AI Actually
Issue No. 24 · Sunday, July 5, 2026
Happy long weekend! Hope you’ve celebrated the 4th and enjoyed the fireworks with your loved ones. For today, it’s a short one : two weeks after AI’s most dramatic swing — a flagship model getting switched off overseas, then switched back on — everyone’s circling back to ask what the whole episode actually proved. Turns out: quite a lot.
Uncle Sam wants a piece of ChatGPT
OpenAI has reportedly proposed giving the U.S. government a 5% stake in the company — worth roughly $42.6 billion at its current $852 billion valuation. The pitch, first reported by the Financial Times, would fold that equity into something like a sovereign wealth fund: a pot the public owns a slice of, the way Alaskans get a check every year from the state’s oil money.
The framing is generous — “share AI’s upside with everyone” — but the mechanics are a little more interesting than the press release. Altman has apparently floated the idea that other major labs, including Google and Meta, do the same. And this isn’t the administration’s first rodeo: it already holds a 10% stake in Intel and takes a cut of Nvidia and AMD’s chip sales to China. A government that owns equity in the companies it regulates is a genuinely new arrangement for the AI industry, and not an obviously clean one.
Here’s the awkward part nobody’s saying out loud: the same officials who’d decide what OpenAI is allowed to release would also, eventually, have a financial stake in OpenAI doing well. That’s not a conflict of interest so much as a conflict of the entire job description.
Why it matters. A 5% stake sounds like a gift to taxpayers, and maybe it is one. But it also quietly redraws the relationship between AI labs and the government overseeing them — from referee to shareholder. Whether that’s the safety net the public actually wants, or just a very expensive PR move, depends entirely on who ends up holding the shares: households, or Washington. Those are two very different bets.
The AI you rent can be turned off. So companies are buying instead.
Quick rewind: on June 12, the U.S. Commerce Department sent Anthropic a letter, and by the next morning, its two newest models — Fable 5 and Mythos 5 — were dark everywhere on Earth, including for the hospitals and researchers the order was never aimed at. Nineteen days later, access was restored. That story’s closed. But the aftershock is just getting started, and it’s showing up in boardrooms, not government offices.
On CNBC last week, Palantir CEO Alex Karp said the quiet part out loud: enterprises are “livid,” paying frontier labs for “tokens that create no value” while, in his words, the labs are “stealing their weights and alpha” — the shorthand for a company’s proprietary data and competitive edge. His argument: if a government can flip a switch and kill your AI vendor overnight, and a business vendor can just as easily change the price or the model underneath you, then renting the frontier was never actually the safe choice. Owning your own stack is.
Companies seem to be listening, or at least hedging their bets — this week alone, Microsoft launched a $2.5 billion, 6,000-person unit to embed engineers directly inside client companies and build AI systems in-house, following nearly identical moves already made by OpenAI, Anthropic, and Amazon. Everyone, apparently, wants to own the means of production now — even the companies that sell the tokens.
Why it matters. For years, the pitch for renting a frontier model instead of building your own was simple: it’s cheaper and someone else does the hard part. The Fable 5 shutdown didn’t change the economics, but it did change the fear — a vendor that a foreign government can shut off with a Friday-afternoon letter isn’t really a vendor. It’s a dependency. Expect “who owns the weights” to become a normal line item in procurement contracts, not just a Palantir talking point.
Safe to ignore this week
Sonnet 5’s launch and the Fable/Mythos comeback — we covered both in full on Wednesday. Old news, deliberately.
The AI Engineer World’s Fair “software factory” debates — a real conference having a real argument about AI-written code, entirely in a language only conference attendees speak.
Anthropic’s early chip talks with Samsung — interesting if it happens. It hasn’t happened yet.
A newsletter claiming Meta’s next model “caught up” to GPT-5.5 — one outlet, one unverified brag, zero benchmarks.
Gemini’s incremental Flash upgrade — a good model got slightly better. We’ll let you know when it gets a lot better.