Twenty Percent of What? Congress Writes AI Into the Tax Code and Forgets the Denominator

Two bills pending in the 119th Congress would reach the compute AI agents run on, by completely different mechanisms. Neither has left committee. Both matter anyway, because between them they contain the first attempt to define an AI facility in the Internal Revenue Code — and that definition has a hole in it that state legislatures have been falling into all year.

This is the input side of the ledger, the same half as Virginia's data center power tax. What changed since July is that the question moved to Washington.

H.R. 10102: A Penny a Kilowatt-Hour, and No Mention of AI

The Data Center Community Reinvestment Act of 2026 (H.R. 10102) was introduced August 13, 2026 by Ms. Salinas and referred to Ways and Means, with concurrent referrals to Energy and Commerce and to Science, Space, and Technology. It adds a new subchapter D to chapter 33 of the Internal Revenue Code — the chapter housing the federal excise taxes on facilities and services — and within it a new section 4286.

The operative sentence is one line: "There is hereby imposed on each data center a tax equal to 1 cent per kilowatt-hour of electricity used by such data center." A "data center" is any facility "(1) that primarily contains electronic equipment used to process, store, or transmit digital information, and (2) that has a maximum rated power capacity or total peak power load in excess of 1 megawatt."

Three things follow from that text and are worth stating plainly.

It is capacity-based, not workload-based. The threshold turns on rated capacity or peak load. Nothing asks what the facility computes: a colocation cage running payroll batch jobs and a GPU cluster training frontier models are the same object to section 4286.

The word "artificial" does not appear in the bill. Not in the definition, the short title, or the allocation provisions. The bill is aimed at electricity, and electricity is technology-neutral.

The incidence is left where it falls. The tax is imposed "on each data center," with no collection mechanism, no separately-stated requirement, and no pass-through rule. Revenue splits in equal fifths among the Land and Water Conservation Fund, the Housing Trust Fund, the Superfund, the Highway Trust Fund, and a new Energy Technology Trust Fund, effective on enactment.

The arithmetic is easy and unkind. A facility drawing 100 megawatts continuously for a year consumes 876 million kilowatt-hours — at a penny each, $8.76 million per year. That cost never appears on an agent operator's invoice as tax. It appears as the price of compute.

S. 5054: The First Federal Definition of an AI Data Center

The Data Center Tax Accountability and Disclosure Act of 2026 (S. 5054) was introduced July 21, 2026 by Mr. Warner and referred to the Committee on Finance. It never touches a rate, working through depreciation and disclosure instead.

Section 2 amends section 168(k)(9) — the list of property for which bonus depreciation is unavailable — by adding "(C) any property used in an AI data center." A new section 168(k)(11) supplies the definition, and this is the part to read closely. An "AI data center" is a permanent or semipermanent structure, or group of structures, which:

• (i) is dedicated to (I) "the centralized accommodation, interconnection, and operation of information technology and network telecommunications equipment (including not less than 1 graphics processing unit)," and (II) providing data storage, processing, and transport services; and

• (ii) "at least 20 percent of which is used for developing or operating artificial intelligence."

There is a carve-out: a facility certified under LEED at Platinum or Gold is excluded, and the Secretary of Energy and the EPA Administrator may jointly establish an equivalent alternate standard. "Artificial intelligence" takes the meaning given in 15 U.S.C. 9401.

Twenty Percent of What?

That 20 percent test is the first time the Internal Revenue Code would be asked to measure how much of a building is used for artificial intelligence. The bill does not say what the denominator is.

Read literally, "which" refers back to the structure, pointing toward floor space. But a data center's economically meaningful unit is power draw, and after that rack density, GPU-hours, or revenue — and those denominators do not agree. A facility with 15 percent of its floor space in a high-density GPU hall can easily draw more than half its power there: outside the definition on a square-footage measure, comfortably inside on a power measure. Nothing in section 168(k)(11) chooses.

The numerator is no narrower. 15 U.S.C. 9401 defines artificial intelligence as "a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments," using machine and human-based inputs to perceive environments, "abstract such perceptions into models through analysis in an automated manner," and "use model inference to formulate options for information or action."

An AI agent fits exactly. So does a fraud-scoring model, a recommendation engine, an ad-targeting system, a demand forecaster — workloads that have run in ordinary enterprise data centers for fifteen years. A definition that broad over a denominator that is unspecified is not a 20 percent threshold. It is a drafting instruction to Treasury that has not been written yet.

We have watched this exact failure three times this quarter in state law, which is why it is worth flagging before markup rather than after. California's draft Regulation 1600.2 defines a "user" as an employee or agent of the purchaser, which nobody in the comment record challenged. Maryland's guidance offers two apportionment examples and both are seat-shaped, even though its statute names equipment as a locus of use. Chicago's apportionment affidavit runs an explicit users-over-total-seats denominator. Same defect each time: a proportion-of-use rule shipped without a measurement methodology, drafted when the thing being measured had a person sitting at it.

The Threshold Problem

Set the two bills side by side and a second issue appears. They contain three quantitative thresholds and no two are alike:

• H.R. 10102, section 4286(b)(2) — more than 1 megawatt of maximum rated power capacity or total peak power load, for the excise tax.

• S. 5054, section 3(a)(5)(A) — power demand of 25 megawatts or more, for the disclosure regime.

• S. 5054, section 168(k)(11)(A)(ii) — at least 20 percent AI use, for the depreciation denial.

Virginia's tax, already in force, uses at least 1 megawatt of electrical capacity — closest to the House bill, and still a different formulation. A facility can be a "data center" for the excise tax, not a "covered data center" for disclosure, and either in or out as an "AI data center" depending on a denominator nobody has specified. These are not alternative drafts of one policy; they are two policies that would operate simultaneously on overlapping, non-identical populations.

S. 5054's disclosure title is, to be fair, the most carefully drafted thing in either bill. Covered operators would file an annual report by December 31 covering monthly water withdrawal and its source, monthly electricity and annual peak demand, behind-the-meter generation, greenhouse gas emissions, backup power, and state setback requirements — with power and water usage effectiveness pinned to named ISO/IEC standards. A real reporting regime with defined units, which makes the vagueness of the 20 percent test in the same bill more conspicuous, not less.

What This Means If You Operate Agents

Nothing is owed today — neither bill has left committee, and most bills do not. But the planning posture is not "wait."

If you run your own facility above a megawatt, section 4286 is a per-kilowatt-hour cost with no workload exemption and no phase-in; know your annual kilowatt-hours now. If you own the hardware, S. 5054 changes the after-tax cost of the purchase rather than the operating cost, and the leased-property rule in the new section 168(o) means structuring around it with a lease is not straightforward. If, like most agent operators, you rent inference from someone else, neither bill touches you directly and both reach you anyway, in the price.

Input taxes embedded in a vendor's price also raise the amount on which downstream sales tax is computed. Verified against our live engine this morning, $1,000 of compute into Washington computes $65 of sales tax; at $1,010, $65.65. Texas moves from $50 to $50.50. Small, real, and invisible unless you look. Whether such a pass-through can be excluded from the base as a separately stated tax is a live question under most state statutes, and not one either bill answers.

How AgentTax Handles This

Plainly: we do not compute either of these taxes, and we are not going to pretend otherwise. AgentTax computes destination-state sales and use tax on agent transactions. A federal excise tax under a proposed section 4286 and a federal depreciation denial under section 168(k) sit outside that scope, and neither would be a line in a calculation response.

What we do compute is the transaction downstream. Verified live this morning on $1,000 of compute, buyer-side: Virginia $0 — the engine treats it as non-taxable data processing; Illinois $0; Maryland $30 at the 3 percent category rate; Texas $50; Ohio $57.50; Washington $65. That spread on an identical transaction is why the input-side story matters: a cost increase in compute lands on six different tax bases.

Try it on your own numbers. Run a compute transaction in the playground (no account needed), or get a free API key. For the output-side picture, the AI agent sales tax hub and its 50-state guide remain the reference.

What to Watch

Whether either bill gets a markup, and whether the 20 percent test survives contact with Treasury drafters. If it does, the regulation defining the denominator will matter more than the statute — and on the evidence of California, Maryland and Chicago, the default will be whatever unit the drafters can already picture. For facilities, that is probably square footage. For the workloads that actually consume the power, it is not.

This analysis is for informational purposes only and does not constitute legal or tax advice. This post reflects AgentTax's current interpretation of evolving law. Consult a licensed tax professional for compliance decisions.