The AI bubble didn’t pop. It sent the bill

Until last week, about 200 organisations could use a preview of Anthropic’s most capable model, Mythos. They got in through a programme called Glasswing, after the system flagged thousands of software vulnerabilities. Then, on the evening of 12 June, the US Commerce Department ordered Anthropic to cut off every foreign national, anywhere.

The company could not separate them from everyone else in real time. So it pulled Mythos and its sibling Fable 5 offline for every user on Earth. Fable 5 had spent three days topping every major benchmark.

By Friday, a few early testers still kept preview access, among them the security firms Dragos and Cisco. Days earlier, Europe’s cybersecurity agency, ENISA, had joined Glasswing. The order pushed it back out.

It is a small, strange detail. It also holds the whole AI economy in miniature. The technology is valuable enough that customers cling to back-channel access. It is sensitive enough that a government will reach into a private firm’s servers and switch it off.

For three years, the story ran on abundance.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!

Models grew. Demos sharpened. The marketing promised intelligence too cheap to meter. This June, that era ends on every front at once. Companies are now repricing the money, the labour market, the security model, and the politics in the same quarter.

The money got ahead of the returns

Start with the balance sheet. That is where the strain shows first. Man Group is one of the world’s largest hedge funds. This week it warned of mounting bubble risks as bond sales boom to fund AI. Its head of client portfolio management, Sriram Reddy, did not hedge. Enthusiasm for the sector, he said, has encouraged significant overreach.

The firm thinks AI-related issuance could approach a fifth of all bond supply. The dot-com peak was roughly $85bn in 2001. The tell is in who borrows now. NextEra Energy, a utility, has overtaken Volkswagen and BP as the largest hybrid-bond issuer worldwide. It sold three this week. The cash helps fund a $118bn move on Dominion Energy and the data-centre load behind it.

Royal Bank of Canada’s chief executive Dave McKay described an insatiable appetite for capital. None of this looks like an industry confident it can pay as it goes. It revives a question TNW has asked before: what happens when the bubble bursts.

The bulls still have a case

The counter-case is real. The evidence does not all point one way. JPMorgan Asset Management runs some $4.3tn, and it tells clients to stay in stocks. Its chief strategist, David Kelly, expects the AI boom to refuel the rally through the second half. He calls it an OK economy for Americans and a great one for the stock market. The unit’s chief executive, George Gatch, sees a rising tide lifting all boats. Nvidia led the S&P 500 this week.

The Philadelphia semiconductor index hit a record.

Betting against any of it has hurt. Polen Capital, a Boca Raton firm, snubbed Nvidia for Adobe. Its assets then fell from about $14bn to under $2bn. Bloomberg reckoned the wrong-way bet cost it some $50bn. A bubble and a boom can be the same object at different distances. In 2026, they are.

The all-you-can-eat buffet is closing

Cost changes the texture of the debate. For years, developers pushed AI into every product and swallowed the compute bill to win users. That buffet is closing. The industry is moving from flat-rate subscriptions to token pricing, which charges for what people actually use. Even Google has started to meter access through an AI Credits system. Its annual capital spending is climbing towards $175bn to $185bn.

The repricing ripples outward in unglamorous ways. Apollo’s concessions on Shutterfly’s debt showed how heavily AI competition now weighs on ordinary borrowers. Enterprises, meanwhile, are learning the real cost of running AI at scale. The free lunch is becoming a metered utility. Metered utilities invite scrutiny.

Jobs taken, jobs postponed, jobs not yet imagined

Labour is where abundance and anxiety meet most directly. Both are happening at once. McKay marvels at how fast AI trains his staff and lifts their output. That is the optimistic reading. The pessimistic one sits in Madrid. There, the private equity firm Portobello has paused the sale of its 76% stake in the legal-services platform Legálitas.

Prospective buyers fear AI will gut the sector before they can exit.

Europe wants to use the technology to plug a hole rather than dig one. Its factory workers are retiring faster than firms can replace them. So the continent wants AI on the shop floor to cover the gap. Mistral, Siemens, and Schneider Electric lead that industrial push.

Critics still warn that its broader sovereignty drive is partly an illusion.

That is the uncomfortable synthesis. AI displaces some work. It freezes investment elsewhere while everyone waits to see who survives. And it quietly takes on jobs that demographics will soon leave undone. South Korea’s central bank added a twist this week. Record bonuses at Samsung Electronics and SK Hynix drove 1.3 points of a 3.4% rise in nominal wages last quarter.

The bank warned those payouts could spread wage demands well beyond chips and stoke broader inflation. The machine sold to cut costs is, in places, pushing them up.

Security is now a product and a liability

As AI spreads, its attack surface grows.

The trigger for the Mythos shutdown was itself a security scare. Someone, the administration feared, had found a way to jailbreak the model. Companies are now racing to lock down their data. FedHIVE’s chief executive, Michael Cardaci, says the US government wants to keep American computing inside American borders as the race intensifies.

The opportunity is large enough to fund a category. Dream, an Israeli firm, raised $260mn at a $3bn valuation, roughly triple its worth a year ago. Its founders include former NSO Group chief Shalev Hulio and Austria’s former chancellor Sebastian Kurz. It sells what it calls sovereign AI for nations.

Anthropic spent the week setting out its own vision for model safety. Co-founder Jack Clark says the firm talks to the Trump administration daily about its two most powerful systems. Security is now both the product on sale and the risk to manage, often at the same company.

The state has entered the chat

That disabled model brings the threads together. Commerce Secretary Howard Lutnick wrote to Anthropic’s chief executive, Dario Amodei. The letter warned that the company would need a licence before giving Fable 5 or Mythos 5 to any foreign national, anywhere. It also threatened criminal and civil penalties.

Axios first reported the intervention. Investors reacted fast. A proxy contract for Anthropic’s pre-IPO shares slipped about 3.7%. The timing stings, because the firm is reportedly eyeing a listing near $965bn.

Political risk now looks like something that can hit an AI position faster than any valuation or spending worry. The same logic runs through hardware, part of the wider chip war between the EU, US, and China. Commerce told ASML it worries that one of its top extreme-ultraviolet machines reached China against export rules. ASML denies it.

The company says it has never shipped an EUV system to China, and Commerce has shown no public evidence of one there. Sovereignty, the word of the year, now means the state deciding who runs which model on whose soil.

Put the week together and the picture looks coherent, not chaotic.

Borrowing is breaking records. Pricing is tightening. The labour market is shifting in both directions. The security bill is coming due. Governments have left the sidelines. Each force is survivable on its own. The harder question is what happens when they converge.

Picture the metered bills, the layoffs, the breaches, and the export orders all landing in one quarter. Then the rising tide has to lift boats that are also taking on water.