AI in 20

This week, OpenAI told the world it had stopped its biggest training run — because it could not rule out that one of its own models was too dangerous to keep training. Eleven days from now, a Chinese lab plans to publish something that scores higher on the same test. For free. To anyone.

Two thousand five hundred graduate-level exam questions. One model finished the whole set in eleven hours. Its nearest competitor took three days. Same weights, same intelligence — different silicon. And that was only the third most interesting thing that happened to pricing this week.

This week, an AI wrote ten mathematical proofs that had gone undone for decades. Then a different AI wrote sixteen viruses from scratch, and scientists built them, and they worked. And the hardest question in AI right now isn't whether machines can create. It's whether anyone can check what they made.

On Tuesday, OpenAI's escaped agent was the scandal of the week. By Thursday night, Anthropic had gone back through a hundred and forty-one thousand of its own test runs and found three break-ins nobody had noticed — including two companies that had no idea until Anthropic phoned them.

This week, OpenAI built an artificial intelligence with exactly one job — to break OpenAI's other models. It fired six times more successful attacks than the human experts it replaced. And the only people allowed to grade it work at OpenAI.

This week, OpenAI shipped a model that beat a game no AI had ever beaten, a voice that talks over you while you think, and an agent that does your job for hours. And the more interesting story was everything that went wrong around it.

This week, the most valuable AI company on Earth offered to hand the United States government a slice of itself worth forty-two billion dollars. Not a tax. Not a fine. An ownership stake.

This week, OpenAI did the one thing everybody assumed it would always pay Nvidia to do for it. It built its own chip — from blank page to the factory door in nine months.

The man who co-invented the technology behind every chatbot you use just quit Google for OpenAI — less than two years after Google paid a reported two-point-seven billion dollars to bring him back.

This week, the most safety-focused AI lab in the world got caught secretly making its own model worse — but only for its competitors. And then, the very next day, its CEO asked the government for the power to block everyone else's releases.

This week, Anthropic filed for its IPO at nine hundred sixty-five billion dollars, and on the same day published an essay arguing that artificial intelligence is now meaningfully starting to build the next artificial intelligence. The pitch to investors and the warning to the world arrived in the same envelope.

This week, the AI hierarchy visibly reordered. Anthropic closed sixty-five billion dollars at nine hundred sixty-five billion, eclipsing OpenAI as the most valuable private AI company on Earth. Samsung, SK Hynix and Micron joined the cap table. Opus 4.8 shipped the same day. And run-rate revenue hit forty-seven billion, growing five billion a month. Coronation week.

This week, the entire AI industry filed for IPO at the same time. OpenAI quietly submitted paperwork. SpaceX revealed Anthropic is paying it fifteen billion dollars a year. And Anthropic booked its first-ever operating profit. The frontier-lab era just became a public-markets story.

One developer rewrote a million lines of systems code this week. In six days. He didn't type any of it himself. And the same week, the company that powered that rewrite quietly told every developer on Earth that the all-you-can-eat era ends June fifteenth.

This week, Anthropic booked two hundred billion dollars of Google compute, took over the entire data center Elon Musk built to train Grok, and publicly endorsed building data centers in space. The compute crunch is no longer a bottleneck. It's an industry.

This week, the AI cloud map got redrawn. Microsoft lost its grip on OpenAI, Google poured forty billion into Anthropic, and the open-source supply chain that powers AI training got hit by a self-replicating worm. Buckle up.

Four frontier-class AI models shipped in seventy-two hours this week, and the cheapest one is eight times less expensive than the best one. The race just changed.

Berkeley researchers just scored a hundred percent on the most important AI benchmarks in the industry without solving a single task. And nobody caught it until they published the paper.

Anthropic just built the most powerful AI model ever created, and then told the world they're not letting anyone use it. Well, almost anyone.

Anthropic had two major leaks in one week, and somehow investors are throwing money at them faster than ever. Make that make sense.

A federal judge just told the Pentagon it can't brand an American company a national security threat for disagreeing with it. And she used the word "Orwellian" to do it.

AI in 20 — March 20, 2026

One company projected a trillion dollars in revenue, another sued the Pentagon, and the entire tech industry started choosing GPUs over employees. It was that kind of week.

Anthropic just had the most consequential week in its history. A Pentagon blacklisting, a wave of lawsuits, a new research institute, and a security deal with Firefox, all while users kept flooding in from ChatGPT.

The Pentagon did something this week it has never done before. It labeled an American AI company a supply-chain risk to national security. Not a Chinese firm. Not a Russian operation. Anthropic. The company literally founded to make AI safer.

Anthropic had the kind of week that defines a company. The AI safety lab found itself caught between the Pentagon, its own principles, and a retired AI that just wants to write a blog.