Two issues ago we watched Apple become a tenant. Weeks ago we watched capital and government elbow their way to the front of the line as the frontier's landlords. This week a third party showed up with its own offer sheet, and its asking price was a lot higher than anyone expected.
OpenAI's proposal to hand Washington a slice of itself is now two weeks old, but this week it picked up a coauthor nobody planned for. A new poll found that most Americans do not just like the idea of a government stake, they want to go further and simply take half. Meanwhile in Washington, the voluntary review process born out of Anthropic's chaotic June shutdown quietly processed its second case, and this time it barely made news because nothing went wrong. And on the other side of the Pacific, a different kind of claim on the frontier played out, with Alibaba retaliating against Anthropic's own theft accusations by banning Claude from its own offices.
Different claimants, same fight. The public, Washington, and Beijing's largest tech company are all trying to formalize a stake in what the labs build, whether that stake is equity, a checkpoint, or control over who gets access. The labs, for their part, spent the week shipping product updates as if none of it were happening.
OpenAI has spent the past two weeks trying to answer a question this newsletter flagged back when Bernie Sanders was still a long shot. Sam Altman is discussing a plan to hand the federal government a 5% stake in OpenAI, worth roughly $42.6 billion at the company's $852 billion valuation, funneled into a Public Wealth Fund modeled on the Alaska Permanent Fund. Altman has pitched the idea directly to President Trump, Commerce Secretary Howard Lutnick, and Treasury Secretary Scott Bessent, and reportedly discussed it with Sanders himself last month.
The mechanism matters more than the headline number. OpenAI would not sell shares or dilute existing cash investors. It would carve out equity for a government-controlled vehicle that, like Alaska's fund, could eventually pay dividends to ordinary citizens. Altman wants the arrangement to extend industry wide, with Google, Meta, and Anthropic contributing matching slices. None of them has agreed to anything yet.
Anthropic in particular has stayed out of the equity talks entirely. The company has floated its own version of wealth sharing, a tax-funded digital dividend that would give displaced workers a direct financial stake in the economy rather than handing Washington shares in Anthropic itself. It is a meaningful difference. A tax applies to everyone equally. An equity stake makes the government a part owner of the very company it regulates, and critics have already flagged the conflict that creates. If a rule Washington writes lowers OpenAI's valuation, it also shrinks the government's own investment.
The timing lines up with both companies' IPO calendars. OpenAI and Anthropic have each filed confidentially with the SEC this year, and analysts warn a pre-listing government stake cuts two ways. Some investors will read it as Washington de-risking the company before the float. Others will price it as a permanent governance headache attached to every share they buy.
That tension has not stopped the idea from finding an audience. A Verasight survey of 1,690 adults, conducted in June and amplified by CNBC this week, found that 69% of Americans support forcing AI companies to transfer half their stock, not 5%, into a public fund. Sanders posted the figure on Sunday, framing it as validation for his American AI Sovereign Wealth Fund Act, the bill this newsletter covered as a long shot back in issue #025. It is still a long shot in this Congress. It just polls better than most legislation does.
Not everyone thinks OpenAI's version goes far enough, and not everyone thinks it should exist at all. Former Trump strategist Steve Bannon dismissed the 5% offer as "tip money" and wants Washington to force firms to hand over half their equity outright. Public Knowledge policy advocate Nat Purser raised the opposite worry, that a government holding both regulatory power and a financial stake in OpenAI's success is compromised no matter the size of the stake.
Context helps explain the timing. Tech layoffs made up close to a third of all US job cuts in the first half of 2026, even as the companies doing the cutting posted record profits. Oracle cut 21,000 positions while its earnings jumped 27%. Meta cut 8,000 jobs while raising its AI capital spending guidance toward $135 billion. Workers can see both numbers on the same chart, and the polling suggests they have already drawn their own conclusion.
My take: Watching OpenAI offer 5% while the public polls for 50% tells you the negotiation has already started, whether or not either side calls it one. Altman's version is the cheaper, faster, friendlier way to get ahead of a mandate instead of waiting for Congress to write one, and it comes with the added benefit of buying goodwill right before an IPO. Anthropic's refusal to join the equity talks is its own kind of bet, that a tax-funded dividend is both fairer and less entangling than a government financial stake in a company Washington also has to regulate. I do not think either company is being purely altruistic here, and I do not think they need to be for the public conversation to matter. A year ago, giving the public a piece of the frontier was a fringe idea from one senator. Now it is something OpenAI feels compelled to offer before anyone forces its hand.
While OpenAI negotiates for a stake in itself, it also became the second company to test a very different kind of government reach over the frontier, the voluntary review process born out of Anthropic's June shutdown. On June 26, OpenAI previewed its new GPT-5.6 model family, called Sol, Terra, and Luna, but held broad release back at the request of the White House. Only about 20 government-vetted partner organizations got early access, chosen by the Office of the National Cyber Director and the Office of Science and Technology Policy.
The trigger was cybersecurity. Sol, the flagship, scored 96.7% on OpenAI's internal cyberattack evaluation and performed on the ExploitBench benchmark roughly in line with Anthropic's tightly restricted Mythos Preview model, at about a third of the inference cost. That is exactly the kind of vulnerability-finding capability that took Fable 5 and Mythos 5 offline for eighteen days last month. Commerce's Center for AI Standards and Innovation ran additional tests, OpenAI sent technical staff to Washington, and on July 9, thirteen days after the gated preview, the review cleared and Sol, Terra, and Luna went fully public.
Thirteen days is the number worth sitting with. The executive order behind this process, signed June 2, gives companies up to 30 days of review before release, down from a 90 day window in an earlier draft. OpenAI's case cleared in under half that time, through a process both sides describe as collaborative rather than adversarial. Compare that to Anthropic's experience three weeks earlier, when a Commerce Department export control order arrived by letter on a Friday evening with no advance testing period, forced a worldwide shutdown of two model lines, and took eighteen days and considerable public frustration to resolve.
OpenAI is not thrilled about any of it, gated or not. The company has said publicly it does not want this kind of access process to become the permanent default, arguing it keeps useful tools away from the developers, enterprises, and cyber defenders who need them. It agreed to the delay anyway, calling it the fastest path to a broader release, and it is already bringing Sol to Cerebras hardware for select customers as a separate, faster track.
My take: Put Anthropic's June and OpenAI's July side by side and you get a preview of what voluntary oversight actually looks like once a lab plays along instead of getting steamrolled. Thirteen days, a named review body, direct technical meetings, and a public timeline is a completely different experience from a Friday-evening letter and a worldwide blackout. That is either reassuring or unsettling depending on your priors. It is reassuring if you think a repeatable process beats an ad hoc punishment. It is unsettling if you notice that the word voluntary is doing a lot of work for a review every frontier lab now has to build into its launch calendar. The checkpoint from issue #028 did not just outlive the crisis that created it. This week it got its first calm, uneventful trial run, and that may be the more important story of the two.
Look at everyone who tried to put a formal claim on the frontier this week and a pattern shows up that has nothing to do with model quality. The public wants equity, or at minimum a mandate that sounds like it. Washington wants a repeatable checkpoint every major release has to clear, and this week showed what that checkpoint looks like when nobody is fighting it. Alibaba wants both extraction and immunity at once, harvesting Claude's outputs through fake accounts while banning the real thing from its own building. None of these are requests for a better model. They are all requests for a say.
The strange part is how normal it all felt from inside the labs. Anthropic put Cowork on your phone. OpenAI folded Codex into a new desktop app and called it ChatGPT Work. Both companies spent the week shipping product updates as if the equity fights, the checkpoints, and the export bans were background noise rather than the actual story. Maybe that is the healthiest way to run a company under this much scrutiny. Maybe it is also how you end up agreeing to give away 5% of yourself before anyone forces the number on you.
What ties Sanders' bill, OpenAI's offer, Washington's checkpoint, and Alibaba's ban together is the same fact this newsletter keeps returning to. The frontier labs do not set the terms anymore, if they ever fully did. The state sets a release calendar. The public is starting to demand a balance sheet. A rival's largest tech company sets its own rules for access in both directions at once. Being the smartest model on the leaderboard increasingly buys a lab very little say in who gets to use it, who gets to copy it, or who gets to own a piece of it.
While Washington argues over who owns a piece of OpenAI, North Carolina is still arguing over who pays for the physical thing that makes AI run. On July 7, the NC Utilities Commission opened hearings on Duke Energy Carolinas' rate case, and the numbers on the table are smaller than a $42 billion stake but a lot more real for the people footing them. Duke already cut its original ask from an 18% residential increase to 11.6% after Attorney General Jeff Jackson and the Public Staff objected, splitting it into 7.5% in 2027 and 4.2% in 2028. Jackson says that is still too high. Commissioners spent the hearing pressing Duke on a detail that lands hardest on low income households, that the Customer Assistance Program's $42 monthly credit expires December 31 unless renewed, which one commissioner noted could mean an effective $52 jump in a single month if the rate increase and the expiration hit at once.
The Ratepayer Protection Act, the bill that would make data centers pay for their own grid upgrades instead of spreading the cost to everyone else, has now sat in the Senate Rules and Operations Committee since June 8 with no floor vote scheduled. A Carolina Journal poll taken as the bill moved through the House found 78% of NC voters want data centers to generate their own power rather than draw off the shared grid, a number not far from the national polling on AI equity even though almost nobody answering it was thinking about anyone's cap table. Environmental justice groups organized a Data Center Lobby Day at the General Assembly on June 24 to push back on the bill's other half, the provision that lets Duke keep aging coal plants running until new nuclear capacity comes online. In rural Edgecombe County, where residents formed a group called Neighbors for Data Center Accountability, that tradeoff is not abstract. A separate bill working through the Senate, House Bill 1213, would end the property tax break for new utility scale solar starting in mid-2027, making the cheap clean power that might otherwise serve new data centers harder to build in the first place.
None of this moves as fast as a Sam Altman press tour, and that is sort of the point. The equity fights happening in Washington this week are still mostly promises on paper. The rate case in Raleigh already has real numbers, a hearing date, and a decision expected by fall. If you want to know what the frontier actually costs, skip the valuation headlines and read a Duke Energy filing instead.
That's the week the tenants asked for equity. See you next Wednesday.
Daniel
BullCity AI · Durham, NC
P.S. If your company has already run a real workload through GPT-5.6's Sol, Terra, or Luna tiers, or moved anything onto Claude Cowork's mobile beta, hit reply and tell me if the marketing matched the experience. I'm collecting before-and-after stories for a future issue.
P.P.S. Forward this to the colleague who thinks a government stake in OpenAI sounds far-fetched. As of this week, 69% of the country wants Washington to take ten times what OpenAI actually offered.