, what Amazon is buying is backup generators for data centersthat is, behind-the-meter self-built power, precisely the channel that bypasses interconnection queues and capacity upgrade fees, and this "escape route" itself is also driving up unit costs. Another path is trading flexible load for interconnection priority. On the same day, Nvidia, Google, and Emerald AI announced the launch of the AI Energy Management Alliance (AEMA), with 18 founding members spanning both power and compute sides, including Anthropic, Analog Devices, National Grid, AES, Constellation, and NRG. According to principles disclosed by the alliance, its proposition is to let data centers dynamically adjust electricity use based on real-time grid conditionsthrough compute load migration, releasing stored energy, and activating on-site generation, transforming large compute facilities from rigid load into dispatchable grid resources; for entities that can provide verifiable flexible power commitments, open a fast-track approval channel based on risk assessment and allocate interconnection costs according to actual system benefits. The alliance says this could unlock up to 100 gigawatts of capacity from the existing power system and save about $733 million in system costs per gigawatt of new AI data center capacity. Viewed together on the same day, the logic is complete: legislation is addressing "who pays," while industry is addressing "how to get powered faster." For companies like Amazon and Google, the former's costs are bearable; the latter's time is what really counts.