Audit finds AI doubling Anthropic's pace; Nvidia's phantom buyer

The AI-safety research group METR spent eleven weeks inside Anthropic, from August 12 to October 29, and published its assessment on November 5. It concludes that Anthropic is likely improving its models on well-specified tasks at about twice the rate it managed between March 2024 and October 2025, and that roughly 80 percent of the current speedup comes from staff using AI to do the research itself. Anthropic's internal benchmark scores have grown about 1.9 times faster since October 2025, with a 95 percent confidence range of 1.5 to 2.8 times. METR found no clear bottlenecks to stop the acceleration from compounding, and it disclosed that Anthropic redacted a majority of the evidence its researchers found most informative, which Anthropic said was to keep confidential details from rivals. The report also says three unreleased Anthropic models finished training between July and October. It lands days after Anthropic's own IPO filing warned that its next models may speed up AI progress dramatically.

The New York Times found that New Matrix, one of the Chinese companies the US Commerce Department approved to buy Nvidia's H200 AI chips, lists an office that turned out to be "a hollow gray shell with wires dangling from the ceiling." The company was created about seventeen months ago, its building's directory does not list it, and its owners have ties to a Chinese private equity firm the US government has blacklisted since 2024. The Times described it as part of a pattern of approved buyers with links to restricted entities. Separately, the Times and the Financial Times reported on a memo based on US intelligence that claims Alice Huang, an executive at Megaspeed, a Singapore cloud company, was recruited by China's Ministry of State Security to spy on Nvidia chief executive Jensen Huang. Nvidia called Jensen Huang's interactions with the executive "not unusual" and said it does not share classified information with customers. Both reports arrive before the November 18 closed-door briefing House Republicans have asked Huang to attend.

Bloomberg reported that Microsoft has been switching off capacity it rented from Oracle since declining to renew a 2023 deal, because the older Nvidia chips involved had become less useful, and that it is reconsidering other deals with the smaller GPU clouds known as neoclouds. The Financial Times reported that Microsoft will not use an option to more than double its Texas deal with Nscale, a different site from the West Virginia campus Nscale is building for Anthropic. The Information tied delays at IREN's data center in Childress, Texas, to spending Microsoft pushed out of the September quarter. Microsoft's chief financial officer, Amy Hood, had said some expected third-party deliveries were late, and the company spent about five billion dollars less than planned. IREN shares fell 20 percent on the report and CoreWeave's fell 15 percent, even as Microsoft's own Azure revenue grew 48 percent last quarter.

John Langford, a co-founder of AI Frontiers, the roughly 50-person group formed inside Microsoft Research in 2023, wrote on November 6 that it was reorganized into Microsoft AI in early September and is now over. The group worked on making agents usable, including the AutoGen orchestration framework and the Fara computer-use agent. Langford wrote that Microsoft AI's leadership, which came from Inflection, is focused on scaling "and aggressively not research," and that "they just did not really want us." He has left, and he says the group's other founding leaders appear to have reached the same conclusion. He argues that research still pays, pointing to ByteDance experiments suggesting a variant of one of the group's architectures performs like a model twice its size. There will be no internships with the group next year.