The current week of AI briefings, Monday to Sunday, in plain English. The weekly summary is on top; each day's full report is linked below. Updated every morning and again at market close on weekdays; a new week starts each Monday.
The week began as a finance story and ended as one. Anthropic's updated IPO paperwork showed about $33 billion in revenue over nine months - roughly twelve times a year earlier. By Friday, Reuters reported potential investors doubting Anthropic can hold next week's roadshow at a $1.5 trillion valuation after the year's sharpest tech selloff, and Ramp's data showed business AI spending growth stalling to 1.5% across August and September. By the weekend, Microsoft was quietly winding down the AI computing it rents from Oracle, part of a wider reassessment of its cloud partnerships.
Hours later, OpenAI closed the first $11 billion of its own new round at a $1.25 trillion valuation - below the $30 billion target reported in September. Around the giants, the money kept moving: Fal raised $1.3 billion at a doubled $16 billion valuation, Salesforce agreed to buy call-center voice-AI maker Cresta, and Anthropic signed the Associated Press as its first licensed news publisher.
The buildout turned political. Google is backstopping a roughly $19 billion lease so Anthropic gets 401 megawatts of Kentucky data-center capacity, and it announced a $17 billion Wyoming data center with its own fuel-cell power plant - promising to pay for 100% of the power it uses. A day earlier, after midterm elections put data centers on the ballot, OpenAI made the same pledge to towns: pay its own power bills, publish site-level usage, and drop official NDAs. By the weekend the election results had sunk in: voters approved more than 20 measures banning or restricting data centers, and the Wyoming project turned out to be Google taking over a stalled Crusoe development.
The research thread looked inward. Outside assessors at METR spent 11 weeks inside Anthropic and found its capabilities improving about twice as fast as a year earlier - largely, they believe, because AI is speeding up the research itself. Anthropic explained how Claude's memory works - agents, like people, consolidate by replaying their day while idle - and DeepMind showed AI agents finding attacks on legacy 1024-bit RSA encryption far faster than known methods.
The week's strangest thread was espionage: a reported US intelligence memo alleges a Megaspeed executive was recruited by China and grew close to Nvidia's CEO - unverified, with no charges, but enough to draw a congressional briefing request.
New models kept coming: OpenAI released GPT-6 Cyber, a security model it classifies in its highest risk tier, built for vetted defenders to find and validate software vulnerabilities. The open tier had a big week too - Tencent launched Hy4, and Moonshot AI released the weights of Kimi K3.1 Pro, a 2.8-trillion-parameter multimodal model, for anyone to download.
And products pushed toward everyday utility: Meta gave its Muse agent a memory upgrade and self-updating project workspaces called Spaces, Claude learned to read your bank account (read-only, through Plaid), ChatGPT learned to make animated GIFs, Google's AI video got longer and cheaper, Apple Intelligence got a China launch date, and OpenAI said ChatGPT now handles a fifth as many searches as Google.