Sign in Subscribe

The AI Token Tax: Who Wants It, and Who Pays

The AI Token Tax: Who Wants It, and Who Pays
Taxing AI tokens

A new phrase is moving through policy and finance circles: the AI token tax. Proposals are emerging from members of Congress, gubernatorial candidates, investors, and AI executives. The mechanisms differ. The shared premise is that AI's economic footprint warrants a new tax base. The rates floated for the token-based versions are fractions of a cent per token, but across an industry running trillions of tokens a day, that math reaches billions in annual revenue. Here's what's actually on the table and how it would land on companies and marketing organizations specifically.

Who's proposing it

This isn't one plan. It's a cluster of token-tax proposals at very different levels of seriousness.

The people with actual policy weight are elected. Rep. Greg Casar (D-TX), chair of the Congressional Progressive Caucus, has pitched a federal token tax framed around worker protection, arguing the tax code rewards automation because firms pay payroll tax on humans but face no equivalent cost when replacing them with AI. Casar's framework pairs a token count with a compute baseline to prevent providers from gaming token definitions, with revenue funding a New Deal-style federal jobs program. Tom Steyer, currently running for California governor, wants a state-level fee on AI usage to fund a Golden State sovereign wealth fund.

Outside Congress, business leaders have proposed their own versions. Mark Cuban proposed a federal charge of under 50 cents per million tokens, structured like a sales tax on commercial providers, with open-source and locally run models exempt, projecting roughly $10 billion in annual revenue at the proposed rate. Vinod Khosla pitched a per-token fee to seed a national wealth fund modeled on Norway's Government Pension Fund, paired with eliminating income tax for households under $100K.

None of this is law anywhere yet, and there's little real momentum in Washington. For now it's a debate that shapes future policy, not a statute about to pass. Worth tracking the way crypto regulation was worth tracking five years ago.

Impact on companies and marketing organizations

AI costs are already climbing due to agentic AI proliferation. A tax adds another layer for heavy users.

Critics argue a token tax would fall hardest on smaller and emerging players, penalizing the experimentation that surfaces AI's most useful applications while well-funded incumbents absorb the cost easily. They also note that tokens measure quantity, not value, so taxing every generation as a flat unit treats raw output as equivalent regardless of whether it produces revenue or margin.

Additionally, a U.S.-only tax on U.S. providers effectively subsidizes foreign AI. It wouldn't touch Chinese, Gulf, or European models, so it risks pushing usage offshore rather than raising the intended revenue. The recommendation from the Brookings Institution, a Washington-based public policy think tank, is to hit retail use and exempt business-to-business transactions so the cost doesn't cascade through production. That design choice limits revenue but also limits the leakage risk. Both are constraints a Treasury would have to balance.

For marketers specifically, the effects are already visible as the economics of AI continue to shift. Agencies are building token costs into client pricing as a pass-through line item, and consultants expect to begin auditing agency token usage the way they audit media spend today. A government levy sharpens that further. "Tokens per campaign" becomes a number CFOs scrutinize.

Will it open the door to taxing AI revenue in general?

That door is arguably already ajar. The proposals aren't unified around safety nets. Khosla's and Steyer's versions fund sovereign wealth funds, which is general wealth accumulation, not targeted social spending. And tax bases tend to be sticky and expandable. Once a working mechanism exists to count and tax tokens (or meter and tax AI data center energy), and once it's generating billions, the temptation to raise the rate or redirect the revenue grows. The income tax started narrow and grew.

The one real brake is the offshore-leakage problem above: push the rate too high and usage migrates to providers the tax can't reach. So "earmarked for the safety net" is a political promise with a genuine economic ceiling, not a structural guarantee.

How a tax of this size shows up in pricing

At Cuban's proposed rate, a typical 2,000-token exchange carries a tax of roughly a tenth of a cent. Per interaction the tax is invisible. Spread across trillions of tokens daily, the rate aggregates to the roughly $10 billion-a-year figures being quoted. At these levels, providers most likely fold the cost into subscription fees and enterprise contracts rather than itemize it. Heavy industrial users feel it clearly. Lighter users mostly don't, especially since it blends into price creep that's already underway.

For marketing organizations, the practical signal is upstream of any statute. The pricing model debate is already happening, agencies are already passing token costs through, and "tokens per campaign" is already moving onto CFO scorecards. The broader picture of why enterprise AI costs are rising even as token prices fall is laid out in Marketing Embeddings' working paper on AI cost economics, which breaks down the five variables driving total AI spend across the enterprise.

The token-tax cluster isn't the only AI-tax conversation happening. Two parallel proposals use different mechanisms entirely.

Senator Elizabeth Warren (D-MA), in a May 27 Time op-ed, proposed a per-kilowatt-hour excise tax on the energy AI data centers consume, calibrated to the scale of operations so the largest players carry the biggest load, paired with a wealth tax on AI-sector billionaires. Revenue would fund worker protections, universal healthcare, education, and unemployment insurance. Warren is taxing AI infrastructure energy and AI-sector wealth, not AI output.

Anthropic CEO Dario Amodei proposed a 3% tax on the revenue AI providers generate, arguing that without a redistribution mechanism, AI will produce mass unemployment without a social cushion. Amodei's mechanism taxes the dollar revenue of token sales, not the tokens themselves.

Both proposals share the redistributive framing of the token-tax cluster but pick different bases. Energy, wealth, and revenue are all candidate tax bases for "taxing AI." The choice of base affects who pays, who escapes, and where the offshore-leakage risk concentrates.

Sources

Neutral / institutional analysis

News reporting

Industry / trade analysis

Advocacy / opinion (read as the proposer's own case)

Critique / opinion

Comments ()