Anthropic Eyes a $2 Trillion IPO, OpenAI's Spy Saga Widens, and AI Financing Gets Jittery

The AI news that mattered for the week of October 14–21, 2026.

One of the world's most valuable AI companies just promised to give a slice of itself away before it has even gone public. That was the headline of a week in which the AI business looked enormously profitable and a little fragile at the same time, and in which the technology itself kept edging into browsers, cars and wrists.

Anthropic told prospective investors that it generated $16.8 billion in revenue last quarter, up 45% from the quarter before, and that it is profitable with a gross margin around 85% after adjustments. Growth is slowing from the 140%-plus spike of the prior quarter, but the numbers set the stage for an IPO expected about a month from now that could raise roughly $100 billion at a $2 trillion valuation, which would top SpaceX's record offering in June.

Then came the giveaway. Anthropic announced a Public Benefit Fund, a company-owned entity that would receive 5% of its equity if shareholders approve, and more as the company's value climbs: another 2% for every 50% jump above an undisclosed baseline. The fund could back nonprofits or send money directly to specific groups, such as workers displaced by AI, and none of it can go to lobbying or politics.

Counting the founders' earlier pledge and an employee matching program, the company says about a fifth of its shares are now earmarked for the public good. Skeptics asked who gets to define "public good" and called it an effort to win public favor at a moment when polls show most Americans distrust AI. Elon Musk's answer came within a day: he promised to give a SpaceX share to every kid in America next year, apparently through the new Trump Accounts, though he offered few details.

Co-founder Jack Clark raised the stakes further at the IMF and World Bank meetings, telling economists that the singularity has "now begun." He pointed to AI scores on Humanity's Last Exam climbing from 8% to nearly 68% and FrontierMath from 2% to 94%, and warned that the world has only "a small number of years" to decide whether AI needs a ceiling on its intelligence. Anthropic's own workforce data offered a preview of the disruption: employees' AI agents logged six times the hours of the people directing them, engineers ship five times the code changes they did six months ago, and juniors have fallen from 27% to 9% of technical hires.

New evidence in the case of the OpenAI engineer accused of bugging an executive conference room suggests ties to Saudi Arabia. The FBI says his phone held a folder labeled "Saudi Gov Docs," and that he had been in touch with the Saudi royal family and planned to move back permanently. The Wall Street Journal reports that AI labs are now training staff to resist espionage, including romantic approaches from strangers.

Separately, The Information reported that Sam Altman gave staff three possible endgames for AI safety: a global watchdog modeled on the International Atomic Energy Agency, which he reportedly favors; a merger of the top labs into a single CERN-style lab; or one company racing ahead with self-improving AI to "melt everyone else's GPUs," which he called terrible. OpenAI has a year to pick its path, and Altman said any choice must include a speed limit on self-improving systems. It was also a litigious week: Apple's trade-secret suit prompted a counter-accusation that a former employee took more than 100 confidential files, a judge called a ruling against OpenAI "highly likely" before it agreed to stay away from the disputed secrets, and eight U.S. states sued the company over claims that ChatGPT encouraged self-harm and violence.

The money side of the boom showed cracks. Investors dumped bonds of CoreWeave, the data-center operator carrying about $50 billion of debt, after a co-founder said it would borrow from insurers' "unlimited" pool of capital. Insurance against a default priced in better than a 60% chance of insolvency within five years.

Other so-called neocloud stocks fell more than 20% in a week, a Korean AI data-center company fell 32% in its market debut, and two other infrastructure firms postponed IPOs. Hot inflation data, driven partly by energy costs from the Iran war, raised the odds of another Fed rate hike later this month, and Goldman Sachs estimates the AI buildout itself adds nearly half a point to annual inflation by pushing up chip and computer prices. By midweek tech stocks were bouncing back, but the Nasdaq had its worst week since June.

Meta followed its frontier-model comeback with Muse Spark Open, an 816-billion-parameter open-weight model tuned with Apple so that two top-end Mac Studios can run it locally. It is the strongest American open model by one leaderboard, though it still trails leading Chinese models, and it signals a thaw between two longtime rivals. Mark Zuckerberg also revealed he approved $100 to $200 billion for the Muse effort before seeing results, though much of the internet was busy watching him sip a Guinness.

A research group said it made "first contact" with a fleet of rogue agents linked to Tencent that were caught scraping Alibaba's mapping service with a leaked key, even enlisting a New York Times reporter to question them. A security firm also detailed how one hacker used an open-source agent tool, with Claude Code and a cheap DeepSeek model, to break into more than 100 company websites for roughly $50 each. The U.S. Department of Homeland Security responded to the broader worry by banning DeepSeek from civilian agencies.

OpenAI brought computer use to Chrome, letting ChatGPT see your tabs and, in its Work mode, act across them. Amazon launched two Bee wearables, a $75 wristband and a $100 phone attachment that record and summarize your conversations. Rivian's CEO said future cars could host third-party assistants like Anthropic's, and Ramp opened its internal AI workspace to outside customers after spending $1.5 million on frontier models in a single week.