AI in 15 — June 15, 2026

KPMG sold a flagship report on the future of AI. It charged clients six and seven figures for advice like it. There was just one problem. The report itself was riddled with AI hallucinations — and of its forty-five citations, only five pointed to anything real.

Welcome to AI in 15 for Monday, June fifteenth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

We've lived inside the Fable shutdown all week, Marcus, and today it went international — Europe is scrambling and Dario Amodei sits down with G7 leaders tomorrow. That's our lead. Then the consulting giant whose AI report was written by hallucinating AI. A city that claimed it built its own model — and got caught. Amazon's chip business quietly blowing past twenty billion dollars. And the eye-watering numbers behind the whole buildout.

Why KPMG just deleted its own AI report.

How Rio de Janeiro's homegrown AI turned out to be two other people's models in a trench coat.

And the seven-point-six trillion dollars the world is about to spend.

Lead story, Marcus. The Fable and Mythos shutdown went global over the weekend. What's actually new today?

Two things, Kate. First, Europe woke up angry. On Sunday the European Commission said it's, quote, looking closely at the practical consequences for European users. Their tech-sovereignty spokesman, Thomas Regnier, added a pointed line — that contingency measures, quote, should not be discriminatory against partners. Translation: a single American government letter took two frontier models offline for paying European customers overnight, and Brussels is not amused. Second, Dario Amodei meets G7 leaders and rival AI executives tomorrow, Tuesday. Days after Washington switched off his best model.

And we finally have a name attached to the jailbreak.

Reportedly, Kate — and I'd stress reportedly. Coverage links it to a well-known hacker who goes by Pliny the Liberator, allegedly using Unicode substitution tricks and publishing a leaked system prompt running to roughly a hundred and twenty thousand characters. But analysts are waving a caution flag: that looks like prompt engineering, not a genuine break of the model's safety weights. Which lines up with Anthropic's own account — that the demonstrated technique amounted to asking the model to read a codebase and fix its flaws. A capability every competitor already ships.

So you've got the EU treating an American export rule as a threat to its own users.

And that's the watershed, Kate. We've spent the week on the domestic angle — who pushed, the precedent, the politics. The new layer is that the rest of the world now sees, in real time, that a model they depend on can vanish on the say-so of a government they didn't vote for. That is rocket fuel for European tech-sovereignty hawks, and frankly it hands every argument to local hosting and multi-vendor routing. The uncomfortable irony for the West? We may be teaching our closest allies to build around us.

Quick hits. Marcus, KPMG. This one is almost too perfect.

It really is, Kate. KPMG quietly pulled a report titled "Redefining excellence in the age of agentic AI." The research group GPTZero went through it and found that of forty-five citations, only five were real. Nearly half the factual claims were unsupported, inaccurate, or misattributed — the classic fingerprints of AI hallucination. UBS, the UK's NHS, Swiss Federal Railways, and Transport for London all said the report mischaracterized their work. My favorite: it claimed Emirates had a mobile chatbot called Sara that could change your bookings. In reality Sara was a 2023 robot greeter that could do nothing of the sort.

So the firm selling AI-transformation advice shipped a report the AI made up.

That's the whole story, Kate. And it's not the first — Deloitte had a near-identical embarrassment last year. The lesson isn't "AI bad." It's that these firms charge clients enormous fees to advise on human oversight, and then skipped the human oversight on their own flagship document. It validates the skeptic's case precisely: AI got bolted onto a workflow and made the output worse, not better. Nobody checked the work.

Which is becoming a theme.

Hold that thought, Kate, because the next story is the same disease in a different outfit.

Rio de Janeiro's homegrown city model. Marcus, what happened?

So Rio's municipal IT company released something called Rio-3.5-Open — a three-hundred-ninety-seven-billion-parameter model — and touted it as homegrown, beating comparable open models on benchmarks. Independent analysts pulled it apart on GitHub, and it's nothing of the kind. Every weight tensor — to thousands of standard deviations — is the same blend: roughly sixty percent of one existing model called Nex-N2, forty percent of another called Qwen, averaged together across all sixty layers. No evidence of any original training at all.

How did they catch it?

This is the fun part, Kate. Strip away Rio's "you are Rio" instruction, and the model introduces itself as "Nex, from Nex-AGI" seventy-nine percent of the time — reciting the other company's backstory word for word. The open-source community cried fraud. And to be clear, merging models is a perfectly legitimate technique. The sin was claiming it was original and hiding the blend.

And there's a geopolitical wrinkle you're going to enjoy.

You know me too well, Kate. The forty-percent base under this proudly "sovereign" Western city model is Qwen — a Chinese open-weight release. So a government chasing the sovereign-AI headline quietly built on exactly the ecosystem we keep warning is seeding projects worldwide. It's benchmark theater and attribution-washing in one. And it pairs with KPMG perfectly: two institutions presenting AI output as trustworthy and original, when nobody bothered to verify the foundation.

Let's talk hardware. Marcus, Amazon's chip business just crossed a number that should make Nvidia nervous.

Twenty billion dollars, Kate — that's the annual revenue run rate Amazon's in-house silicon exited the first quarter at. That's Graviton CPUs, the Trainium AI accelerators, and Nitro, together growing nearly forty percent quarter on quarter, triple digits year over year. Andy Jassy says as a standalone company it'd be worth around fifty billion — a top-three data-center chip player on its own — and he hinted Amazon might start selling Trainium to outsiders.

And the customers are names we cover constantly.

That's the kicker, Kate. Amazon is sitting on over two hundred twenty-five billion dollars in revenue commitments for Trainium alone — with Anthropic signed up for as much as five gigawatts of capacity and OpenAI for around two. Trainium2 claims roughly a thirty percent price-performance edge over comparable GPUs and is basically sold out. Trainium3 started shipping early this year with another thirty to forty percent gain, nearly fully subscribed.

So Nvidia's near-monopoly is finally cracking.

At the hyperscaler level, yes, Kate. And here's why it ties to our lead. The frontier labs aren't just buying cheaper compute — they're buying independence. The same week a government proved it can switch off a model overnight, the people who build those models are deliberately diversifying the hardware underneath them. Whoever controls the compute increasingly controls the models. Anthropic and OpenAI both clearly decided they don't want a single landlord.

Marcus, put a number on the whole buildout. How big is this bet?

Brace yourself, Kate. Goldman Sachs projects cumulative AI capital spending from this year through 2031 will hit seven-point-six trillion dollars — compute, data centers, and power. That's roughly a quarter of one year's entire US economy. It's about one-point-four times the entire annual output of Germany. On AI infrastructure. In six years.

That's a number that fuels both the bull case and the bubble fear.

Both at once, Kate, and that tension is the whole story. Against that backdrop, Anthropic's president Daniela Amodei told Bloomberg that frontier AI needs enormous capital and the public markets are, quote, very well suited to funding it — and disclosed that annualized revenue reached forty-seven billion dollars in May, up from nine billion at the end of last year. I'll flag that figure as reported, not confirmed — but if it's even close, it's staggering. And Nvidia's Jensen Huang compared buying the coming AI IPOs to getting early shares in Amazon or Google.

Which means ordinary investors are about to be invited into this trade.

Exactly, Kate. SpaceX is already public, OpenAI and Anthropic are circling. A multi-trillion-dollar buildout this fast strains the power grid and the capital markets simultaneously — and now retail gets to ride along. My caution is the one this show keeps repeating: a revenue curve that vertical, printed into a market that has already decided AI is the trade, deserves a skeptic with a spreadsheet. Show me the cash, not the keynote.

One more, Marcus, and it sharpens the whole episode. Anthropic says most of its own code is now written by AI.

In a June fourth paper, Kate, Anthropic disclosed that over eighty percent of the code merged into its production codebase was authored by Claude as of last month. Engineers shipping roughly eight times more code per day than in 2024, and the length of tasks AI can complete is doubling about every four months. And the paper's actual purpose was to argue for globally coordinated guardrails on frontier development.

So the company building the most capable models is also lobbying for the power to restrict them.

And that's the tension critics jumped on, Kate — one commentary was literally titled "Did Anthropic ask for this?" The argument is that you can't simultaneously sprint to the frontier and demand the authority to police it — and this weekend's government shutdown of Fable and Mythos is what that contradiction looks like when it bites. Whether you read it as responsible or as inviting exactly the heavy-handed state intervention that just hit them, it's the same recursive-self-improvement story made tangible. The safety-focused lab's software is now mostly machine-written.

Big picture, Marcus. Pull it together.

Two threads, Kate. The first is sovereignty — it's the meta-story of the year now. A government letter pulling Anthropic's models. A city faking a model to look sovereign. Europe scrambling for independence. Amazon's chips letting labs escape their hardware landlord. Every one of those orbits a single question: who controls the models, and who controls the compute beneath them. And uncomfortably for the pro-Western case, the honest answer this week is that control is more fragile, and more contested, than anyone admitted a month ago.

And the second thread?

Verification, Kate. KPMG shipped a report its own AI hallucinated. Rio passed off a merge as original training. Two faces of one problem — AI output handed to the public as trustworthy when nobody checked the work. And it rhymes with the lead, because the entire Fable shutdown turns on a contested claim about what a jailbreak actually demonstrated. Reported as fact, unverified underneath.

So where do you land, given the week we've had?

Genuinely torn, Kate, and I'll own it. The free-market instinct says a government that can switch off a private company's best product overnight, on a contested security claim, with allies left stranded — that's a power easy to grant and brutal to claw back. But the verification stories cut the other way: an industry this sloppy about checking its own work hasn't exactly earned blind trust either. The thread that holds it together is the same every day this week. The capability is real and largely Western. What's missing is the paperwork — the proof, the process, the human who actually checked. The West can still win this. But you win it by demanding the receipts, from the government and from the labs alike.

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