A bill now before U.S. Congress proposes taxing the companies that build the largest AI models and spending the money on work for people. Its motivation is sound. As AI takes on more of the work people are paid to do, the income that used to reach households as wages flows instead to a handful of companies - preventing that drift is a good reason to legislate.

Economists have a name for a state that lives on income from a single source rather than from the work of its citizens, a rentier state, and such states tend to answer less to their people because they no longer depend on them. In this instance, the concern isn't limited only to the concentration of income from a single source (such as oil) but also a concentration of power in a citizen's everyday life via compute, infrastructure, and the terms on which everyone else is allowed to use them.

Drafting a bill is the right mechanism in our current democracy, but the question is whether this design addresses these concerns. Let’s step through a thought experiment of what might happen with this bill and follow the money.

Step 1: Who pays

The tax falls on any company that develops a foundation model, sells access to one, or modifies an open-weight one, provided it earns money from doing so. A foundation model here means one trained on broad data for general use with at least 10²⁵ operations of computing power, a threshold that captures the frontier labs and those building on their models. Government, university and charitable research is exempt.

So far, so good.

Step 2: How it works

Each covered company pays whichever is larger: two per cent of the value of the tokens its models process, or three per cent of its revenue from AI services. Those rates hold while unemployment stays at or below five per cent. Between five and seven per cent the rate rises point for point with unemployment, and above seven it rises twice as fast. The figure used is the highest annual unemployment rate of the previous three years. The bill does allow the Treasury to discount unemployment caused by war, a pandemic or another shock unrelated to AI, so it is trying to isolate the impact of the technology.

The money goes into a trust fund run by a new Work Protection Administration, which awards grants to states, local government, schools and non-profits to create permanent, full-time jobs in childcare, education, health, elder care, housing construction, public infrastructure and a long list of other public services.

Potential cracks start to show here as we look at where the money goes. The jobs it funds are in care, construction and public services, while the displacement is happening in knowledge work. A paralegal whose role has been automated is not a ready-made care worker.

Most countries would approach this differently. They would tax a booming industry that is concentrating wealth and fund care, housing and public services from general revenue because society needs them, without claiming that one caused the need for the other. This bill makes that claim and wires both the tax and the spending to it.

Its supporters argue the jobs it funds are needed regardless of AI, which rather makes the point: if they are needed anyway, tying their funding to an unemployment figure guarantees they are paid for when the rate is high and cut when it falls.

Step 3: Who really pays?

Start with the companies being taxed. Most of them do not yet make a profit: OpenAI reportedly lost more than a dollar for every dollar it earned in early 2026, and Anthropic recorded its first quarterly operating profit (on an adjusted basis) only this summer. A company running at a loss has little room to absorb another two or three per cent of its revenue, so the cost travels down the line in the price of each token, to the businesses using AI: the logistics firm, the insurer, the regional accounting practice trying to keep pace with larger competitors. This is primarily the knowledge-work sector - not construction or childcare - and many medium-sized businesses that are currently seeing competitive advantage from these tools.

Now stand in the position of one of those businesses. Its AI bill has gone up slightly. An AI agent still works around the clock, takes no holiday, isn't unionised and needs no training, and even with the tax added it will almost always come in cheaper than a person doing the same work. So the business keeps the agent and looks for the savings elsewhere. For most businesses that elsewhere is the largest remaining cost - wages. A tax meant to protect jobs gives the firms that are downstream of it one more reason to cut them.

There is also a door left open. A business that downloads an open-weight model and runs it on its own servers is not developing, selling or modifying anything, so it pays nothing, and several of the most capable open models come from China. A levy on the American route to AI, with none on the alternative, nudges the very firms it hopes to influence towards models built elsewhere.

It can be argued that the tax paid is small as the price of each token has been falling quickly over the last few quarters, and two or three per cent will barely register against that. But the objection cuts both ways: a levy too small to slow adoption is also too small to protect anyone, and it still lands in the wrong place.

Step 4: Why the money arrives too late

Suppose, even so, that jobs are lost and the tax begins to collect. The money is released according to a single number, the unemployment rate, which counts people out of work, including those who have given up looking. It misses most of what is happening now: reduced autonomy in roles, posts not being filled when they fall vacant, and the job that carries on with fewer hours or lower pay. The bill asks the Bureau of Labor Statistics to study that kind of degradation, but nothing the study will find feeds back into the rate, because the rate measures unemployment only.

The tax takes effect a year after the bill becomes law, the rate reads back across three years of figures, and the money then passes through a new agency and a competitive grant process before anything is built. What it builds is housing, roads and public services, which take years to design and deliver. 

The disruption itself is immediate, and growing with each new generation of model. A worker whose role disappears this year needs something this year, not a bridge that opens several years from now.

The timing proposed by the bill compounds the problem it is trying to solve.

A thermostat in one room

The mechanism being suggested is not dissimilar to a thermostat. It reads the temperature in one spot, switches the heating on when the reading falls below its setting, and off again once it recovers. It has no idea why the room went cold, or what is happening anywhere else in the house. This bill works the same way: unemployment is the reading, five per cent is the setting, and the tax and the grants are the heating. It is a thermostat fitted in one room, reading one number, while most of what matters happens in rooms it cannot monitor.

Where the cost lands

Follow the money to its end and the picture is uncomfortable. The cost lands on the businesses adopting AI and, through them, on the workers they no longer need. What stays untouched is the thing that made the bill necessary. The wealth concentrating in AI is held as ownership: of the models, the data centres, the chips, and shares whose value - close to a trillion dollars at each of the leading labs - is a bet on profits still to come. The bill cannot reach any of it, and the power that comes with that ownership - over who gets access, at what price and on whose terms - stays exactly where it was.

For anyone holding capital in the businesses downstream of the labs, this matters in a practical way. The businesses in your portfolio that are adopting AI would carry the cost of the tax, passed down to them in the price of every token. And what the tax is meant to buy, a stable labour market and the household spending, tax base and price stability that rest on it, is the thing this design is least likely to deliver.

What this problem needs is not a better thermostat in one room, but a connected house: one that can see where power is being drawn, by whom and in which rooms. It would read more than one number: how many junior roles are being filled, what is happening to hours and pay inside the jobs that remain, and who controls access to compute and on what terms. And it would balance what it sees for the good of every company, frontier labs included, the economy and the people living there, now and for the generation that inherits it. That is far harder to legislate than a tax rate tied to one familiar number. It is also the only version that reaches what the bill was written for.

Sources

The link has been copied!