ALTMAN PITCHES WASHINGTON A PIECE OF THE PIE — AND WANTS EVERY AI LAB TO DO THE SAME
Sam Altman walked into the White House and proposed something that sounds like a riddle: what if the company that might one day make you obsolete could also pay your mortgage? OpenAI is reportedly pitching the idea of handing the US government a 5% stake in the company, which at the current $852 billion valuation works out to roughly $42.6 billion worth of equity sitting in the hands of Uncle Sam. And Altman does not just want this for OpenAI. He reportedly wants every major AI lab in the country to do the same thing, with all of it flowing into a fund modeled on the Alaska Permanent Fund, which sends annual checks to every Alaskan resident simply for living there.
This is the kind of idea that sounds absurd until you sit with it for five minutes and realize it might actually be clever. OpenAI has been under serious pressure from Washington ever since the GPT-5.6 export control situation last week. When the government delays the global rollout of your most powerful model over national security concerns, you have a problem. When they can also hold up future releases and threaten to cap where your technology goes in the world, you have a much bigger problem. So what do you do? You offer to cut them in.
Think about the logic here. If the US government owns 5% of OpenAI, suddenly OpenAI’s success is Washington’s success. When Altman goes to testify before Congress, he is no longer just a tech CEO trying to avoid regulation. He is showing up to talk about an asset the government directly owns. That changes the dynamic completely. Politicians who want to crack down on AI are now politicians who want to crater the value of a national asset. Nobody in Congress wants to own that vote.
The Alaska comparison is also not an accident. The Alaska Permanent Fund is one of the most broadly popular government programs in American history because it literally sends people money. Altman is essentially saying: imagine if AI made every American slightly richer every year just by existing. That is a political argument, not a business one, and it is aimed directly at the people currently deciding how hard to regulate him.
The details are still fuzzy. The Financial Times, which broke the story, described the discussions as conceptual and early stage, which in Washington speak means somebody floated this at dinner and now it is in the newspaper. Any actual deal would likely need Congress to act, which means it could take years or go nowhere at all. There is also an obvious question about whether Anthropic, Google, or Meta would agree to a similar arrangement. None of them have said yes yet.
But even if this never closes, the fact that it was proposed changes the conversation. Altman is reframing AI not as a threat to be regulated but as a public resource to be invested in. That is a genuinely smart move, and you can bet the other labs are watching closely to see how Washington responds.
FABLE 5 IS BACK AFTER 19 DAYS IN THE GOVERNMENT PENALTY BOX
Source: VentureBeat | The Hacker News
On June 12, something unusual happened in the AI industry. The government pulled a product off the market. Not because it was dangerous in the traditional sense, but because researchers at Amazon had found a jailbreak in Anthropic’s Fable 5 model, the most powerful Claude to date, and the Department of Commerce decided that was enough reason to require an export control license before Anthropic could offer it to anyone outside the United States.
Anthropic, facing the impossible task of instantly screening millions of users by country of residence, did what any company in that position would do: it turned the model off for everyone. For 19 days, Fable 5 was simply gone. Not degraded, not restricted. Gone.
If you were an enterprise customer who had built workflows around Fable 5, those 19 days were a wake-up call delivered at full volume. Entire legal and finance teams were suddenly working with an older model. Customer service pipelines built on the more capable version reverted to something slower and less accurate. The shutdown taught every enterprise customer in the world a lesson they did not ask for: AI is now infrastructure, and infrastructure can be taken away by a government memo with about 48 hours of notice.
On July 1, the Department of Commerce lifted the export control restriction. Commerce Secretary Howard Lutnick sent a letter withdrawing the requirement, Anthropic flipped the switch, and Fable 5 came back online for global users. It returned with something new attached: a cybersecurity classifier designed to detect attempts to extract sensitive security-related outputs. That classifier is, in essence, the direct response to the jailbreak that caused all of this.
A few things worth noting here. First, the original jailbreak was found by Amazon researchers, and Amazon is both a cloud partner of Anthropic and one of its biggest investors. The awkwardness of your own investor discovering the flaw that got your model banned has a certain dark comedy to it. Second, Anthropic added new restrictions even after getting the model back. Fable 5 now has guardrails it did not have before, which makes it technically safer but also somewhat more limited for legitimate security researchers who need to probe AI systems for vulnerabilities.
Third, and this is the part that matters most going forward: a precedent has been set. The government now knows it can require export controls on AI models, and it knows companies will comply quickly. Whether that authority gets used thoughtfully or recklessly is entirely up to whoever is running the Commerce Department at any given moment. That is either reassuring or terrifying, and either way it is now a permanent feature of the AI landscape that every enterprise buyer needs to price into their planning.
NEWSOM GIVES EVERY CALIFORNIA GOVERNMENT WORKER AN AI ASSISTANT — ALL RUNNING ON CLAUDE
If you want to understand how fast AI is moving into government, look at what California just announced. Governor Gavin Newsom signed a deal with Anthropic that gives every single state agency in California, plus every city and every county in the state, access to Claude at a 50% discount. That is roughly 200,000 state employees, plus local government workers across the most populous state in the country, getting access to one of the most capable AI assistants available starting right now.
The deal flows through something called the Statewide Information Technology Shared Services portal, which is already up and running. The first use cases are the obvious ones: drafting documents, analyzing policy language, handling internal workflows that eat up hours of government worker time every week. The California DMV has already been running Claude for customer service improvements. The Department of Health Care Services has been using it for internal operations. This deal just makes both of those programs official and scales them statewide.
Newsom has been an interesting politician to watch on the AI question because he has managed to be both skeptical and enthusiastic depending on the audience. Last year he vetoed a major California AI safety bill, arguing it was too aggressive and would push AI companies out of the state. Now he is signing a deal to put one of those same companies’ products inside every arm of state government. The critics who said he killed the safety bill to protect the industry he was about to partner with now have some material to work with.
That said, the deal itself is not unreasonable on its face. Government agencies across California are drowning in paperwork and manual processes that have not changed in decades. If Claude can cut the time it takes a state worker to process a routine document from three hours to thirty minutes, that is a real improvement in how government actually functions. The 50% discount also suggests Anthropic is treating this partly as a flagship reference deal. If California rolls this out successfully, every other state government in the country starts calling next quarter.
The more interesting piece is a program called Poppy, a state-built AI assistant designed specifically for government workflows. Poppy has already piloted with over 2,800 state employees across 67 departments and is on track for statewide rollout this month. Poppy is built on top of Claude, which means what California is really doing is building its own government AI product using Claude as the engine and deploying that product to every government worker in the state.
That is a different kind of AI adoption than what we usually talk about. This is not a startup buying enterprise seats. This is a state of 40 million people deploying AI through its government machinery. We are about to learn a lot about what institutional AI adoption at genuine scale looks like, and California is going to be the case study.
THE STARTUP THAT CLONES YOUR VOICE JUST DOUBLED ITS PRICE TAG IN FIVE MONTHS
In February 2026, ElevenLabs raised $500 million at an $11 billion valuation. It is now July 2026, five months later, and the company is exploring a secondary offering that would value it at $22 billion. It doubled. In five months. That sentence deserves a moment to breathe.
To understand why that is either remarkable or troubling depending on your perspective, you need to understand what ElevenLabs actually does. The London-based company builds AI voice technology: text-to-speech, voice cloning, audio generation, the kind of tools that can take a script and produce audio that sounds indistinguishable from a specific human being saying those words. The applications range from audiobooks and customer service automation to the kind of synthetic audio that keeps election security researchers awake at night thinking about what happens when nobody can tell real from fake anymore.
The business case is clear. Voice is everywhere. Customer service calls, podcasts, audio advertising, video narration, educational content, accessibility tools, video game characters, navigation systems. Every piece of recorded speech that exists represents a potential use case for ElevenLabs technology. The company crossed $500 million in annual recurring revenue in just the first four months of 2026, up from $350 million at the end of 2025. That is a serious jump in a short window of time.
A tender offer, which is what is being explored here, works by letting current employees sell some of their shares to outside investors. This is a common mechanism among AI startups right now because it lets a company reward and retain talent without going public. Employees get liquidity, outside investors get access to a fast-growing company before any IPO, and the company itself does not have to deal with the chaos of public markets while it is still in a growth phase. Everyone gets something they want, at least on paper.
The doubling of the valuation in five months is where things get philosophically interesting. Either the company’s growth genuinely justifies a $22 billion price tag, or the AI market is repricing everything upward in a way that will look embarrassing in retrospect. Honest answer: probably some combination of both. ElevenLabs has real revenue growing at a real rate. It also exists in an environment where investors are paying premium prices for anything AI-adjacent, and the premium has been getting higher every quarter without pause.
What makes ElevenLabs worth paying attention to is that voice AI remains the most underrated segment of the entire artificial intelligence landscape. Text generation gets all the coverage because it is the most visible thing AI does. But voice is where the real volume of human-machine interaction is heading. ElevenLabs is not a niche player. It is positioning itself as core infrastructure for an audio internet that is just getting started. If the $22 billion tender closes, you will have a company that barely existed three years ago now worth more than most companies on the Fortune 500. That is the kind of fact that requires you to either update your assumptions about what AI is worth or brace for a significant correction. Possibly both.
WASHINGTON IS WRITING AI RULES — AND IT IS POLITELY ASKING COMPANIES TO FOLLOW THEM
Source: Yahoo Finance / Financial Times | White House
Washington is getting serious about AI rules, which in practice means Washington is having a lot of meetings where serious people discuss what serious rules might eventually look like.
Here is the situation as of this week. The US government is in advanced talks with OpenAI, Google, and Anthropic about creating voluntary standards for releasing new AI models. An announcement could come as soon as next week. The framework being discussed would require companies to submit their most powerful models to the government for testing up to 30 days before releasing them publicly, giving agencies time to evaluate what the models can actually do before they hit the open market and start reshaping industries.
This follows a Trump executive order from June that laid out a similar structure on paper. The order called for a classified benchmarking process to assess the advanced capabilities of AI models, particularly around cybersecurity vulnerabilities, and established a threshold above which a model would be considered a covered frontier model requiring this kind of pre-release review. The executive order gave the framework a legal basis. These talks are about getting the companies to actually show up and participate.
The word “voluntary” is doing a lot of work in this story and it deserves scrutiny. Voluntary standards in an industry where the major players are also the entities helping write those standards have a complicated track record. The companies that agreed to voluntary AI safety commitments in 2023 under the Biden administration faced essentially zero enforcement and varying degrees of actual compliance. Whether this round is different depends entirely on how the government structures the consequences for companies that choose not to participate, which is something the public reports have been notably quiet about.
There is also a timing element here that matters. The Fable 5 export control situation demonstrated that the government already has blunt instruments for intervening in AI deployments when it decides there is a security concern. The question is whether you want to keep using that blunt instrument on a case-by-case basis or build a more transparent system that creates real-time visibility before problems become public crises. The voluntary standards discussion is the more nuanced option, at least in theory.
What is interesting about the current moment is that this conversation is happening at all and that the companies appear to be genuinely engaging with it rather than lobbying against it. A year ago, the idea of AI companies submitting their models to government review before launch would have generated significant industry pushback. The fact that it is now described as advanced talks with an imminent announcement suggests the labs have decided that a predictable review process is preferable to the unpredictable alternative. Which, given what just happened with Fable 5, is probably the rational conclusion.
Whether the framework that eventually emerges has any enforcement mechanism is the thing to watch. Voluntary, well-designed standards with no consequences attached are still just suggestions. And in an industry moving at this pace, suggestions tend to age poorly.