Artificial intelligence (AI) could raise productivity across a wide range of occupations and tasks. But it could also replace workers who currently perform some of those tasks. How many workers AI might replace, and how quickly, remains an open question.
These questions matter for the federal budget because the U.S. tax system relies heavily on income earned by workers. Individual income and payroll taxes generate roughly 85% of federal tax receipts. The individual income tax draws heavily from wages, salaries, and tips and is the single largest source of federal revenue. The payroll tax that funds Social Security and Medicare applies only to wage and self-employment income. If AI shifts income from workers to business profits and other capital income, some of that income could face lower effective tax rates. A sustained shift could erode federal revenues amid growing demands for public spending to assist displaced workers.
In response, lawmakers are proposing AI-specific taxes with a range of stated goals: maintaining tax revenues given the changing economy; raising additional revenues to support displaced workers; addressing the power and water infrastructure investment AI facilities require, and slowing down the pace of AI activity and growth. Several distinct tax designs have been proposed or discussed, each aimed at a different point in the AI production process, from the computing power used to train AI models to the profits they will eventually generate. The proposals differ substantially, however, in timing and scope, potential revenue and economic costs, and ease of administration.
An AI-specific tax is targeted and easy to explain to the public. That is a fair political consideration, but it is not an economic one. Achieving the above goals does not strictly require a tax specific to AI. Reforming how capital income is taxed, introducing a broad-based consumption tax such as a value-added tax (VAT), a carbon tax, or a corporate cash-flow tax could address some of the same concerns.
And federal revenue isn’t the only issue at stake. Several of these ideas sit in tension with existing U.S. international tax and domestic investment strategies. By shaping where AI infrastructure gets built, they touch national security as well.
The following examination of the arguments for taxing AI and five leading proposals highlights their differing objectives, revenue potential, administrative feasibility, and trade-offs, as well as their broader implications for the U.S. tax system.
Why tax AI?
Whether, how, and when AI might reshape the economy are uncertain, and the arguments for taxing AI rest on several possible scenarios rather than a single potential trajectory.1 Three come up most often.
AI will erode the tax base. In this scenario, as AI displaces workers, income and payroll tax will fall immediately, while any resulting offsetting tax revenue from the profits and capital gains would arrive later, if at all. At the same time, demand will rise for the public programs that help displaced workers, including unemployment insurance, retraining, and other safety-net supports. An AI tax could help fund this spending at a time when federal revenues from income and payroll taxes decline.
AI infrastructure imposes costs on people who don’t use it. This scenario reflects a bipartisan concern about who pays for the consumption required by AI infrastructure. Data centers use large amounts of electricity and water, and lawmakers in both parties have objected that AI facilities do not directly pay those costs—instead surrounding communities bear them. The rapid buildout of data centers occurs in places where resources are scarce: Two-thirds of new data centers built since 2022 are in regions already short on water. A tax on AI facilities could ensure that they bear more of the costs they impose on surrounding communities.
AI firms may be investing more than is efficient. This scenario concerns the pace of AI investment itself. Firms may be spending more on AI than is efficient in the hopes of locking in market share, potentially diverting capital and other resources from more productive investments elsewhere in the economy. It is hard to verify whether this is happening, and if that investment would otherwise generate substantial economic benefits, the tax could discourage productive activity and leave the economy worse off. But if the spending is excessive, a tax that raises its cost could reduce it and improve how capital is allocated, increasing overall economic efficiency.
What’s on the table
Below is a summary of five leading proposals to tax AI, illustrating what would be taxed, when revenue would be received, how easily the tax could be collected, and what compromise, or trade-off, might be required.
A closer look at the mechanics and limitations of each tax proposal’s key trade-offs follows.
Training compute tax: Charge AI developers for the computing power used to train a model
During a training run, computers process large amounts of data to develop a model’s capabilities. The tax would be based on the computing resources used during that process, rather than on the subsequent use of the model.
The main appeal of this tax is timing, since training activity is happening right now. Leading AI labs may not become profitable for some time, so a profit tax might raise little revenue in the near term. A training compute tax targets the development of AI capabilities rather than its use. By making training more expensive, the tax could slow the development of models capable of replacing workers. However, it could also delay improvements in AI that help both AI and workers become more productive.
This tax also has narrower reach than most of the alternatives. Because relatively few frontier AI labs conduct large-scale training runs, the government would have a handful of taxpayers rather than an entire industry to monitor. However, enforcing the proposed tax could be difficult because it would rely on self-reported data, absent a standardized, independent way to verify how much computing power a lab used for training. Enforcement would also require distinguishing between taxable training runs and routine model refinement, which would be difficult.
The growth of the tax base is hard to predict. Algorithmic efficiency, which allows AI models to be trained using less computing power, improves steadily over time. But rather than use those efficiency gains to train the same models with less computing power, labs often use them to train larger, more capable models. Total training compute has grown rapidly as a result. It is unclear which of these trends will prevail over the next decade, making the potential revenue from this tax difficult to predict.
Data center electricity tax: Charge data centers for the electricity they demand
Unlike training compute, electricity use can be independently verified in a standardized manner. A tax on electricity used by data centers could be assessed, for example, on a per-kilowatt-hour basis. Utilities would independently report to tax authorities how much power the data center drew from the grid using existing meters.
Simplicity of administration comes with a catch: Metering tells you how much electricity was used, not how much of that use supported AI activities. Data centers also perform non-AI activities, including cloud storage, video streaming, and financial services, which run on the same infrastructure and draw on the same power supply. Isolating AI’s share of the power use would, like the training compute tax, require a self-reported estimate.
Two approaches have emerged for taxing data centers for their electricity use, with very different revenue implications. One would generate stable and possibly growing revenue through a straightforward per-kilowatt-hour charge that scales with total AI energy use (if we can separate it from non-AI use). The second would require data centers to pay for the electricity upgrades needed to accommodate their demand, rather than passing those costs on to other ratepayers. These charges would recover infrastructure costs rather than generate public revenue. The second approach is closer to what federal and state lawmakers have been calling for.
Usage tax: Charge consumers for what an AI model produces
The electricity tax simplifies administration relative to a training compute tax, but it isn’t well targeted. A usage tax narrows the tax base back to AI itself, namely, the tax base is the tokens a model processes or produces, which could turn into a line of code, a generated image, an answer. Each is charged at a per-unit rate, similar to how an excise tax operates.
But a narrower base isn’t necessarily a better-targeted one. A usage tax falls on all uses, discouraging both the applications that displace certain jobs or tasks and those that make workers more productive.
Administering a usage tax requires defining its base, or what counts as a unit of use. To bill customers, AI providers already measure usage in units called tokens. A token represents a piece of text, such as a word, part of a word, or punctuation. However, tokens vary in length, and different AI models may divide the same text into different numbers of tokens. The computing power needed to process or generate a token also varies across models. As a result, taxing tokens at a uniform rate would not necessarily reflect the amount of computing power used, raising questions about whether tokens are an appropriate basis for taxing AI usage.
How much revenue a usage tax would raise over time is uncertain. Rapidly growing AI use could increase the number of taxable tokens and generate more revenue. However, improvements in efficiency could allow AI models to perform the same task using fewer tokens, potentially reducing revenue per task. At the same time, new models may generate longer responses that require more tokens. The tax’s long-term revenue potential would depend on how these trends affect the total number of taxable tokens.
Windfall profit tax: Charge AI firms for excess profits
Unlike a windfall profit tax, the three proposals mentioned previously would tax AI development, electricity consumption, or usage rather than profits. These taxes could increase businesses’ production costs, particularly when businesses use AI in their operations. Some of those costs might be passed along to consumers through higher prices.
A windfall profit tax works differently. Unlike taxes on AI development or use, a well-designed tax on excess profits could raise revenue without increasing the marginal cost of producing goods and services, because it does not apply to what a company spends or sells. It applies only to profit, and only to the profits that exceed the “normal” return on investment, or the minimum return investors need to justify an investment. Because the normal return is not taxed, a windfall tax on “supernormal” returns to investment has a minimal effect on the decision to invest. In theory, it is one of the few taxes that can raise revenue without changing behavior.
In practice, however, separating normal from supernormal returns is difficult. The normal rate of return varies from investment to investment. A risky bet on a new AI model architecture might reasonably require a higher return to be worth undertaking than a safer investment in additional server capacity. As a result, even one company’s own investments in a single year might not share the same normal return benchmark. These differences make it difficult to determine which profits should be subject to the tax without potentially discouraging worthwhile investments.
Revenue from a windfall tax would also likely be delayed. Few early-stage investments post large profits while they are still in a capital-intensive growth phase, and in the case of AI, many companies are firmly in that position. A windfall tax could also encourage companies to continue investing to reduce taxable profits, potentially delaying revenue collections further.
Automation tax: Charge businesses that use AI to ultimately displace workers
Unlike the first four proposals, an automation tax not tied to AI production nor targeted at AI providers. Instead, businesses that use AI or robotic systems would pay the tax, because these technologies displace human workers. In some proposals, the revenue would be earmarked for public spending to support those workers. Designs vary, from taxing the capital a company invests to automate a task, to taxing the performance of a subsequently automated task.
Administering the tax raises new problems. The research and development (R&D) tax credit offers a useful parallel. It requires the IRS to define, through regulation, what activities qualify as research. That exercise has generated decades of guidance, disputes, and incentives for companies to structure their activity to maximize the credit. An automation tax would face the same pressures, with an added wrinkle: distinguishing AI-driven automation from the many other forms of automation already embedded in production processes. The distinction could be particularly difficult when AI and conventional automation perform similar functions.
How much revenue this tax could raise depends on the extent of AI-driven job displacement. Labor’s share of national income held remarkably steady for several decades after the late 1940s, when the federal government first tracked these data. It has fallen since, through the first decade of the 2000s, and again in recent years. Whether AI accelerates or slows that decline remains uncertain, as does to the potential revenue from an automation tax. Ironically, the tax could generate the most revenue when AI displaces the greatest number of workers—but the revenue intended to help displaced workers might not become available until after they have lost their jobs.
Taxing AI would introduce tensions with other policy goals
Each of the five proposals comes with its own design questions: what to tax, how to administer it, and how much revenue to expect. But two additional considerations beyond conventional tax design are relevant to any AI tax proposal: domestic AI development and national security.
First, some AI-specific taxes could work against existing federal tax incentives that encourage domestic investment and discourage certain offshore activities. The 2017 Tax Cuts and Jobs Act established international tax provisions that Congress modified through the 2025 One Big Beautiful Bill Act. Current law provides a deduction for qualifying foreign-derived income earned by U.S. corporations through the Foreign-Derived Deduction-Eligible Income (FDDEI) provision. It also imposes a minimum tax on certain foreign income through the Net CFC Tested Income (NCTI) provision.
Together, these provisions aim to influence where multinational corporations invest and report their profits. An AI-specific tax could work against these incentives by raising the cost of operating in the U.S. For example, a training compute tax could encourage AI firms to move model training to offshore data centers, potentially shifting activity outside the U.S. tax base and creating additional enforcement challenges.
Second, an offshoring incentive does more than just undercut existing tax preferences for domestic activity. A 2026 Brookings report describes several strategic risks associated with frontier AI data centers built outside the U.S. Overseas facilities are physically exposed to adversarial action, the data and models inside them fall under foreign legal jurisdiction, and hosting AI infrastructure abroad can create long-term dependency on countries whose interests don’t always align with those of the U.S.
That risk isn’t hypothetical. In March 2026, Iranian drone strikes damaged AWS data center facilities in the Middle East. Iranian state media claimed it chose a target in Bahrain specifically because of its role supporting U.S. military and intelligence operations. Any AI tax that could push training or infrastructure offshore would also need to be weighed against the potential national security risks, including exposure to foreign threats, laws, and governments. Those risks would vary depending on where facilities are located and the security and legal protections governing their operations.
Broader tax reforms could address some AI tax proposals’ concerns
The concerns driving proposals for AI-specific taxes could also be addressed through broader changes to the tax system.
- Reforming the taxation of capital income, for example, could address a shift from labor income to capital income regardless of which industries are responsible.
- A cash-flow tax could be designed to replace the corporate income tax with one that falls only on supernormal returns across the economy, rather than isolating a single industry.
- A carbon tax would price the environmental cost of AI’s infrastructure alongside every other source.
- A value-added tax would raise revenue from a much broader tax base than an AI-specific tax, while generally avoiding taxes on business inputs.
Policymakers might still prefer an AI-specific tax but for a different reason. A tax aimed squarely at AI is easier to explain to the public than broad capital income tax reform or a new consumption tax. That advantage is real, but it’s an advantage in politics rather than economics.
Leave a Reply