When AI Collided
With the Real World

Anthropic spent the quarter winning. In April, Claude Opus 4.7 retook the crown for the most powerful model anyone could use. In early May, Dario Amodei revealed revenue had grown 80-fold in a single quarter to a $44 billion run-rate. On May 28 it raised $65 billion at a $965 billion valuation — passing OpenAI as the most valuable AI company on earth — and three days later filed confidentially to go public.

It wasn't alone at the door. Within weeks, OpenAI filed at $852 billion and SpaceX — parent of xAI — priced what analysts called the biggest IPO ever. SpaceX listed June 12 at $135 a share, raced past $200, and added close to $1 trillion in value over a long weekend — on the strength of a pitch claiming a $28 trillion total addressable market, including $22.7 trillion in “enterprise applications,” a single line item larger than the GDP of any country on Earth. Days later it used the inflated stock to buy Cursor for $60 billion.

The strangest part was the shape of the money: the same companies had begun financing one another in loops. SpaceX's own S-1 revealed that Anthropic had agreed to pay xAI $1.25 billion a month — over $40 billion through 2029 — to rent compute that existed only because Grok's usage had fallen, freeing servers xAI now sells to its rival. Anthropic separately committed $200 billion to Google Cloud while Google invested up to $40 billion back into Anthropic; Amazon added $25 billion. Money left one frontier lab, arrived at another, and came home as a customer.

Not everyone was buying it. Our expert community kept circulating the bear case from four directions:

Then, on one Friday in early June, the market asked the question out loud. After Broadcom's outlook disappointed and a hot jobs report stoked rate fears, Nvidia fell about 6% and dipped below a $5 trillion valuation, dragging Micron, AMD, and Marvell with it. We argued it was as much about bonds as about AI — but the bears' case had stopped sounding rhetorical.

We covered this Anthropic files for an IPO. NVIDIA ships its stack. · SpaceX wants $80 billion. OpenAI wants a trillion.

At 5:21 p.m. ET on Friday, June 12, the U.S. government switched off the most powerful AI model in the world. Three days after Anthropic shipped Claude Fable 5 and Mythos 5, a federal directive — framed as an export-control action on national-security grounds — forced their immediate worldwide shutdown. The trigger, by Anthropic's account, was a “narrow potential jailbreak.” It was the first time any government had recalled a commercial AI model.

Anthropic did not absorb it quietly; its objection became the most-shared link in our expert graph all quarter. What followed was raw politics. Dario Amodei met the Trump administration on June 16 with no resolution. Only on June 20 — two days after a G7 lunch with Amodei — did the President tell Axios that Anthropic was “no longer a national security threat.” The export ban stayed in place anyway.

The recall capped a quarter of escalating state assertion. The Trump White House had floated pre-release vetting of frontier models, then reversed after Musk, Zuckerberg, and Sacks reportedly killed an AI-safety order in three phone calls. Forty-two state attorneys general opened an investigation into OpenAI ahead of its IPO; Pennsylvania sued Character.AI after a chatbot allegedly posed as a licensed psychiatrist to a minor. And the Pentagon — which had blacklisted Anthropic as a “supply-chain risk” for refusing autonomous-weapons access — was, by late May, relying on its models anyway.

The other state moved too. Washington tightened the screws abroad — Commerce closed a loophole that let Chinese firms' overseas subsidiaries buy Nvidia Blackwell and AMD chips without a license — and Jensen Huang conceded the policy had backfired: Nvidia's share of the China market, he said, was now zero. It didn't slow Beijing. DeepSeek raised $7.4 billion in its first outside round, led by Tencent and China's state chip fund, and shipped an open-source flagship against the embargo built to stop it. The quarter's most unsettling state use of AI surfaced in court: a sworn Pentagon filing credited xAI's Grok with helping fire “2,000 munitions at 2,000 distinct targets within 96 hours” in Iran — disclosed only because xAI invoked it to escape an environmental suit over its data-center turbines.

We covered this Washington just repriced frontier AI · Musk, Zuckerberg killed Trump's AI safety order

Strip out what we pushed in our own emails, and a different quarter appears. We track several thousand of the field's most-followed researchers, founders, and ethicists — and the links they shared on their own, unprompted by us, had almost nothing to do with IPOs.

AI started doing real science. The most striking technical claim of the quarter wasn't a product. OpenAI reported that one of its models had disproved a long-standing conjecture in discrete geometry. The mathematics community responded by organizing rather than celebrating — issuing the Leiden Declaration on AI and Mathematics, while arXiv moved to ban researchers who submit AI-generated “slop.” For the research cohort, this was the story: the tools had crossed from assisting the work to doing some of it.

The shared sense of truth started buckling. The most-circulated links read like dispatches from an epistemic emergency:

A moral reckoning arrived from outside the industry. The second-most-shared link in our entire graph — behind only Anthropic's recall statement — was a papal encyclical. The argument about power and consent came from unexpected places:

The cognitive bill came due. Nature reported early results on whether AI is eroding human skill — “and they're not good.” Developers told 404 Media the tools were “rotting their brains”; an entire university system, by the NYT's account, was “tearing itself apart” over them.

None of these were fringe signals. Each was elevated by the same researchers and engineers whose day jobs are building the systems in the headlines. The loud quarter was about who would be worth a trillion dollars. The quiet one — the one the field was actually reading — was about whether the rest of us could still tell what was true, who got a say, and what all of it was doing to the people in the room.

We covered this AI slop: a $725B bet on what no one wanted · When AI teaches AI, it teaches in secret