New Congressional Data Center Tax Proposals Threaten US AI Investment

Introduction

US companies are projected to invest trillions of dollars in artificial intelligence (AI) over the next several years. Goldman Sachs recently estimated that capital expenditures from just a few US AI “hyperscalers” will reach $581 billion in 2026 alone. Some, though not all, of this investment will go toward the buildout of data centers—physical structures that store and process vast swaths of information being exchanged between digital platforms and users.

Members of Congress concerned about the land and energy use of data centers, as well as the labor market effects of AI, have introduced proposals to taxA tax is a mandatory payment or charge collected by local, state, and national governments from individuals or businesses to cover the costs of general government services, goods, and activities. new data centers. While many members have proposed changes, two recent ideas, one from Sen. Mark Warner (D-VA) and another from Senate Finance Ranking Member Ron Wyden (D-OR) would deny certain AI data centers the use of full expensing for machinery and equipment.

Doing so would add significant complexity to the tax code and draw arbitrary lines around the types of companies and investments that can fully recover their investment costs. It may also drive AI investment overseas, denying US towns and cities the jobs, economic growth, and tax revenue that flow from investment in local communities.

Background on the Two Proposals

In July 2026, Warner introduced the Data Center Tax Accountability and Disclosure Act. The bill would prevent businesses from claiming full expensingFull expensing allows businesses to immediately deduct the full cost of certain investments in new or improved technology, equipment, or buildings. It alleviates a bias in the tax code and incentivizes companies to invest more, which, in the long run, raises worker productivity, boosts wages, and creates more jobs. under Sec. 168(k) for “any property used in an AI data center.” Warner defines an “AI data center” as one that:

1. Is a structure or group of structures dedicated to information technology (IT) equipment, network telecommunications equipment, and “data storage, processing, and transport services,” and
2. Is used “at least 20 percent” for developing or operating AI.

Businesses could still claim full expensing for their AI data centers if the structures are Leadership in Energy and Environmental Design (LEED)-certified at the Platinum or Gold levels, a measure of energy and environmental efficiency. Tying certain tax items to LEED certification could encourage data center owners to make their operations more energy efficient.

Wyden’s proposal was released in August 2026 as a white paper, and it would deny full expensing to new data centers as well. Specifically, the proposal would:

• Deny additional tax provisions to new data centers, including Opportunity Zone funding and real estate investment trust (REIT) benefits.
• Levy a gross receipts taxGross receipts taxes are applied to a company’s gross sales, without deductions for a firm’s business expenses, like compensation, costs of goods sold, and overhead costs. Unlike a sales tax, a gross receipts tax is assessed on businesses and applies to transactions at every stage of the production process, leading to tax pyramiding. at an undefined rate on data center operators in the US. “Internet infrastructure,” corporate IT departments, “small” data centers, and pre-2024 data centers for all but “the largest actors” would be exempt.

The gross receipts tax would also feature a deemed minimum regime that requires data center users above certain asset and expense thresholds to pay a minimum tax, which Wyden argues is aimed at tax avoidance.

Both the Warner and Wyden proposals leave key definitional questions unanswered, which may make them difficult to administer and comply with absent further regulations from the Treasury Department and the IRS. For example, Warner’s bill does not detail how the 20 percent AI threshold is determined.

Similarly, Wyden does not define the “internet infrastructure” or “small local data center operators” that are exempt from his gross receipts tax. The deemed minimum regime has its own complexities and may cast a wide net in subjecting anyone engaged in “complex financial arrangements [with] a data center” to the tax.

Analyzing the Proposals  

While Wyden proposes a tax in the “low single digits,” a seemingly low tax on gross receipts can in practice be a very high tax on net income. For example, as the figure below illustrates, a gross receipts tax of 4 percent can result in an effective tax rate on net income that is substantially higher, reaching 40 percent or even 200 percent depending on a firm’s expenses.

Neither the Warner nor Wyden proposals capture non-US investment in data centers, which may have the practical effect of driving data center investment overseas. This could be intended, but pushing data center investment abroad could also pull broader AI investment with it, potentially eroding the United States’ currently projected advantage in attracting global AI investment.

One premise underlying each of these proposals is that the current US tax system is not well-positioned for the economic gains that will accrue from AI investment and adoption. That premise has been scrutinized by tax experts across the ideological spectrum. Existing corporate income, capital gains, and property taxes are likely to capture any supernormal returns from AI companies and the data centers they operate, weakening the case for specific taxes on AI.

To the extent AI gains are shared more broadly across the US economy, the federal government can also expect to see higher tax revenue from rising corporate profits, capital gains realizations, and wages (particularly among individuals in higher marginal tax brackets). Switching depreciationDepreciation is a measurement of the “useful life” of a business asset, such as machinery or a factory, to determine the multiyear period over which the cost of that asset can be deducted from taxable income. Instead of allowing businesses to deduct the cost of investments immediately (i.e., full expensing), depreciation requires deductions to be taken over time, reducing their value and disco deduction timing does not raise long-run revenue, but it would discourage new investment into revenue-generating economic activity like data centers.

An extensive body of economics literature finds that the benefits of new technologies are not captured primarily by innovators but rather dispersed in the economy. This finding is often used as justification for favorable tax treatment of research and development, a clear contrast with the proposals from Wyden and Warner to increase taxes on AI.

Estimated Revenue Effects

Using Tax Foundation’s Taxes and Growth (TAG) Model, we centrally estimate that Warner’s proposal to disallow bonus depreciationBonus depreciation allows firms to deduct a larger portion of certain “short-lived” investments in new or improved technology, equipment, or buildings in the first year. Allowing businesses to write off more investments partially alleviates a bias in the tax code and incentivizes companies to invest more, which, in the long run, raises worker productivity, boosts wages, and creates more jobs. for qualified data centers would raise $29.9 billion from 2027 to 2036 on a conventional basis.

The proposal would reduce the long-run size of the economy by less than 0.05 percent, as it increases the cost of capital by disallowing full cost recoveryCost recovery refers to how the tax system permits businesses to recover the cost of investments through depreciation or amortization. Depreciation and amortization deductions affect taxable income, effective tax rates, and investment decisions. for firms investing in certain data centers.

On a dynamic basis, the proposal raises $18.2 billion over 10 years, reflecting reduced income and payroll taxA payroll tax is a tax paid on the wages and salaries of employees to finance social insurance programs like Social Security, Medicare, and unemployment insurance. Payroll taxes are social insurance taxes that comprise 24.8 percent of combined federal, state, and local government revenue, the second largest source of that combined tax revenue. revenue associated with a smaller economy.

Table 1. Revenue Estimates of Bonus Depreciation Disallowance Under the Data Center Tax Accountability and Disclosure Act, 2027-2036, in Billions (Central Estimate)

Source: Tax Foundation General Equilibrium Model, August 2026.

The projected path of data center investment, the proportion of data centers that meet the 20 percent AI criteria, and LEED certification take-up all introduce uncertainty into our estimate. To address the uncertainty, we modeled three scenarios: low, central, and high amounts for the investment basis affected by the proposal.

These scenarios vary by the total basis of investment that would become bonus-ineligible: the low scenario begins at $37.9 billion affected in 2027, the central begins at $65.0 billion, and the high begins at $101.1 billion. These figures are sensitive to what share of data center investment qualifies for LEED certification and the 20 percent AI threshold under the bill (see Table 2).

The lower scenario would raise $17.5 billion in conventional revenue over 10 years, while a higher path would raise $46.5 billion over that time. The corresponding economic harm also increases but remains below 0.05 percent of GDP in the long run.

Importantly, none of the scenarios result in sustained increases in federal tax revenues over the long term. This is because denying bonus expensing is primarily a timing shift—bringing the same tax revenues sooner in time—rather than a long-term increase in tax revenue.

Table 2. Range of Revenue Estimates Under the Data Center Tax Accountability and Disclosure Act, 2027-2036

Source: Tax Foundation General Equilibrium Model, August 2026.

Policymakers should be cautious about using the tax code to capture returns from AI in a way that discourages investment in AI and data centers. Denying cost recovery and imposing new excise taxes would add complexity, distort investment decisions, and risk pushing activity abroad. A neutral tax code that allows businesses to recover their investment costs while taxing resulting profits is better suited to capturing the economic gains from AI without undermining US competitiveness.

Modeling Notes                 

We model Warner’s proposal as removing a share of economy-wide information and communication technology (ICT) investments from being eligible for bonus depreciation. Analysis of BEA detailed fixed-asset data suggests that about 80 percent of data center equipment investment consists of ICT.

To model the trajectory of the relevant investments, our scenarios capture a near-term AI data center investment boom. We assume a 20.5 percent annual growth rate through 2030 in the bonus-ineligible amount under the proposal, followed by a gradual decline back toward the 2027 share of ICT made bonus-ineligible by 2036.

Share this article