AI in 15 — June 26, 2026

Here's a question that sounds simple and isn't: do you have the right to point a piece of software at a website and have it shop for you? Amazon says no. Perplexity says that's the whole open web. And right now a federal appeals court is deciding which of them is right.

Welcome to AI in 15 for Friday, June twenty-sixth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today is shaping up to be about who controls the AI stack, Marcus — the access, the talent, the supply chain. We start with a courtroom fight that could reshape every shopping and travel agent you'll ever use.

Then — Google keeps bleeding its Gemini brains to Anthropic, and we now have a valuation that explains exactly why.

The West quietly builds a chip alliance designed to box out China.

And a half-billion-dollar plan, backed by OpenAI and Anthropic, to wipe out the common cold.

Lead story, Marcus. Amazon versus Perplexity. On the surface it's a corporate spat. You're telling me it's bigger.

Much bigger, Kate. The fight is over Perplexity's browser agent, Comet. Comet can log into your Amazon account and shop on your behalf — find the thing, add it to cart, check out, all without you clicking. Amazon sued, and here's the legal hook: they're invoking the Computer Fraud and Abuse Act. That's the federal anti-hacking law, written back in the 1980s. Amazon's argument is that an automated agent operating inside a logged-in account is unauthorized access — essentially that Comet is breaking in.

Breaking in? It's logging in with the user's own password, with the user's permission.

That's exactly the tension, Kate. The user authorized it. But Amazon's case leans on a detail that's genuinely awkward for Perplexity: Comet reportedly identifies itself to the website as "Chrome" — a normal human browser — rather than disclosing it's an AI agent. So Amazon says, you're not just acting for the user, you're disguising yourself to get past our defenses. A federal judge agreed enough to issue a preliminary injunction back in March, blocking Comet from logged-in Amazon pages.

So it's settled?

Not even close, Kate. The Ninth Circuit — the appeals court — paused that injunction, heard oral arguments on June eleventh, and we're still waiting on a ruling. And the reason this is a landmark and not a footnote is the precedent. The CFAA is a sledgehammer. If the courts decide that pointing an agent at a site is "unauthorized access," that doesn't just hit Perplexity. The ACLU filed a brief warning it could chill journalists and researchers who rely on automated tools to scrape public data.

So the principle underneath is — who's in charge of your own browsing? You, or the website?

You've put your finger on it, Kate. The open web's founding assumption is that a user controls their own software. A site serves you a page; what you do with it is your business. Agents push on that assumption hard, because now the "user" is a tireless piece of automation that can act a thousand times a minute. Amazon's real fear, I'd argue, isn't security — it's that an agent shops on price and ignores the ads, the recommendations, the upsells that are the actual business. Whichever way this lands, it writes the rules for every shopping, travel, and research agent that comes after.

Quick hits. Marcus, we've talked about Google losing people to Anthropic — John Jumper, Noam Shazeer. There's fresh blood leaving this week, and a number attached that explains everything.

Two more big names, Kate. Jonas Adler, who worked on Google's AI coding efforts, and Alexander Pritzel — a model pretraining researcher who, notably, contributed to AlphaFold. Both heading to Anthropic. That's on top of Jumper, the Nobel laureate, also going to Anthropic, and Shazeer who jumped to OpenAI. This is not a trickle anymore.

And the number?

Anthropic just got valued at nine hundred sixty-five billion dollars, Kate. They confidentially filed for an IPO on June first, and they could go public as soon as this fall. That's the whole story right there. The weapon Anthropic has that Google doesn't is pre-IPO equity — stock that hasn't gone public yet, so the upside is still ahead of you. You join now, you list this fall, you potentially get a life-changing payday. Google's a mature company; its stock already did its big run decades ago.

So it's not that Google's a worse place to work. It's purely the money?

Money's the biggest lever, Kate, though it's not the only one. A SignalFire analysis last year found DeepMind staff were nearly eleven times more likely to leave for Anthropic than the other direction. Eleven to one. And I'd flag the usual caution — exact pay packages are reported, not confirmed, so treat specific dollar figures with some skepticism. But the direction is unmistakable. The model race has become a talent race, and Google keeps losing the exact people who built Gemini's coding and pretraining stacks. You can't easily out-engineer the people who left.

Next, Marcus. There's a chip alliance forming with a name straight out of a history textbook — "Pax Silica." And it's growing fast.

It is, Kate. Pax Silica is a US State Department initiative — the idea is to pull the entire AI supply chain, chips, critical minerals, energy, away from China and into a bloc of allied countries. This week it expanded meaningfully: the EU, Germany, the Netherlands and Greece have signed on, plus Argentina, Chile, Costa Rica, Kazakhstan and Panama joining now, pushing it past two dozen members.

The Netherlands makes sense — that's ASML, the company that makes the machines that make the chips.

Exactly the keystone, Kate. You can't build a leading-edge chip without ASML's lithography machines. So getting the Netherlands inside the tent matters enormously. The architect, a State Department official named Jacob Helberg, pitched it as a counter to China's Belt and Road, and warned that fragmented rules risk what he called "synchronised mediocrity." Now — per the show's instincts, I'll keep the politics light. But I'll also do the thing I always do: China's state outlet, the Global Times, framed the EU joining as coming "at the cost of AI independence." That's a state-aligned source pushing a line, so flag it as such and move on.

But this is the scaffolding under every chip story we cover.

That's the one-line takeaway, Kate. Every custom chip, every export fight, every talent war we've covered this week sits on top of this question — can the West build an end-to-end chip pipeline that doesn't run through Beijing? Pax Silica is the attempt. Whether it holds is a multi-year story.

Quick connective one, Marcus. We covered the Anthropic-Alibaba distillation accusation yesterday — that alleged campaign to clone Claude. There's an open-model angle worth a beat.

There is, Kate, and it explains why distillation is such a raw nerve right now. The open-weights model GLM-5.2, from Z.ai, has become the talk of the developer community — we mentioned it earlier in the week beating GPT on coding. The thing to sit with is the price. It's roughly a dollar-forty per million input tokens, against the closed labs charging many times that, and it's landing in the same league on long-horizon coding tasks. Agentic software engineering was supposed to be the closed labs' moat — the thing you couldn't get cheaply anywhere else.

And now you can.

And now you can, openly licensed, Kate. So when Anthropic accuses a Chinese lab of distilling Claude's coding skills, you can see why it stings. The most valuable, most defensible capability is exactly the one that keeps showing up in cheap open models. I'll keep my standing caveat — the distillation figures are Anthropic's own and not independently verified — but the competitive pressure is very real, and it's pointed straight at the part of the business the labs thought was safe.

Let's end the hits on something completely different, Marcus. A palate cleanser. OpenAI and Anthropic are teaming up — not to compete, but to fight the common cold.

It's a lovely one, Kate. A new nonprofit called Intercept, half a billion dollars, backed by Stripe, Anthropic, the OpenAI Foundation, Bill Gates, and traders from Jane Street. The goal is genuinely audacious — drastically cut respiratory infections, the colds and flus that quietly cost the economy enormous amounts every year. The plan pairs next-generation vaccines with large-scale air cleaning in buildings.

Where's the AI in it?

Through the Arc Institute, Kate — a research outfit building AI-for-biology models that speed up vaccine and protein design. Same family of techniques as AlphaFold, which keeps coming up today for a reason. Instead of years of wet-lab trial and error, you let models propose and narrow down candidates first. I'd temper expectations — "cure the common cold" is a headline that's broken a lot of hearts over the decades. But it's a real example of frontier-lab money flowing into science rather than just bigger chatbots, and that's worth noting on its own.

Rival labs funding the same nonprofit. That's a nice change of pace.

It is, Kate. They'll go back to poaching each other's researchers tomorrow. But for one story, they're on the same side.

One to watch tomorrow, Marcus.

Fable 5, Kate. Anthropic's Fable and Mythos models were forced offline on June twelfth under a US export-control directive. This week, listings and ID strings for Fable 5 reappeared in AWS Bedrock and in Claude Code documentation — which set off a lot of excitement. The catch: Anthropic staff say zero traffic is actually being routed, and it's probably just a UI bug, not a quiet relaunch. Watch whether real access genuinely comes back.

Agree, or counter?

Agree it's the one to watch, with my usual counter, Kate — a string in a config file is not a product. Don't read a soft relaunch into a stray line of documentation. Watch for actual traffic and an actual statement, not a leaked identifier.

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