A token tax bills the meter that keeps getting cheaper

Yesterday Bill Gates made the case for taxing robots and AI tokens. Six days before that, Raghuram Rajan, who used to run India’s central bank, proposed a tax on the tokens businesses consume, starting small and rising as the displacement data comes in. Bloomberg ran a guide to the whole menu a week ago. Two people who rarely agree on tax policy reached for the same instrument within a week of each other.

They are right about the problem they are pointing at. Gates puts it plainly: “The tax system nudges you toward replacing people with machines.” Hire a person and you pay payroll tax on their earnings. Buy the machine that does their job and you write it off. Rajan makes the same observation from the social security side. A firm contributes for every worker and contributes nothing for the software that replaced one.

The asymmetry is real. The meter they picked to fix it is the wrong one.

The base shrinks while the harm grows

A tax on tokens is a tax on compute spend, and compute spend per unit of work is falling faster than almost anything else in the economy. Frontier model pricing per million tokens in 2026 is roughly one fifth of what it was two years ago, and the cheap tiers have fallen further than that. An agent loop that cost forty cents a task in 2024 runs around eight cents now.

Point a tax at that and you have indexed worker compensation to a number the entire industry is racing to drive down. Every efficiency gain, every smaller model that does the same job, every prompt cache, cuts the fund. Meanwhile the thing you actually care about, the number of people whose jobs went away, moves in the opposite direction. Cheaper inference is exactly what makes automating a six person support desk affordable for a company that could not justify it last year. The tax base falls as the displacement spreads. You would be funding the response to a problem with a meter that gets quieter the worse the problem gets.

Rajan half concedes this. He suggests starting the rate low and raising it as evidence accumulates. So the rate has to climb just to stand still, and every increase is a fresh political fight against an industry that will be much larger by then.

Tokens do not know what they did

The deeper issue is that a token has no idea whether it took a job.

Picture two companies with the same customer volume. The first keeps its thirty support reps and gives each of them an AI assistant that drafts replies, pulls order history, and summarizes threads. Those reps are faster and happier and still employed. That company burns an enormous number of tokens, because thirty humans in the loop generate far more model calls than an automated pipeline does.

The second company cut the desk to five people and runs an agent that answers most of the mail on its own. Cached context, a small fast model on the common questions. Fewer tokens.

Under a token tax, the company that kept twenty five jobs pays more than the company that eliminated them. That is not a corner case. It is the normal shape of AI assistance versus AI replacement, and it points the incentive backwards. Rajan’s own proposal quietly admits it, since he pairs the tax with a retention credit to undo the effect. When the fix needs a second policy to cancel out what the first one does, the instrument is doing the wrong job.

Meter the thing that grows

If you want a fund that scales with the harm, tie it to something that goes up when the automation works. Revenue goes up. Profit goes up. Those are the numbers that rise precisely because a payroll line went away, and unlike token counts, they are already audited.

That is the shape we picked. Our commitment, the Dividend Standard, is the greater of 5% of our qualifying revenue or 70% of our adjusted profit, every year, directed to the workers our software displaces and granted as ownership held in a trust rather than a check that a future board can stop writing. The revenue floor matters most early, when a company can reinvest everything and show no profit at all. The profit figure takes over later. Neither one falls when inference gets cheaper.

The other half is who gets paid. A tax collects into a general pool, then argues for years about distribution. We tied payment to a named person: a covered front line service role, at a company running Celeste, with a documented separation inside a twelve month window. When those line up, eligibility is presumed, and the former employer can object with cause but never has to approve. The claim process is open now, before the fund is large, because a promise nobody has built the plumbing for is not a promise.

None of this argues against a tax. Something public will be needed, and Gates is right that budgets will be squeezed from both sides at once. But whatever gets built should be metered against a number that grows when people lose work, not against the one input everyone in the industry is competing to make cheaper. Get the meter wrong and the fund runs dry exactly when it is needed most.