Mark Cuban has proposed a federal tax on AI tokens — under 50 cents per million tokens — aimed at pushing large model operators toward efficiency while raising roughly $10 billion annually. The proposal, which exempts open-source models and local inference, enters a debate over AI energy consumption and government oversight. Cuban made the pitch this week, tying it to lessons from crypto regulation.
How the tax would work
The rate is tiny: less than half a cent per million tokens processed by large language models. Cuban says the fee would push big operators to optimize their compute, cutting waste. He estimates it'd generate about $10 billion in the first year, with that number growing as inference demand climbs. The motivation isn't just revenue — it's energy. Data centers for large models are draining U.S. power grids, and Cuban argues a usage tax forces efficiency better than voluntary pledges.
Open-source models and any inference running locally — on a user's own machine — are exempt. That carve-out is meant to avoid stifling small developers or hobbyists.
Critics push back
Not everyone is buying it. Anduril founder Palmer Luckey argues the tax would hurt American AI firms, driving customers to foreign models that face no such fee. He also warns it'd create the infrastructure for government tracking of who's using AI and for what. The tracking concern, Luckey says, amounts to a surveillance mechanism disguised as a green policy.
Cuban's regulatory framing
Cuban himself draws a parallel to the early crypto days. He notes that resistance to regulation was fierce at first, but eventually the industry started lobbying for clear rules rather than fighting every proposal. He's suggesting the AI sector might follow the same arc — better to engage now than get hit with something worse later.
Congressional appetite — or lack of it
The proposal currently has little support in Congress and is unlikely to advance this session. Lawmakers are still figuring out their stance on AI broadly, and a new tax isn't a priority. But the energy problem isn't going away — data center buildouts are only accelerating. Whether Cuban's idea gains traction later or some other mechanism emerges, the debate over who pays for AI's power hunger is just getting started.
