Taxing AI Compute Would Be a Mistake

The rise of artificial intelligence (AI) has led to growing calls for a tax on AI computing power, or a “compute tax.” But implementing such a policy would be a mistake.

The idea of something akin to a compute tax dates back to 2017, when Bill Gates proposed a tax on robotics. Gates and fellow advocates argued that robots would displace workers and shrink the tax base, thereby requiring a new source of revenue to make up the shortfall. Pundits and outspoken public figures are now applying similar logic to AI, asserting that a tax on computing power is necessary to slow AI's rise and ensure the tax base doesn’t shrink as wages potentially account for a smaller portion of taxable income in America. But these arguments overlook the substantial economic and social benefits AI can generate.

There are several reasons an AI compute tax would be misguided.

First, a compute tax would slow the productivity gains AI can deliver.

AI is a productivity-enhancing technology that, in and of itself, presents significant benefits for the American public. Over the past several decades, U.S. labor productivity has stagnated, falling below projections and far below the highs of the 1960s, when it often grew by over 3 percent annually. From 2025 to 2035, productivity is expected to grow by only 1.4 percent annually, according to projections from the Congressional Budget Office. Returning productivity growth to levels closer to those experienced in the 1960s would raise per capita GDP and improve consumer welfare. Automation and AI-enabled tools will play a central role in achieving those gains.

Placing a tax on AI would impede, potentially significantly, AI’s ability to power productivity growth. Whether imposed on firms operating AI data centers or on consumers purchasing AI-enabled products and services, such a tax would increase the cost of deploying AI systems. Higher costs would discourage adoption, particularly for smaller firms with fewer resources. For firms and industries that have already adopted AI, including in pharmaceutical development, epidemiology, weather forecasting, and fraud detection, a compute tax would increase costs, slowing diffusion among industries that rely on the technology for accurate predictions and trend analysis.

Second, a compute tax would push investment and innovation abroad, weakening U.S. AI leadership.

Some of the largest proponents of an AI tax claim that such a tax would slow down AI development, giving policymakers and workers time to adjust to this new AI-driven world. But AI doesn’t exist solely in the United States, nor will its benefits necessarily be concentrated there. Implementing an AI tax would not hinder global AI development; it would only ensure that AI investment and innovation are directed elsewhere. Such a shift would weaken U.S. leadership in AI innovation and infrastructure while doing little to reduce the pace of global AI advancement. AI firms have already begun looking abroad for data center construction sites as U.S. policymakers at the state and federal levels push back against domestic development. Brazil and Mexico have become the newest hotspots for investment, while the United Arab Emirates has long aimed to become a data center hub.

Third, a compute tax would do little to preserve the tax base.

The mathematical logic for the tax is patently false. Proponents argue that machines are increasingly substituting for human labor, reducing labor’s share of national income and thus shrinking the tax base. However, the share of income going to productive capital, such as machinery, AI, and robotics, has not increased over this same period. Instead, the decline in the labor share has been offset by an increase in the share of income going to housing capital (rent paid by renters and mortgage payments by homeowners) and self-employment income (another form of labor income). In other words, the decline in labor’s share has not primarily been driven by automation replacing workers. If automation-related capital income is not actually replacing labor income, taxing AI compute would do little to address the underlying cause of changes in the tax base.

Additionally, because economic growth increases the size of the U.S. economy, and therefore the tax base, labor’s share of income would need to fall dramatically and persistently over many years to erode the tax base. ITIF finds that when labor’s share of income is about 57 percent, it would need to fall by 5 percentage points in a single year for tax revenue to decline. A decline of this magnitude has no historical precedent, as labor’s share of income has fallen by about 0.1 percent per year over the past 73 years.

Fourth and finally, a compute tax would not help solve a mass unemployment problem AI is unlikely to create.

It is extraordinarily unlikely that AI will result in mass unemployment. It’s estimated that about 8 percent of jobs, at most, are at high risk of automation, far below some of the more alarmist projections commonly cited. And just because 8 percent of jobs are at risk does not mean that 8 percent of the labor force will remain permanently unemployed. Technological change has always created periods of adjustment, but it has never led to sustained declines in overall labor demand. Higher productivity from automation will reduce prices, increasing consumption and investment. Ultimately, that growth will require more workers, not fewer—creating jobs rather than eliminating them.

Policymakers should not focus on slowing the development of AI through an AI compute tax; they should focus on helping workers transition to a labor force increasingly augmented by AI. This includes increasing workforce and digital literacy training and improving worker adjustment programs such as unemployment insurance.