A Dollar a Year, Then Half Off, With No Minimum

In August 2025, OpenAI sold its workplace tools to the executive branch for a dollar a year. Anthropic then sold its government model to all three branches for the same dollar ^[1]. Thirteen months later, the General Services Administration signed a blanket agreement cutting the price of OpenAI's most advanced models in half for every federal, state, and local agency, with no minimum spending commitment ^[2]. Read those as commercial terms. A seller granting half off with no volume floor is a seller with nothing left to bargain with. The customer, in these documents, concedes nothing on its side. The seller has reasons to bend. Over the past year OpenAI cut the price of one of its models by 80 percent and Anthropic released a top model at half its usual cost, both moving against cost-efficient Chinese rivals and open-source systems that have nearly closed the performance gap — and doing it while racing toward profitability ahead of possible public offerings ^[3]. Morgan Stanley put the squeeze in writing: open-weight models will keep pressing token prices down, and the durable returns in this industry are migrating to proprietary chips like Amazon's Trainium and Alphabet's TPUs, not to the models themselves ^[4]. An August analyst note then wrote out what a moat for these models would have to be made of, since the models themselves are "easily replicated": restricted distribution, premium pricing, Know Your Customer verification, and "government-issued licensing and accreditation to manage the risks associated with powerful models" — ending with the most powerful models "reserved exclusively for government use" ^[5]. The prescription is half-executed. The gating and licensing lines show up in the record; the pricing line runs the opposite direction, the labs cutting prices rather than charging a premium for their moat ^[3][2]. The same companies briefing the United Nations Security Council about the end of the species ^[6] are the ones now on sale to that government — a dollar a year, then half off, with no minimum ^[1][2]. The most direct route to the state's money was closed first. In November 2025, asked for federal support, the administration said no.

We believe that governments should not pick winners or losers, and that taxpayers should not bail out companies that make bad business decisions or otherwise lose in the market. — Sam Altman

Altman, whose company had requested the support, matched the line at the time.

There will be no federal bailout for AI. — David Sacks

That door has stayed shut. What opened instead were the side channels. The discount: the GSA's half-off agreement carries no minimum, so the government pays only for what it uses and can walk away owing nothing ^[2]. The gate: an export-control directive forced Anthropic to abruptly shut two of its models off to foreign nationals, and India — its IT giants cut off overnight — is now building its own sovereign models, one Indian lab warning users not to mistake access for ownership ^[7]. The insurance: by late September the ask had moved from discount to underwriting, and OpenAI's finance chief called for a government backstop to keep the industry's financing flowing ^[8].

Given the magnitude of what I expect A.I.’s economic impact to look like, the government should serve the role of "insurer of last resort." — Sam Altman

Ten months later, he asked the same taxpayer to stand on the other side of the bet. The reading of that rhetoric as leverage is on the record too, and it belongs to Alex Karp of Palantir. Karp accused the labs of using apocalypse narratives to secure a Section 230-style liability shield — protection from the lawsuits over training on clients' data that would otherwise hang over their valuations — and offered the only alternative he sees: a nationalization at a 50 percent government stake ^[9]. The bid is contested at every layer, which is the honest way to say it is not yet a capture. The leading candidate for the new AI czar, Treasury Secretary Scott Bessent, argues for developer liability rather than federal shields ^[6]. Nvidia's Jensen Huang counters that safety is an engineering problem, not a legal one, and that the market is already doing the work ^[9]. And the company at the center of all this is not failing: OpenAI passed a $10 billion annualized revenue run rate — a business growing fast while still deeply unprofitable, a seller with a real product and a pricing problem, not a rescue case ^[10]. The underwriting, though, is unsigned. Every term so far is an ask: the insurer role a sentence in the record, the backstop not yet a statute ^[8]. The customer still holds discount terms with no minimum attached ^[2]. And the counter-signatures on the document are refusals — the administration's no-bailout line, India building its own models rather than renting access ^[7]. The terms that would settle it are specific: a future discount that arrives with a minimum attached, a backstop that moves from ask to law, and the new AI czar's first ruling on who pays when the model fails.