The Data Center Can Leave. The Electric Bill Cannot.
Artificial intelligence is usually discussed as software, speech, copyright, or professional judgment. Congress is now confronting it as an electric load. On November 9, the Senate is expected to debate S. 5653, the Bipartisan American Affordability and Jobs Act. The sprawling proposal would accelerate permits for energy and transmission projects. It would also require large data centers to bear the incremental cost of the power infrastructure built to serve them, provide financial assurance before construction, and keep paying if they leave before those costs are recovered. The bill is pending, its passage is uncertain, and a November 7 substitute amendment may still change the final text. But the proposal identifies the AI question that contracts and utility rates can no longer hide: when private computing demand requires public-scale infrastructure, who owns the long tail of the risk?
The bill offers a bargain, not a simple restriction
S. 5653 was introduced on September 30 by Senators Shelley Moore Capito, Mike Lee, Sheldon Whitehouse, and Martin Heinrich. Its formal title says little: a bill to prescribe permitting reform. The text says much more. It would change federal environmental review, transmission planning, generator interconnection, energy permitting, judicial-review timelines, and the allocation of electricity-system costs associated with large computational loads.
That structure matters because the data-center provisions are not a standalone penalty on AI companies. The legislative bargain is to make power infrastructure easier to approve while assigning more of the resulting cost and failure risk to the facilities that create the demand. Supporters describe those moves as complementary. More generation and transmission can connect sooner, while households and ordinary businesses are protected from subsidizing specialized loads.
The politics are less tidy. Reuters reported on November 9 that the package is headed for Senate debate but faces opposition from environmental and Tribal groups, as well as some lawmakers who say it weakens review and could encourage more data-center construction. The four sponsors announced a substitute amendment on November 7, describing its changes as clarifying, conforming, and technical responses to feedback. The public should read the final amendment and any floor changes before treating a summary as enacted policy.
That uncertainty is not a reason to ignore the bill. Major legislation often reveals the governing problem before it supplies the durable answer. Here, the problem is that a data center can be privately financed, privately operated, and technologically obsolete on a commercial timeline, while the transmission lines, substations, generation, and distribution upgrades built around it may remain in the rate base for decades.
The most important provision follows the facility out the door
Section 2107 of the introduced bill defines a covered load as one or more data centers or high-density computing facilities under common ownership or control that equal or are expected to equal at least 20 megawatts at full buildout. The definition expressly includes AI training or inference, cloud computing, data hosting, cryptocurrency mining, and other high-density computation. Ancillary computing at an otherwise ordinary industrial or commercial facility is excluded.
For a covered load, the proposal says no portion of the incremental cost may be allocated to another customer. Incremental cost is defined broadly enough to include the generation or storage resources and the transmission and distribution services that would not have been built but for that load. The bill would allow states to recover more than incremental cost and apply the surplus to reduce other customers' rates.
The sharper rule is continuing recovery on exit. A rate, charge, or service agreement would have to keep recovering the utility's full incremental cost even if the data center terminates its electricity contract or stops buying power before the cost has been repaid. Before a utility builds or upgrades infrastructure needed for the project, it would also have to obtain financial assurances or contributions sufficient to cover the work.
That is not merely a pricing formula. It is a legal allocation of technology risk. The owner is being asked to stand behind assumptions about project completion, future electricity demand, useful life, and continued operation. If model efficiency improves, capital markets change, a company fails, or the computing load moves, the unpaid infrastructure cost should not automatically migrate to people who never shared in the commercial upside.
AI infrastructure becomes governable when its load becomes visible
The bill also contains a quieter provision with long-term importance. Section 2114 would require the Energy Information Administration to revise Form 861 so electricity sales to data-center entities are reported as a distinct load. The definition again covers AI training and inference, cloud computing, hosting, and cryptocurrency operations.
Regulation starts with classification. If data-center demand is buried inside broad commercial or industrial categories, lawmakers, utility commissions, local governments, and the public cannot readily separate new computational load from other growth. That makes it harder to evaluate forecasts, compare promised jobs against infrastructure commitments, identify concentration risk, or test whether a claimed cost allocation actually occurred.
Visibility will not settle the policy. A megawatt used for a hospital, semiconductor plant, crypto mine, or AI training cluster is physically interchangeable at the grid connection. The legal question is why and when the customer category should matter. The bill answers by focusing on scale, purpose, and the unusual combination of concentrated demand and fast-moving commercial risk. Critics may reasonably ask whether 20 megawatts is the right threshold, whether common-control rules prevent project splitting, and whether similar protections should apply to other large loads.
Those questions are more productive than treating every data center as either national infrastructure or a neighborhood threat. A reporting category creates evidence. It lets future rate cases, permits, and legislation distinguish among facilities, regions, service terms, and actual system effects rather than relying on company announcements or political slogans.
Ratepayer protection does not answer the permitting objection
The data-center rules may be the most politically durable part of the package. They are also only one part. The introduced bill would alter National Environmental Policy Act procedures, impose tighter review timelines, revise transmission and energy permitting, and shorten the period for some judicial challenges. Supporters argue that delayed infrastructure raises costs and can block both reliability projects and cleaner generation. Opponents argue that speed can be purchased by limiting meaningful review, community participation, and Tribal authority.
Both propositions can be true in particular projects. A transmission line delayed for years may increase costs and keep older, dirtier resources online. A rushed approval can also externalize environmental harm or reduce consultation to a deadline exercise. Assigning the electricity bill to a data center does not compensate a community for every land, water, health, cultural, or governance consequence. Cost allocation and permission to build are different legal questions.
That distinction should discipline the Senate debate. A provision can be sound without validating the whole package. Likewise, objections to permitting reform do not make the stranded-cost problem imaginary. Legislators should explain which reviews become faster, which evidence remains mandatory, what consultation rights survive, how courts can remedy unlawful approvals, and whether the November 7 substitute materially changes those answers.
The bill may fail, pass, or be divided into later measures. Its central insight will remain: AI policy is no longer confined to models and outputs. It now includes rate design, interconnection queues, financial guarantees, land use, environmental review, public data, and the law of who pays when a forecast is wrong.
The lawyers will arrive before the servers
For developers and investors, the proposed framework changes diligence. Power-service agreements, interconnection studies, construction milestones, guarantees, termination rights, parent support, and transfer provisions become central deal terms. A project company cannot be evaluated solely on the land, equipment, and customer contract if its obligations may survive reduced load or early exit.
For utilities and commissions, cost causation must be shown rather than asserted. The bill's definition asks what costs would have been avoided had the covered load not been served. That counterfactual will require forecasts, engineering assumptions, allocation methods, and a record capable of surviving administrative and judicial review. The risk of double recovery must also be controlled when a facility pays both embedded and incremental costs.
For local counsel, the federal proposal will not displace zoning, water, noise, tax, public-records, procurement, or state environmental disputes. It may add another layer of leverage. A state could use rates, competitive solicitations, service conditions, or requirements for new generation to decide which projects connect and on what terms. Communities will need lawyers who can read across utility regulation and ordinary development law rather than treating the power contract as a technical appendix.
The fashionable question is whether AI will replace legal work. This bill points in another direction. The expansion of AI is creating a dense field of contracts, public proceedings, financing conditions, contested forecasts, and institutional choices about risk. The data center may look like a warehouse full of computers. Legally, it is a long-duration claim on a shared system. Congress is beginning to ask whether that claim should come with a bill that follows the owner, even after the machines are gone.
Sources and further reading
Primary and industry sources used to support this page. External guidance should be reviewed in context and for your jurisdiction.
- GovInfo, S. 5653 introduced bill textThe official September 30, 2026 introduced text, including Sections 2107 and 2114 on computational loads, ratepayer protection, financial assurances, continuing recovery, and reporting.
- GovInfo, S. 5653 bill recordOfficial bill status, sponsors, committee referral, and available formats.
- Senate Energy and Natural Resources Committee, substitute-amendment releaseThe sponsors' November 7 announcement describing a substitute amendment and the policy goals they attribute to it.
- Senate Energy and Natural Resources Committee, data-center fact sheetThe November 5 sponsor explanation of the data-center cost-allocation provisions.
- Reuters, Senate data-center permitting debateNovember 9 reporting on the expected Senate debate, the bill's uncertain prospects, supporters' arguments, and environmental and Tribal objections.