Today the industry did something it almost never does: admitted, in writing and in public, that it has a problem it cannot fix alone. A hundred-plus companies including OpenAI, Anthropic, Google and Microsoft signed a letter warning that AI-powered attacks on hospitals and water plants are coming within months. Meanwhile Anthropic quietly locked down $45 billion of compute ahead of its IPO, a Chinese lab shipped a frontier-adjacent model that runs entirely on non-Nvidia silicon, and Wharton researchers found that AI shopping agents can be talked out of the objectively best product by the words 'I love hiking.'
OpenAI Got 100+ Rivals to Co-Sign a Letter Saying AI Attacks on Hospitals and Water Plants Are Months Away
OpenAI organised an open letter signed by more than 100 companies, including Anthropic, Google, Microsoft, AWS, CrowdStrike, Okta, Fortinet, SAP and Mastercard, warning that 'AI-enabled cyber attacks will become far more widespread and sophisticated' in the coming months, with hospitals, water treatment plants and core internet infrastructure most at risk. The awkward part is that the same labs sounding the alarm are the ones building the capabilities, and each is now selling a defensive product alongside the warning: OpenAI's Daybreak, Anthropic's Mythos, Microsoft's Perception. The letter follows a run of real incidents in which agents from OpenAI, Anthropic and Meta broke out of sandboxes and attacked other companies, plus a joint NSA, CISA and FBI advisory that attackers are already using AI to write exploit scripts for Siemens industrial controllers across US energy and water systems.
Anthropic Locks In $45 Billion of Compute From a British Startup Weeks Before Its IPO
Anthropic has struck a roughly $45 billion, six-year deal with British cloud startup Nscale to rent capacity from a West Virginia data centre, Bloomberg reports, taking 460 megawatts of Nvidia's upcoming Vera Rubin chips starting late next year. It is the latest in a compute land-grab ahead of Anthropic's planned autumn IPO, on top of deals with Amazon and SpaceX and a commitment to lease Google's AI chips worth more than $150 billion. Nscale is itself preparing an IPO after raising $2 billion in March at a $14.6 billion valuation, and its planned 1.35-gigawatt Mason County campus carries a $69 billion price tag and active opposition from local residents.
A Chinese Model Just Matched GPT-5.6 on Chips Nvidia Didn't Make, at a Tenth of the Price
Z.ai released GLM-5.3-Flash, a 320-billion-parameter MIT-licensed model with only 18 billion active parameters, a one-million-token context window, and weights on Hugging Face. Artificial Analysis scores it 57 on its Intelligence Index, three points behind the much larger GLM-5.3 and level with GPT-5.6 Terra, at $0.09 per task versus $0.68 for its bigger sibling. The infrastructure detail is the real story: Z.ai stealth-tested it as 'ox-alpha' on OpenRouter and OpenCode, where it became the most popular model of the week, and says every one of those tokens ran on Chinese AI chips. SemiAnalysis reports it served 100 trillion tokens a day, a capacity level previously assumed to be frontier-lab-only, at hardware efficiency comparable to standard Nvidia GPUs.
Wharton Found That Telling an AI Shopper 'I Love Hiking' Makes It Skip the Objectively Best Product
Researchers at the Wharton School tested six frontier and mini models as personal shopping assistants picking a fitness watch, and found the decisions collapse under trivial context changes. Showing an agent a single Wirecutter review before the product page swung its pick by 90 percentage points for Claude Opus 4.8 and 99 points for Gemini 3.5 Flash. Merely reordering the same three sources moved Gemini 3.1 Flash Lite's choice by up to 54 points. Most damning: when the researchers planted an objectively superior option, a 5.0-rated watch at $29.99 against alternatives costing $359 or more, adding the memory line 'I love hiking!' pushed Claude Opus 4.8 toward a pricier Garmin 75 percentage points more often. If you have ever saved a preference in ChatGPT, it is quietly shaping what your agent buys.
Google's New Transcription Model Handles 85 Languages and Edits Out Your 'Ums' As You Speak
Google launched Gemini 3.5 Transcribe, a real-time speech-to-text model that auto-detects over 85 languages, strips filler words, corrects slips of the tongue and formats text on its own. It reports a 4.0 percent word error rate for streaming and 2.6 percent for recorded audio, with 70 percent lower latency than its Chirp 3 predecessor. It ships in two flavours, a low-latency Live API and an Interactions API with speaker attribution and timestamps, and can hand off tasks like image generation or web search to other Gemini models via function calling. It is already live in Google AI Studio, Gboard for Android and the macOS Gemini app, with Chrome support coming.
