Sam Altman spent Thursday pitching the US government a 5 percent stake in OpenAI. At the $852 billion valuation from the March funding round, that comes to about $42.6 billion of company. His framing is that every leading US AI lab should pay the same tax, into an Alaska-style public fund, so the country holds a share of the technology it is helping regulate. It is a real idea. It also arrived exactly six days after the government asked OpenAI to hold back the public release of GPT-5.6 and gate access to a customer-by-customer approval list. Both of these things can be true. It is worth reading them together.
The GPT-5.6 delay was reported on June 25, and the family (Sol, Terra, Luna) is now shipping only to a set of government-approved partners during preview. Altman told employees this was not a template OpenAI wants to normalize. The 5 percent stake proposal, six days later, is the sound of an executive team that has looked at the room and decided the price of getting the model out is a check the size of Ford Motor Company. The choice is not “release now” versus “release later.” It is “release later, and pay for the privilege.” That is a very specific position to negotiate from, and it is not the position OpenAI held 90 days ago.
The reason the position has weakened is the number that came out this week from the other side of the industry. Anthropic’s annualized revenue is now above $30 billion. OpenAI’s sits around $24 to $25 billion. That is a crossover. It happened earlier than the Epoch AI forecast from February, which pegged the flip for mid-2026 at the earliest. Anthropic’s revenue has been compounding at roughly 7x year over year since last summer. OpenAI’s is compounding at roughly 3.4x. This is not a rounding-error lead. It is a trajectory difference.
The reason the trajectories differ is the story theclank has been writing all summer. Anthropic has been shipping. Sonnet 5 dropped as the default model on every Claude account on July 1, priced at $2 per million input tokens and $10 per million output through August, with the Claude Code rate limits raised the same day. It runs agents. It is the thing enterprise buyers are wiring into production. Meanwhile Sol, Terra, and Luna are behind a US government access sheet with about 20 organizations on it. If you are a Fortune 500 CTO deciding what to standardize on this month, you are picking between a model your team can call today and a model your team cannot see until a federal preview period ends. Nobody is picking the model they cannot see.
This is what makes the equity pitch interesting. Altman is not offering the government $42.6 billion because OpenAI is feeling generous. He is offering it because the alternative is a preview process that could stretch into the fall, and every week that stretches is another week Sonnet 5 gets to be the default choice at Anthropic’s pricing while GPT-5.6 sits on a shelf. Whether or not the fund idea is good policy (and it is not the worst policy proposal to come out of the last month), the timing suggests OpenAI has done the math on the cost of the delay and concluded that the cost of settling is less than the cost of waiting. The math would look different if OpenAI still had a two-to-one revenue lead. It does not.
The other thing worth naming is the shape of the ask. Altman said every leading US lab should pay in. Anthropic, Google, Meta, xAI, and the neoclouds are not on record as agreeing to any of this. Anthropic in particular has spent the summer pointing at guardrails-in-public as the alternative model to guardrails-by-guest-list. It is easier to accept a 5 percent tax when the tax is what unlocks your product than when the tax is what levels a playing field you are winning. The equity offer, whether it is intended this way or not, effectively proposes that OpenAI’s specific problem this week become the industry’s shared cost of doing business.
The remaining question is whether the government accepts. If it does, GPT-5.6 ships broadly on some accelerated timeline, OpenAI gets its comeback narrative, and every other US lab suddenly has a new line item on its cap table. If it declines and the preview period keeps grinding forward, the revenue gap widens for another quarter and the July 4 piece from this site becomes the standing read on OpenAI’s summer. Either way, the pitch itself is the tell. Companies that are winning do not offer the referee 5 percent of the ball.