The Week in AI: Oct 19–23, 2026

Anthropic pitches investors on $16.8 billion quarter ahead of possible IPO

Anthropic is preparing investors for a possible public offering as soon as next month, after reporting third-quarter revenue of $16.8 billion, up 45% from the previous quarter. The company has discussed a potential IPO raising as much as $100 billion at a valuation near $2 trillion, though the timing could change, according to the New York Times and Wall Street Journal. The scale of the offering could influence investor confidence in AI companies and other technology IPOs.

Anthropic told investors it was profitable on an adjusted basis. Its reported figures included an 11% adjusted operating margin, excluding stock-based compensation, according to Simon Willison’s summary of investor disclosures. The company also said rival labs’ model distillation had accounted for 5% to 7% of revenue before it materially reduced the activity in the third quarter. Competition, market volatility and AI safety concerns remain potential obstacles.

Power delays and debt expose strain in AI data-center boom

The AI data-center build-out is facing a practical setback and rising financing scrutiny. Microsoft pulled out of a lease at Nscale’s planned U.K. site after grid problems delayed its expected opening; Nscale is seeking other power sources, while both companies say their broader partnership stands despite the site setback. The episode highlights the risk that delays can postpone revenue even as companies take on large commitments for infrastructure projects

Investors are separating data-center debt from that of computing providers, but neither is risk-free: CoreWeave’s bonds have come under pressure, and construction and tenant risks remain visible to bond buyers. Rising borrowing costs and inflation add strain across the sector for neocloud firms. The IMF warns a sharp AI pullback could threaten financial stability and global growth beyond stock markets worldwide. Debate persists over overbuilding versus whether customers will pay for useful AI, while investors await returns on heavy spending before confidence holds.

Judge detains former OpenAI engineer in alleged bugging case

A judge has ordered former OpenAI engineer Abdulrahman Alfozan held pending trial, with prosecutors citing his planned move to Saudi Arabia and his family’s departure from the United States as flight risks. They also point to Saudi contacts and assistance with travel and bail, while his lawyer says he is not charged with stealing trade secrets or acting as a foreign agent. The judge’s detention order Prosecutors’ account of Saudi assistance The alleged plans to relocate

The case began with allegations that Alfozan hid recording devices in a restricted conference room and planned to place more; he was arrested and fired. The FBI arrest and OpenAI’s response The alleged concealed devices Prosecutors describe Saudi advisory ties, but his lawyer disputes broader interpretations of the charge. His reported Saudi work The lawyer’s position OpenAI says it found no evidence sensitive technology or customer data was accessed. The reported lack of compromise The allegations have sharpened scrutiny of security at fast-growing AI companies, where insider threats could threaten valuable technology and investor confidence. Industry security concerns Risks facing AI firms

Meta and Apple bring Muse Spark Open to Mac Studios

Meta’s Muse Spark Open is now available as an open-weight model for coding, complex reasoning and agent-based tasks, with Apple’s MLX toolkit enabling local inference across multiple Mac Studios. The partnership could let developers run a very large model on their own hardware, though the setup is costly and the model has acknowledged limitations. Meta and Apple’s partnership

Meta says the 816-billion-parameter model can run across two M5 Ultra Mac Studios and is also offered for download and through cloud providers. Apple’s researchers report that four linked Mac Studios produce more than 36 tokens per second using a quantized version. Meta’s release details Apple’s performance tests

The local setup requires at least one terabyte of combined memory, with hardware costing more than $40,000, according to the Wall Street Journal. Meta also says Chinese open-weight models remain stronger in some areas, including software development. Thus, the release expands options for private, on-device AI, but access and performance remain constrained by cost and capability.

Anthropic launches Claude Duet for work tasks

Anthropic has introduced Claude Duet, a research-preview agent that monitors connected work apps, sends morning briefings and carries out tasks—pushing Claude beyond answering prompts toward acting on users’ behalf. The tool is aimed at professional work, not personal errands, and can connect to more than 6,000 services, according to the New York Times report.

Duet can draft replies, schedule tasks and prepare dashboards or forecasts, while continuing work in the cloud. It is available to US Pro and Max subscribers through a waitlist, and its background activity counts toward existing usage limits, The Verge reported and Anthropic’s help center explains.

Users must approve consequential actions such as sending messages or making purchases; safeguards also restrict some sensitive tasks. The preview can make mistakes, making oversight essential as AI systems take on more workplace responsibilities, Anthropic says.