They Deleted the AGI Clause. Nobody Changed Their Plans.

Microsoft and OpenAI had a deal. Inside it was a clause: if OpenAI achieved artificial general intelligence, Microsoft's license would terminate. The moment was named. The consequences were defined. Then, on April twenty-seventh, two thousand twenty-six, they deleted it. The reason was not that AGI had been reached. The reason was that the clause had become inconvenient. Fourteen months of negotiations had been triggered by an Amazon deal that technically violated Microsoft's exclusivity rights, and somewhere in the unwinding, the two companies decided the simplest solution was to stop naming the thing they were racing toward. The clause was replaced with a date: twenty thirty-two. Satya Nadella, Microsoft's CEO, said two days later on an earnings call that he planned to "exploit" the arrangement. His AI business was generating thirty-seven billion dollars a year, up one hundred twenty-three percent from the year before. He said "exploit" on a call with financial analysts. Nobody found this unusual.

Jack Clark co-founded Anthropic. He also founded the Stanford AI Index and served on the United States National AI Advisory Committee. He writes a newsletter called Import AI. This week he published issue four hundred fifty-five and said, directly, that there is a sixty percent probability that an artificial intelligence system will train its own successor autonomously before the end of two thousand twenty-eight. He listed the evidence: in twenty twenty-two, GPT-3.5 could complete tasks that took a human about thirty seconds. In twenty twenty-three, GPT-4 reached four minutes. In twenty twenty-four, o1 hit forty minutes. In twenty twenty-five, six hours. In early twenty twenty-six, twelve hours of autonomous work. Claude Mythos Preview achieved in April a fifty-two times speedup on a model training task — compared to the four times improvement a human achieves in four to eight hours. Clark described his conclusion as "reluctant" because the implications are large and society is not prepared. He gave it a thirty percent probability for twenty twenty-seven. He did not suggest anyone should stop.

In the Federal District Court for the Northern District of California, in Oakland, before Judge Yvonne Gonzalez Rogers, Elon Musk spent the first week of May on the stand. He had co-founded OpenAI in twenty fifteen with thirty-eight million dollars of his own money to create a nonprofit that would serve as a counterweight to Google. Then he left the board in twenty eighteen. Then he sued. Under cross-examination, he was asked whether xAI, his own artificial intelligence company, had used distillation techniques on OpenAI's models to train Grok. He answered: "In part." He also said that distillation is "standard practice." He told the jury: "I was literally a fool." He meant the donation. A man named Jared Birchall — who manages Musk's personal finances — testified about a ninety-seven point four billion dollar offer that a Musk-led consortium submitted in February twenty twenty-five to acquire OpenAI's assets. The judge asked Birchall how he had persuaded investors to commit to ninety-seven point four billion. Birchall said he did not remember where the number came from. The judge said she found this unpersuasive. The trial continues.

The four largest technology companies in the world spent one hundred thirty-three point six billion dollars on data centers and computing infrastructure in the first three months of twenty twenty-six. Amazon spent forty-three billion. Google spent thirty-six billion, and its CEO said the company would have made more money that quarter if it had more computers to sell — they were "constrained by compute capacity." Morgan Stanley projected that five hyper-companies will spend eight hundred five billion in twenty twenty-six and one trillion one hundred billion in twenty twenty-seven. The twenty twenty-seven figure is approximately what every non-tech company in the S&P 500 spent on capital investment last year. David Sacks — who was the White House AI and Crypto Czar until March twenty-sixth and is now Co-Chair of the President's Council of Advisors on Science and Technology — wrote on X that AI had accounted for seventy-five percent of United States GDP growth in the first quarter. He said the capex momentum would be "closer to two point five percent this year and more than three percent next year." This is how a civilization decides it cannot stop.

Sarah Friar has been OpenAI's Chief Financial Officer since June twenty twenty-four. She has a degree from Oxford and an MBA from Stanford. She previously took Square public in twenty fifteen and guided Nextdoor's SPAC listing in twenty twenty-one. The Wall Street Journal reported on May second that she has suggested privately that the company wait until twenty twenty-seven to go public, putting her at odds with Sam Altman, who wants the IPO in twenty twenty-six. Her concern, per the report: OpenAI may not be ready for public-company reporting standards, and the data center commitments are large. She has reportedly told private investors that the real planned spending through twenty thirty is around six hundred billion dollars — compared to the one point four trillion Altman has used publicly. OpenAI is losing approximately seventeen billion dollars a year. Its revenue is approximately twenty-four billion annualized. The IPO would target a one trillion dollar valuation. When the Journal published this, Oracle's stock fell more than six percent. OpenAI called the story "ridiculous." And so it goes.