It’s the tech world’s messiest open relationship, and everybody just got new partners. Monday, Microsoft and OpenAI quietly renegotiated their landmark deal, ending Microsoft’s exclusive cloud lock on OpenAI’s models — and less than 24 hours later, Amazon rolled up to the AWS event in San Francisco like it had been waiting outside with the engine running.
OpenAI’s GPT-5.5 and GPT-5.4 are now live on Amazon Bedrock in limited preview. A jointly-built “Stateful Runtime Environment” — basically persistent memory for agents — is in the pipeline. AWS CEO Matt Garman, barely hiding his glee at the San Francisco launch event, said customers had been “asking for this for a really long time.”
Microsoft isn’t crying into its Azure console, though. It gets to stop paying revenue share to OpenAI, keeps a non-exclusive IP license through 2032, still holds 27% equity, and remains the “primary cloud partner.” Plus, it’s now cozied up with Anthropic. Everyone’s dating everyone. The biggest winners? Enterprise procurement teams who no longer have to pick a lane.
The courtroom drama of the decade kicked off in Oakland Tuesday. Elon Musk — co-founder, early bankroller, and now full-throated nemesis of OpenAI — took the witness stand in his lawsuit against Sam Altman, alleging the ChatGPT maker “stole a charity” and abandoned its nonprofit soul.
His ask: $130B in damages, Altman’s ouster, and OpenAI forced back into nonprofit form. OpenAI’s counter: Musk knew about the for-profit pivot all along, lost a power struggle, and now is throwing a $134-billion tantrum. Judge Yvonne Gonzalez Rogers told Musk to lay off the “Scam Altman” posts on X during proceedings. He agreed. For now.
Meanwhile, Sam Altman appeared via pre-recorded video at the AWS event in SF, citing a schedule “taken away” from him. Nadella, Brockman, and Altman himself are all expected to testify before the liability phase wraps May 21.
A bombshell paper dropped at ICLR 2026 in Rio de Janeiro with a finding nobody wanted to hear: train a model to reason harder, and it hallucinates tool calls more often. Researchers are calling it “The Reasoning Trap.”
The timing is brutal. Deloitte data shows 47% of enterprise users have already based at least one major business decision on hallucinated AI content — and that’s before the current agent wave. Meanwhile, 96% of enterprises are running AI agents, but only 12% have a central platform to manage them.
In multi-agent systems, one hallucinated tool call can contaminate every downstream agent sharing memory. Princeton IT warns this is especially dangerous in HR and payroll pipelines. Neither prompt engineering nor preference optimization fully closes the gap.
Snap cut ~1,000 jobs and 300 open roles, citing AI efficiency. AI now writes more than 65% of Snap’s new code. Stock rose 11% on the news.
A defense startup raised $82M to build drone factories inside shipping containers. The U.S. Air Force also debuted its AI wargaming tool, WarMatrix, in live use.
Per Stanford’s 2026 AI Index: as of March, Anthropic leads global model rankings, trailed by xAI, Google, and OpenAI. Chinese models are closing fast.