Taxing artificial intelligence: What’s on the table, and what’s at stake
Gary Lloyd |
Oct 27,26 | 01:31 EST
The rapid advancement of artificial intelligence (AI) presents a dual challenge and opportunity: while it promises to significantly boost productivity across numerous sectors and tasks, it also carries the potential to automate and displace a substantial number of human workers. The exact scale and pace of this displacement remain subjects of ongoing debate, yet the implications for national economies, particularly the U.S. federal budget, are profound and immediate. The current U.S. tax system is heavily reliant on income derived from human labor, with individual income and payroll taxes collectively accounting for approximately 85% of federal tax receipts. Should AI-driven automation lead to a significant shift of income from wages and salaries to corporate profits and other forms of capital income, these new income streams might be subject to different, potentially lower, effective tax rates. Such a sustained structural change in the economy could result in a significant erosion of federal revenues, coinciding with an anticipated increase in demand for public services designed to support displaced workers, such as unemployment benefits, retraining programs, and broader social safety nets.
In anticipation of these economic shifts, policymakers and lawmakers are increasingly exploring the concept of AI-specific taxes. These proposals are being floated with a multifaceted set of objectives. Primarily, they aim to stabilize or augment federal tax revenues in a rapidly evolving economic landscape shaped by AI. Secondly, a portion of these new revenues could be earmarked to fund essential programs that assist workers whose jobs are disrupted or eliminated by AI. Thirdly, there is a growing bipartisan concern regarding the environmental and infrastructural demands of AI, particularly the intensive electricity and water consumption of large data centers. Advocates suggest that AI taxes could help internalize these external costs, ensuring that AI facilities contribute more directly to the communities they impact. Lastly, some proponents argue that taxation could serve as a mechanism to temper the pace of AI development and growth, particularly if there is a perception of inefficient or speculative investment in the sector.
The various AI tax designs currently under discussion target different points within the AI value chain. These include proposals to tax the immense computing power utilized in training sophisticated AI models, the electricity consumed by data centers that house AI infrastructure, the usage of AI models (e.g., based on processing "tokens"), the "windfall" or "supernormal" profits generated by AI companies, and even a broader "automation tax" levied on businesses that deploy AI to replace human labor. Each of these proposals comes with its own unique set of considerations, offering varying degrees of potential revenue, administrative complexity, and economic impact. For instance, a tax on AI training compute could generate revenue early, but faces challenges in verification and defining taxable events amidst continuous model refinement. A data center electricity tax offers simpler administration through existing utility metering but struggles to isolate AI-specific energy use from other cloud services. A usage tax ties directly to AI output but must grapple with the evolving nature of "tokens" and potential disincentives for productive AI applications. Windfall profit taxes theoretically avoid distortion but are notoriously difficult to implement due to the challenge of defining "normal" returns and are unlikely to yield significant revenue in AI's capital-intensive early growth stages. Automation taxes directly address worker displacement but introduce immense definitional and enforcement complexities given the broad spectrum of automation.
Beyond their direct fiscal and economic effects, the article emphasizes that introducing AI-specific taxes carries broader implications, potentially creating tensions with existing U.S. policy goals. Current federal tax policies, such as the Foreign-Derived Deduction-Eligible Income (FDDEI) and Net CFC Tested Income (NCTI) provisions, are designed to encourage domestic investment and discourage offshore profit shifting. An AI-specific tax, if not carefully designed, could inadvertently increase the cost of AI operations within the U.S., thereby incentivizing AI firms to relocate their training and infrastructure to overseas data centers. This outward migration of critical AI capabilities not only undermines domestic economic policy objectives but also raises significant national security concerns. As highlighted by a 2026 Brookings report, offshore AI facilities are susceptible to adversarial actions, operate under foreign legal jurisdictions, and could foster long-term dependencies on countries whose geopolitical interests may not align with those of the U.S., as exemplified by recent incidents involving damage to AWS data centers abroad.
Ultimately, the article posits that many of the core concerns driving the push for AI-specific taxes—such as the potential erosion of the tax base or the need to internalize environmental costs—could also be addressed through more comprehensive reforms to the broader tax system. Options like overhauling capital income taxation, implementing a corporate cash-flow tax that targets supernormal returns across all industries, introducing a carbon tax to uniformly price environmental externalities, or adopting a broad-based value-added tax (VAT) are presented as economically sound alternatives. While these broader reforms might lack the political appeal of an AI-specific tax, which is often easier to communicate to the public, their potential for wider economic efficiency and revenue stability should not be overlooked in the ongoing policy debate.
Why tax AI? The article outlines three main reasons for taxing AI. First, AI could erode the tax base as it displaces workers, reducing income and payroll tax revenues while increasing demand for social safety nets. Second, AI infrastructure, particularly data centers, imposes costs (e.g., electricity, water) on surrounding communities that are not directly borne by AI users, suggesting a need for taxation to cover these externalities. Third, there's a concern that AI firms might be over-investing in AI to secure market share, potentially diverting capital from more efficient investments, which a tax could help correct. What’s on the table This section introduces a comparative summary of five leading proposals for AI taxation, detailing what would be taxed, revenue timing, ease of collection, and potential trade-offs. The subsequent content elaborates on each of these proposals individually. Training compute tax: Charge AI developers for the computing power used to train a model This tax targets the computing resources used to train AI models. Its main advantage is immediate revenue generation, as training is ongoing. It aims to slow the development of worker-displacing models but might also hinder beneficial AI advancements. Administration is challenging due to the reliance on self-reported data and the difficulty in distinguishing between taxable training and routine model refinement. Revenue prediction is uncertain, influenced by algorithmic efficiency gains and the trend towards larger models. Data center electricity tax: Charge data centers for the electricity they demand This proposal taxes data centers based on their electricity consumption, which is easily verifiable. The challenge lies in distinguishing electricity used for AI activities from other data center functions. Two approaches exist: a per-kilowatt-hour charge for stable revenue or requiring data centers to cover infrastructure upgrade costs. Usage tax: Charge consumers for what an AI model produces A usage tax would be levied on the "tokens" an AI model processes or produces, similar to an excise tax. While narrowing the tax base to AI, it risks discouraging both job-displacing and productivity-enhancing applications. Defining a consistent unit of "use" (tokens) is problematic due to variations in token length and computational intensity across models. Future revenue is uncertain, depending on AI adoption, efficiency improvements, and output length. Windfall profit tax: Charge AI firms for excess profits This tax targets "supernormal" profits exceeding a normal return on investment, theoretically without affecting investment decisions. However, practically differentiating normal from supernormal returns is difficult due to varying risk profiles of investments. Revenue generation is also likely to be delayed, as many AI firms are currently in capital-intensive growth phases with low immediate profits. Automation tax: Charge businesses that use AI to ultimately displace workers This tax targets businesses utilizing AI or robotics that lead to worker displacement, with revenues potentially funding support for those workers. Administration is complex, similar to R&D tax credits, requiring clear definitions of what constitutes AI-driven automation. Revenue potential is tied to the unpredictable extent of AI-driven job displacement, meaning funds might only become available until after jobs are lost. Taxing AI would introduce tensions with other policy goals Beyond design and revenue, AI tax proposals face two key considerations: domestic AI development and national security. Such taxes could counteract existing federal tax incentives for domestic investment, potentially pushing AI training and infrastructure offshore. This offshoring creates strategic risks, including physical exposure to adversarial actions, foreign legal jurisdiction over data, and dependence on countries with differing interests, as highlighted by a 2026 Brookings report on overseas AI data centers. Broader tax reforms could address some AI tax proposals’ concerns The article suggests that concerns addressed by AI-specific taxes could also be tackled through broader tax system changes. Examples include reforming capital income taxation, implementing a cash-flow tax on supernormal returns across all industries, introducing a carbon tax to price environmental costs, or adopting a broad-based value-added tax (VAT) for broader revenue generation. While these broader reforms are economically sound, AI-specific taxes might be politically more appealing due to their direct and easily explainable nature.
Why tax AI?
The article outlines three main reasons for taxing AI. First, AI could erode the tax base as it displaces workers, reducing income and payroll tax revenues while increasing demand for social safety nets. Second, AI infrastructure, particularly data centers, imposes costs (e.g., electricity, water) on surrounding communities that are not directly borne by AI users, suggesting a need for taxation to cover these externalities. Third, there's a concern that AI firms might be over-investing in AI to secure market share, potentially diverting capital from more efficient investments, which a tax could help correct.
What’s on the table
This section introduces a comparative summary of five leading proposals for AI taxation, detailing what would be taxed, revenue timing, ease of collection, and potential trade-offs. The subsequent content elaborates on each of these proposals individually.
Training compute tax: Charge AI developers for the computing power used to train a model
This tax targets the computing resources used to train AI models. Its main advantage is immediate revenue generation, as training is ongoing. It aims to slow the development of worker-displacing models but might also hinder beneficial AI advancements. Administration is challenging due to the reliance on self-reported data and the difficulty in distinguishing between taxable training and routine model refinement. Revenue prediction is uncertain, influenced by algorithmic efficiency gains and the trend towards larger models.
Data center electricity tax: Charge data centers for the electricity they demand
This proposal taxes data centers based on their electricity consumption, which is easily verifiable. The challenge lies in distinguishing electricity used for AI activities from other data center functions. Two approaches exist: a per-kilowatt-hour charge for stable revenue or requiring data centers to cover infrastructure upgrade costs.
Usage tax: Charge consumers for what an AI model produces
A usage tax would be levied on the "tokens" an AI model processes or produces, similar to an excise tax. While narrowing the tax base to AI, it risks discouraging both job-displacing and productivity-enhancing applications. Defining a consistent unit of "use" (tokens) is problematic due to variations in token length and computational intensity across models. Future revenue is uncertain, depending on AI adoption, efficiency improvements, and output length.
Windfall profit tax: Charge AI firms for excess profits
This tax targets "supernormal" profits exceeding a normal return on investment, theoretically without affecting investment decisions. However, practically differentiating normal from supernormal returns is difficult due to varying risk profiles of investments. Revenue generation is also likely to be delayed, as many AI firms are currently in capital-intensive growth phases with low immediate profits.
Automation tax: Charge businesses that use AI to ultimately displace workers
This tax targets businesses utilizing AI or robotics that lead to worker displacement, with revenues potentially funding support for those workers. Administration is complex, similar to R&D tax credits, requiring clear definitions of what constitutes AI-driven automation. Revenue potential is tied to the unpredictable extent of AI-driven job displacement, meaning funds might only become available until after jobs are lost.
Taxing AI would introduce tensions with other policy goals
Beyond design and revenue, AI tax proposals face two key considerations: domestic AI development and national security. Such taxes could counteract existing federal tax incentives for domestic investment, potentially pushing AI training and infrastructure offshore. This offshoring creates strategic risks, including physical exposure to adversarial actions, foreign legal jurisdiction over data, and dependence on countries with differing interests, as highlighted by a 2026 Brookings report on overseas AI data centers.
Broader tax reforms could address some AI tax proposals’ concerns
The article suggests that concerns addressed by AI-specific taxes could also be tackled through broader tax system changes. Examples include reforming capital income taxation, implementing a cash-flow tax on supernormal returns across all industries, introducing a carbon tax to price environmental costs, or adopting a broad-based value-added tax (VAT) for broader revenue generation. While these broader reforms are economically sound, AI-specific taxes might be politically more appealing due to their direct and easily explainable nature.
Read More
