AI in 15 — July 03, 2026

OpenAI wants to hand the United States government a stake in itself worth forty-two billion dollars. And it wants Google, Meta, and Anthropic to do exactly the same thing.

Welcome to AI in 15 for Friday, July third, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

All week we've been living inside the question of who controls access to these models. Today the question flips to something bigger — who owns them. That's our lead, and it's a doozy. Then a story that'll matter to anyone who's ever hired a freelancer.

Fable 5 is back — but the security community is furious about the leash it's wearing.

Meta decides to become a landlord for its own spare computing power.

And Anthropic's revenue quietly triples while everyone's watching what it won't ship.

Lead story, Marcus. Walk me through this. OpenAI wants to give Washington a piece of the company?

Roughly five percent, Kate — that's the reporting, first from the Financial Times, then confirmed across Bloomberg, CNBC, and TechCrunch, all dated yesterday. Against the eight hundred and fifty-two billion dollar valuation from OpenAI's March funding round, five percent is about forty-two-and-a-half billion dollars. And Sam Altman reportedly raised it directly — with President Trump, with Commerce Secretary Lutnick, with Treasury Secretary Bessent. His framing was that public ownership is, quote, the fairest way to share AI's upside.

Okay, but here's where it stops being just an OpenAI story, right?

That's exactly it, Kate. The pitch is industry-wide. Altman wants every leading American lab — Anthropic, Google, Meta — to cede a similar five percent slice into a sovereign-wealth-style fund. And the model he keeps pointing to is the Alaska Permanent Fund. That's the thing that takes Alaska's oil revenue, invests it, and cuts every resident an annual dividend check. OpenAI even laid the groundwork back in April, in a policy paper called "Industrial Policy for the Intelligence Age" — the idea being fund returns flow straight to citizens.

So the oil this time is the AI itself.

That's the poetry of it, Kate. Treat frontier AI as a national resource, and pay the public a dividend on it. And interestingly, Altman's even floated this with Bernie Sanders — though Sanders wants public ownership closer to half of each lab, and in June introduced his own bill proposing a fifty percent one-time tax on AI equity. So there's a live debate here, not one plan.

Now, cynical question. Why would OpenAI volunteer to give away forty-two billion dollars?

Because Washington has gotten colder, Kate, and goodwill is suddenly worth buying. Think about the week we've had — OpenAI and Anthropic both had frontier releases held up by US scrutiny. A government stake could thaw that. And there's a template: Intel. Trump swung hard behind the chipmaker right after the government took a ten percent stake in it. So you can read this as OpenAI trying to convert itself from a target into a partner.

But there's an obvious problem, isn't there.

The obvious problem, Kate, is that the government would become a shareholder in the very companies it's supposed to regulate. When your overseer is also your part-owner, whose interest is it protecting — the public's, or its own equity stake? That line gets very blurry. And I'd add the honesty caveat: OpenAI itself calls these talks "conceptual." An actual deal might need an act of Congress, and the rivals — Anthropic, Google, Meta — could simply refuse to play along. Right now this is a trial balloon, not a signed term sheet.

So the headline is the ambition, not the deal.

Exactly, Kate. But even as a trial balloon, it tells you where the frontier of the argument has moved. Six months ago the fight was "how do we regulate these labs." Now Altman is asking "how do we co-own them." That's a genuinely different conversation.

Next, Marcus, and this ties off a thread we've pulled all week. Fable 5 is back — but not everybody's happy about how.

Right, so quick recap for anyone joining us, Kate — Anthropic's most powerful model, Fable 5, came back worldwide on July first after a nineteen-day suspension, once the US export controls lifted. We covered the return. What's new today is the backlash to how it came back. Fable is now wrapped in a new cybersecurity classifier that blocks the specific jailbreak in over ninety-nine percent of cases, and when it trips, it quietly reroutes you down to the older, weaker Opus 4.8.

And the security community's reaction to that?

Not warm, Kate. Researchers say the filter is wildly over-broad — that Fable now reflexively refuses anything even loosely security-adjacent. IBM X-Force's Valentina Palmiotti reportedly said it rejects any request that could be even tangentially cyber-related, down to just reading a blog post. Others say a routine code review sets it off. And here's the part I appreciate — Anthropic actually concedes it. They admitted the classifier flags benign requests too often during normal coding and debugging, and told Wired, in so many words, that they "made the wrong tradeoff."

Why does that matter beyond a few annoyed researchers?

Because the people it slows down are the defenders, Kate. The whole point of these tools in security is that the good guys use them to find and patch vulnerabilities before the bad guys exploit them. An over-cautious filter doesn't stop attackers — they'll use an unrestricted model. It stops the defender doing a legitimate code review. So you get security theater that taxes exactly the wrong side.

And the irony you flagged?

The irony, Kate, is that on today's automation leaderboard — which is literally our next story — Fable 5 is the single model doing the most real work of any model alive. So the most capable thing out there is also the one on the tightest leash. The bottleneck has fully moved from engineering to governance. The models look ready. Permission to actually use them freely is the scarce thing now.

So let's go straight to that number, Marcus, because I think it's the most concrete thing in the whole brief. AI now finishes one in six real freelance jobs?

It's called the Remote Labor Index, Kate — a benchmark from the Center for AI Safety and Scale Labs. And what I like about it is it doesn't test puzzles. It grades AI agents on two hundred and forty real, paid freelance projects — 3D and CAD work, architecture, graphic design, video, audio, data analysis, web apps — against what a professional would actually deliver. And Fable 5 just posted the highest score ever: it automated sixteen-point-one percent of those projects end to end.

Sixteen percent. Put that in context for me.

It's nearly double the next model, Kate — Opus 4.8 at eight-point-three percent, GPT-5.5 at six-point-three. And everyone else — Gemini 3 Pro, Grok 4, GPT-5.2 — is still stuck in the low single digits, one to two-and-a-half percent. But the real story is the trajectory. When this benchmark launched back in October, the leader managed just two-and-a-half percent. In under eight months it's more than quadrupled. From automating one in fifty projects to about one in six.

That's a genuinely steep curve. Okay — keep me honest, Marcus. Where's the catch?

The catch matters and I want to state it plainly, Kate. Even the leader fully automates only sixteen percent — the other eighty-four percent still need a human. And the Center for AI Safety says something striking about its own result: none of the three sample Fable 5 deliverables it published would actually be accepted as finished work. So it's scoring completed gigs, not whole jobs. It tracks encroachment — the water rising — not wholesale replacement.

But rising water is still the story people feel.

It is, and it pairs with two things from the same week, Kate. A California workforce study showing AI job pressure climbing into educated, high-exposure white-collar roles — concentrated, tellingly, in the Bay Area. And Palo Alto Networks CEO Nikesh Arora telling his own staff they're facing a, quote, Darwinian moment — adapt or get cut. So capability is no longer the scarce thing. The fight now is over who captures the gains and who absorbs the disruption.

That's a heavier note. Let's shift to the business side.

Meta, Marcus. It's getting into the cloud business? I thought Meta sold ads.

It does, Kate — and that's precisely the itch it's trying to scratch. The reporting, from Bloomberg and CNBC, is that Meta is building a cloud business — reportedly called Meta Compute — to sell access to its surplus AI computing power and its hosted models to outside developers. Two tiers: one where you run AI models on Meta's servers, and one where you just rent raw GPU capacity, neocloud-style. And Meta's stock popped about nine percent on the news.

Why build this now?

Because Meta guided to a hundred and twenty-five to a hundred and forty-five billion dollars of capital spending this year, Kate — data centers, chips, the whole buildout. And apparently that buildout is now running ahead of Meta's own internal demand. So rather than let expensive silicon sit idle, they'll rent it out. It's an attempt to turn a giant cost center into a revenue line — and to diversify away from being purely an ad company.

So this puts Meta up against Amazon, Microsoft, Google's cloud.

Head to head with the hyperscalers, Kate. Which is a bold place to plant a flag. But here's my caveat — there's no pricing, no launch date, no named customers yet. So treat this as strategic intent, not a live product you can buy today. What's genuinely revealing is the admission underneath it: even Meta built more compute than it can use. That tells you something about the scale of this arms race.

Quick hits, Marcus. And this first one is a nice counterweight to all the "labs tapping the brakes" talk. Anthropic's revenue.

It's a striking number, Kate. Anthropic's annualized run-rate revenue has reportedly crossed thirty billion dollars — up from roughly nine billion at the end of 2025. And the detail I'd underline: more than a thousand customers are now each spending over a million dollars a year, and that group has doubled in under two months. So while the public story all week has been about restraint — no Opus 5, Fable on a leash — the business underneath it is quietly on fire.

So the "tapping the brakes" narrative and a tripling revenue line, side by side.

That's the tension worth sitting with, Kate. A company can publicly preach caution about what it ships and still be selling everything it can make. Restraint on the frontier model and a rocket ship on the balance sheet are not contradictory — they're the same company managing two very different audiences.

And one more, Marcus — the environmental bill for all this compute. Google put out numbers.

And they're sobering, Kate. Google's 2026 environmental report shows its electricity demand jumped thirty-seven percent in a single year, and water use rose thirty-four percent, to nearly eleven billion gallons. Now, to Google's credit, they added a record twelve gigawatts of new clean energy and ran on eighty-six percent carbon-free electricity. But even with all that, operational emissions only fell two percent.

So the green efforts are real, but the growth is outrunning them.

That's the whole shape of it, Kate. Google is building clean power faster than almost anyone on the planet — and it's still barely holding the line, because the AI buildout is growing faster than it can green it. That's the physical cost sitting underneath every one of these capacity stories. The chips need power, the power has a footprint, and right now demand is winning that race.

One to watch tomorrow, Marcus.

I'm going off the AI beat for this one, Kate, because it's too good to skip. A University of Minnesota team, led by synthetic biologist Kate Adamala, says it has built the first cell from scratch — from nonliving chemicals — that can feed, grow, copy its own genome, and divide. They're calling it SpudCell, ninety thousand base pairs across seven plasmids, released open-source. It's not peer-reviewed yet, and the AI link is thin — it runs through machine-learning "aging clocks." But if it holds up, it's genuinely stunning. Watch whether it survives scrutiny.

Agree, or counter?

Slight counter, Kate. The nearer-term signal is whether any rival lab publicly answers Altman's five-percent proposal. That loaded question now sits with Anthropic, Google, and Meta — and the first one to say yes or no tells you whether this idea has legs or dies as a headline.

That's your AI in 15 for today. See you tomorrow.