Last week the story was price. This week it's power, as in who has the frontier and who has to rent it.
On Monday, Apple walked on stage at its final Tim Cook keynote and unveiled a rebuilt Siri that thinks with Google's model, not its own. On Tuesday, Anthropic shipped its most powerful public model yet, five days after telling the industry it should build a way to slow down. SpaceX priced the largest IPO in history. And a US senator proposed simply taking half of the big AI companies and handing them to the public.
Pull it together and one shape emerges. The frontier is now owned by a handful of labs, and almost everyone else, even the most valuable hardware company on earth, is becoming a tenant. This week Apple signed the lease.
Apple opened WWDC on Monday with the announcement it had been promising and delaying for two years, a Siri that finally works like a modern assistant. It can hold a real back-and-forth, read what's on your screen, pull details out of your messages and photos, and take actions across your apps. The headline, though, wasn't any single feature. It was the engine. The rebuilt assistant, now called Siri AI, runs on a custom version of Google's Gemini, a model reported at roughly 1.2 trillion parameters, running inside Apple's own data centers. Apple is paying Google somewhere around a billion dollars a year for it, and analysts put the full multi-year deal at up to $5 billion.
Sit with how strange that is. Apple has spent more than a decade insisting it does the hard things itself, on its own chips, with its own software, behind its own walls. It evaluated proposals from OpenAI and Anthropic, and then it picked Google. The most valuable hardware company on the planet looked at the cost of building a frontier model in time and decided to rent one instead. The deal was actually struck back in January. Monday was the moment it walked on stage.
Apple spent a chunk of the keynote insulating itself from the obvious privacy question. The Gemini model runs on Apple's Private Cloud Compute servers, the company says no user data is shared with Google, and the model weights sit on Apple's own hardware rather than Google's. Craig Federighi put it bluntly, that privacy in AI is "non-negotiable." Whether that framing survives contact with regulators is a separate question.
The money flow is the part I can't stop thinking about. Google already pays Apple something like $20 billion a year to be the default search engine on the iPhone. Now the cash runs both directions, with Apple paying Google about a billion a year for the intelligence behind Siri. Two of the biggest companies in the world are quietly renting from each other. And it cements the comeback story of the year, because Gemini now powers Google's own products and Apple's flagship feature at the same time.
There was a second, quieter headline. This was Tim Cook's last WWDC. He hands the company to hardware chief John Ternus on September 1 and moves to executive chairman. It's a fitting bookend. The CEO who turned Apple into a four-trillion-dollar machine is walking off the stage at the exact moment Apple concedes the one race it could not win on its own.
A few caveats keep the hype honest. Siri AI launches as an English-only beta later this year, it won't ship in China while Apple works through regulators, and it won't be available in the EU at launch, where the Digital Markets Act is already circling Google's double role as both Apple's search and AI supplier. And plenty of people watching wondered the obvious thing, which is whether a great Siri in late 2026 is simply too late after ChatGPT and Gemini spent two years redefining what people expect.
My take: Apple just told us the most important thing about this era without quite saying it. Building a frontier model has gotten so hard and so expensive that the company with the most cash, the best silicon, and two billion devices decided renting beats building. That isn't weakness, it's arithmetic. Apple spends around 3% of revenue on capital projects while Google is pouring something like $90 billion into AI this year. But it does change what Apple is. For fifteen years the whole pitch was that Apple controlled the entire stack, top to bottom. Starting this fall, the most personal feature on your phone will think with a competitor's model. The thing to watch isn't the demo. It's whether the promise that your data never leaves Apple's servers holds up once regulators and researchers start pulling on it.
The sequence here is the whole story. On June 4, Anthropic published a report called "When AI Builds Itself," urging the industry to build what its co-founder Jack Clark called a "brake pedal," a coordinated, verifiable way to slow or pause frontier development if models start improving themselves without humans in the loop. Clark compared the moment to Cold War arms control, the idea being that even bitter rivals have managed to agree on ways to pull back before. To make the point concrete, the report noted that more than 80% of the code merged into Anthropic's own codebase is now written by Claude, not its engineers, and that those engineers now merge roughly eight times as much code per day as they did in 2024.
Five days later, on Tuesday, Anthropic shipped Claude Fable 5, the most powerful model it has ever released to the public. It's the first of a new tier the company calls Mythos-class, which sits above its Opus models. If the name rings a bell, it should. Mythos is the model Anthropic held back in April because it was unsettlingly good at finding software vulnerabilities, the same capability we watched an attacker actually use in last week's Sysdig story. Just last week the company had widened access to Mythos to hundreds of organizations across 15 countries, most of them tied to critical infrastructure. Fable 5 is the softened version it decided was safe enough to sell to everyone.
By Anthropic's own numbers it leads every other Claude on nearly every benchmark, pulls further ahead the longer and more complex the task, and can work on its own for longer than any prior model. It runs about $10 per million input tokens and $50 per million output, double the price of Opus. It's free on the Pro, Max, Team, and Enterprise plans through June 22, after which using it burns usage credits until capacity catches up.
The safety story is bolted on at the edges. In the riskiest areas, cybersecurity, biology, chemistry, and teaching other models, Fable 5 refuses and quietly hands the question to the older, weaker Opus 4.8. Anthropic says those guardrails kick in less than 5% of the time. The fully unlocked version, Mythos 5, only goes to vetted partners through Project Glasswing, the group that includes AWS, Google, Microsoft, Cisco, JPMorgan, and, yes, Apple.
So in a single week Anthropic told the world to design a brake pedal, then floored the accelerator. Scientific American captured the skeptical read cleanly, noting that to critics these warnings can look like strategy, a way to invite regulation onto the frontier while Anthropic keeps sprinting toward it. The same skeptics question whether the company has overstated what its systems can really do. One researcher called a real pause flatly impossible. And all of it lands while Anthropic is racing toward a trillion-dollar IPO, the filing we covered last week.
My take: I want to take the safety argument seriously, because it might be correct. If models really are closing in on the ability to improve themselves, then designing a coordinated brake before you need one is just good sense. But you can't separate the message from the messenger's incentives. A company warning that the frontier is too dangerous to race on, while shipping the most capable public model yet and raising money at a trillion dollars, is asking us to hold two ideas at once, that this is genuinely risky and that you should buy theirs. Maybe both are true. The cleaner way to say it is that "this is dangerous" has quietly become a product category, and Fable 5 is the first one you can expense.
There was a pattern under this week's headlines. Apple, with more cash than most countries, decided it can't build a frontier model and is renting Google's. Anthropic shipped a model so capable it had to bolt the safety on at the edges. The distance between the few labs that build the frontier and everyone else who buys it stopped looking like a gap and started looking like a moat.
We spent the last three years talking about AI as a product, then as infrastructure, then last week as a price. The quieter story is dependence. The intelligence inside more and more of the software you touch now comes from three or four companies. When Apple's assistant thinks with Google's model, and your coding tool and your competitor's both bill against the same lab's inference, the real power in tech is concentrating faster than anything I can remember. And the dependence runs deeper than money. The same few labs set the price, the capabilities, and the limits of the intelligence everyone else builds on, and they're the same names now racing each other to the public markets.
That's the backdrop to both Anthropic's brake-pedal warning and Sanders's tax bill. Strip away the details and they're asking the same question. If a few labs are going to own the most important technology of the decade, who gets a say, and who gets the upside? Apple answered for itself on Monday. It signed a lease. The rest of us haven't decided yet.
Last week I flagged the Ratepayer Protection Act moving through Raleigh. Since then the House passed it, 69 to 44, and sent it back to the Senate. So this is no longer a committee story. It's a real bill headed for a real fight.
The data-center half of the bill is the part most people like. Large facilities would have to pay for the grid upgrades and new generation they require, use closed-loop cooling that recycles water, file environmental and noise studies, and they'd be blocked from using eminent domain or collecting local tax-break incentives. The idea is simple enough, that data centers should pay their own way instead of leaning on everyone else's power bill.
The fight is over the other half. Democrats and clean-energy groups argue the bill's energy provisions, which would let Duke delay retiring coal plants while it lines up new nuclear and would reopen the state's 2050 carbon-neutral goal, will actually push bills up, not down. They point to nuclear projects in Georgia and South Carolina that ran billions over budget. Meanwhile Attorney General Jeff Jackson is fighting Duke's pending 18% rate hike with a counter-proposal he says would save customers $1.4 billion over two years, and he wants data centers put on their own rate class so the rest of us aren't subsidizing them.
Here's the thread back to the top of this issue. The frontier's value is concentrating in a few labs, mostly on the coasts, and this week even Apple lined up to rent from one of them. The frontier's physical cost, the power and water and transmission and higher bills, lands in places like North Carolina. The Triangle's edge was never going to come from out-building those labs. It comes from applying their models inside regulated industries and from getting the unglamorous infrastructure rules right before the buildout outruns them. This bill is that exact fight, playing out a few miles down the road.
That's the week the frontier became a landlord. See you next Wednesday.
Daniel
BullCity AI · Durham, NC
P.S. If you switched your assistant, your coding tool, or your stack to a different model this week, whether that's Fable 5, Gemini, or anything else, hit reply and tell me what changed and whether it was worth it. I read every one, and the honest reports are better than any benchmark.
P.P.S. Forward this to the person on your team who still thinks Apple builds everything itself. As of Monday, Siri thinks with Google's brain.