OpenAI's free licences for Washington are a purchase, not a gift

This week OpenAI said that American governments—federal, state, local and tribal—can now take ChatGPT Enterprise for free, with usage charged at half the standard price, under an agreement that runs from October 2026 to the end of 2028. The licence normally costs $15 per user per month. Washington has been handed the seats; the bill, at a 50% discount, will come for how much the machines are actually used. On its face the arrangement reads as a subsidy to the state. Look again: this is how a private company buys itself the largest single customer of information-technology services on the planet, and it is a useful window onto what OpenAI's backers believe its value is.

The world's biggest buyer

Start with the customer, because scale is the whole point. The American federal government is the biggest single purchaser of IT services anywhere, which is precisely why the frontier AI firms fight over it. The predecessor to the new deal shows how large the prize became in only a year. In August 2025 OpenAI struck a pilot through the General Services Administration under which agencies paid a nominal $1 to give their staff ChatGPT. It claims 3.5m federal employees came to use it, generating $1.4bn of cost savings. The new agreement extends that reach to every level of government and to a public-sector workforce of roughly 23m.

Free seats, half-price tokens

Read the pricing and the strategy reveals itself. OpenAI has split its price into two parts: the fixed licence, dropped from $15 a user a month to nothing, and the variable usage, discounted by half. With no minimum commitment and no procurement hurdle, the on-ramp is frictionless. That is deliberate. A firm that sells a machine cheaply and then the consumables minted in volume is using a two-part tariff: the giveaway removes the obstacle to adoption, and the money is made on scale. For OpenAI, whose marginal cost of answering a query falls as its datacentres mature, the usage pool is where the revenue now has to live. Free seats, half-price tokens: the seats are bait, the tokens are the trade.

The gift is aimed at a rival, not at the taxpayer. Anthropic, maker of the Claude models and the firm that had come to dominate classified federal and defence work, collided with the Pentagon early in 2026, when the Department of Defense labelled it a supply-chain risk; only in August did it win a court ruling against the government. OpenAI, by contrast, sealed a classified-information agreement with the department in February. In a market built as much on trust and procurement plumbing as on model quality, the free deal converts a political opening into an installed base—and does so at the moment the rival best placed to contest it is wounded.

The market is being asked to price the habit

Here the story stops being a curiosity for taxpayers and becomes one about expectations, which is the part that should interest an investor. OpenAI is not public, so there is no direct ticket to buy. But it is preparing for an initial public offering that reports value at around $1 trillion, and the disclosed fundamentals give the giveaway its context. Revenue reportedly runs at about $2bn a month on a product used weekly by more than 900m people; in February a capital round raised $110bn at a $730bn pre-money valuation, much of it for compute. For all that revenue, the firm still burns through capital on an industrial scale.

Against that arithmetic, giving the product away is a purchase of a form of earnings OpenAI cannot yet show: durability. An installed default inside the machinery of the state, revisited daily by tens of millions of public servants, is a defensible rent of exactly the kind that justifies a premium before the profits exist. The state's habit, once set, becomes a switching cost that competitors must pay to dislodge—and the price of that exclusion now falls on Anthropic and the rest, not on OpenAI's own books.

The trouble is that the accounting does not yet prove the point. The $1.4bn in claimed savings is a benefit to the government, not revenue to OpenAI. Free seats produce no licence bill, and half-price usage must scale enormously before it pays for the compute. Whether the bet converts into durable, paid usage is exactly the question no near-term disclosure will answer. None of this makes the strategy foolish—it is the standard playbook of a firm buying a distribution monopoly while the window is open—but it is a bet on volume arriving, made before the volume has arrived.

For the retail investor the lesson is about expectations rather than positions. Because OpenAI is private, the tradeable echo runs through Microsoft, its biggest shareholder and long its partner in compute, which rises and falls with the firm's worth even after the two loosened their exclusive Azure ties this year. The larger point outlasts any ticker. When a money-losing company gives its product to the world's biggest buyer for free, it is not being generous. It is pricing its dominance in the only coin it has—a daily habit, inside a government that changes its software slowly—and asking the market to value the habit before the profit arrives. The state gets a discount. Investors get the bill.