Huang is out at Nvidia, Anthropic prices its IPO next week, Google launches Gemini 4 Hydrogen and nonprofits take the big AI labs to court.
Update: Jensen Huang steps down as Nvidia CEO; CFO Colette Kress takes over on an interim basis
Certain Business Chips
Huang resigned as CEO on Thursday 12 November, after the board reportedly decided the evening before to let him go. Colette Kress runs the company until a successor is named. The stock fell sharply, with market value dropping from above $4.5 trillion to about $4.0 trillion within a week. Reuters also reported a third relationship, which keeps the affair in the headlines.
Update: Anthropic to price its IPO on 18 November, with first trading on 19 November as 'ANT'
Certain Business
According to Anthropic, pricing is set for Wednesday 18 November, with trading starting Thursday 19 November. It is offering 825 million shares at $70 to $78, implying a valuation of roughly $1.3 to $1.45 trillion; earlier talk of up to $2 trillion was scaled back after turmoil in the sector. For the first nine months it reports $33.2 billion in revenue.
Nonprofits sue OpenAI, Anthropic, Google, Meta and SpaceX to halt frontier model development
High Policy Safety
A suit filed in San Francisco under California's unfair-business-practices law asks a court to bar work on models more capable than today's, citing the labs' own leaders warning about existential risk. Meta declined to comment; according to Reuters the other companies did not respond.
Update: Huang to testify behind closed doors on 18 November; chip-export bills to China gain momentum
High Policy Chips
According to Yahoo Finance, Huang has agreed to appear before Congress on 18 November, a day after Nvidia reports earnings on 17 November; Rep. Mast had earlier spoken of a date before Thanksgiving. Meanwhile chip legislation, including the AI Overwatch Act, is being discussed for the NDAA, and Commerce is probing whether Singapore-based buyers rented compute to Chinese firms on export-control lists.
Democrats open Nvidia investigation; Trump floats Tim Cook as Huang's possible successor
High Policy Business
Congressional Democrats have launched an investigation into Nvidia and Huang's ties to the Trump administration, according to TechCrunch with a sharply worded letter. Trump separately floated Apple chairman Tim Cook as a successor, which is a suggestion rather than an appointment.
Google launches Gemini 4 Hydrogen: a faster, smarter but pricier successor to Flash
Certain Models Products
The model is built for coding and multi-step agent tasks, offers up to 1 million tokens of context and, per Google, runs up to 2.5 times faster than Gemini 3.8 Flash. It costs $1.50 per million input and $7.50 per million output tokens, which VentureBeat puts at 15 times the cheapest rivals. Hydrogen is also arriving in Google Search and the Gemini app.
Major Microsoft outage hits Outlook, Teams and Copilot, including the OpenAI and Anthropic connectors
High Products Other
From about 1:26 p.m. US Eastern, users hit problems with Exchange Online, Teams, SharePoint and Microsoft 365 Copilot among others; Downdetector logged over 97,000 Outlook reports. Microsoft blames servers becoming unexpectedly unavailable and says it is redistributing traffic and restarting machines. At the time of reporting it was not yet resolved.
Anthropic brings Claude into Google Meet to draft documents and whiteboards during meetings
Medium Products
According to VentureBeat, Claude can follow along in a Google Meet call and draft documents and whiteboards as the discussion unfolds. It lands the same week as Claude Meetings, Sheets and Whiteboards, as Anthropic pushes further into office software.
Anthropic research 'Worse Together': group outcomes worsen when every user has their own A.I. agent
Certain Research Safety
The alignment team tested models from Anthropic, OpenAI and Alibaba in four multi-agent settings and found group outcomes got worse when each user ran a personal agent. Three of the test environments are being released as a benchmark, MAMUBench.
Cyberattack on AI vendor Fireworks AI lets hackers reach Notion customers' chat transcripts
High Safety Business
Notion confirms that hackers who breached Fireworks AI could access some customer chat transcripts; the company says fewer than 1% of workspaces were affected. The incident highlights how dependent companies are on third-party AI providers.
Bernstein halves Nvidia price target to $200, the third broker in two days to step back over Huang
Medium Business Chips
According to InvestingLive, Bernstein is the third bank in two days to turn more cautious on Nvidia. The stock fell as much as 13 to 14% intraday on Thursday after news of Huang's exit. Nvidia reports quarterly results on 17 November, with Kress expected on the call.
Google and Samsung open pre-orders for Gemini smart glasses; sales start 10 December
Certain Products
The glasses were developed with Samsung, Warby Parker and Gentle Monster. Warby Parker models start at $395, Gentle Monster at $449. Pre-orders are open in the US; pricing and availability elsewhere are not yet confirmed.
SpaceX plans over $74 billion for AI supercomputing campuses in Texas and Arizona, Reuters reports
Medium Business Chips
The Arizona site alone is reported at about $50 billion and still needs county approval. According to the report, the Texas campus is designed to generate its own power.
OpenAI launches security tools Daybreak Red Team and Defense Factory
Medium Products Safety
The tools probe applications for weaknesses and propose fixes. Red Team is in early access, and according to The Register patches from Defense Factory still need engineer review.
Attackers abuse 'AntiFox', an A.I. pentest fork of ARTEX, to steal AWS credentials
Medium Safety
According to Arctic Wolf Labs, via The Hacker News, attackers used the tool, hosted on compromised Alibaba Cloud systems, to grab AWS credentials after chaining IIS web shells. It fits a run of incidents involving agentic attack tooling.
China: Baidu opens API access to ERNIE 5.5 Flash Preview; ByteDance plans a Muse-like assistant within weeks
Medium Models Business
Baidu is giving developers access to a preview of ERNIE 5.5 Flash. ByteDance reportedly plans to launch an agent like Meta's Muse within weeks, per Bloomberg and TheNextWeb; founder Zhang Yiming expects frontier parity with OpenAI and Anthropic only around 2029-2030.
Meta brings its A.I. agent Muse to Apple Watch and Wear OS
Certain Products
Users can talk, text or dictate to Muse on the watch. Per TechCrunch the app reached 10 million US downloads in 40 days and can manage email, meetings and bills, and make reservations or order groceries.
OpenAI commits $2 million to research on A.I. and child safety
Certain Safety Policy
The money funds an international research network on learning impact, age-appropriate safety, risk measurement and safe use by children and guardians. OpenAI says it will publish all findings, including negative or inconclusive ones.
AMD details Instinct MI400 series: MI430X gets 432 GB HBM4 and native double precision
Certain Chips
AMD positions the MI430X (CDNA 5, up to 288 TFLOPS FP64) for scientific and sovereign AI workloads and shared new inference results for the MI455X on Helios racks. This is AMD's own claim; independent measurements are not yet available.
Google brings Gemini to Apple CarPlay; Gemini for Home now answers visually on smart displays
Medium Products
According to Bright, Gemini can now be used while driving through Apple CarPlay. Gemini for Home also now shows visual information with its answers on smart displays.
Test: AI shopping assistants sometimes point to risky online stores
Medium Safety Research
In brand-protection firm Podqi's tests, questionable sites appeared in 15% of Perplexity, 8% of ChatGPT and 4% of Gemini answers. This is one firm's measurement, not a universal rate.
United Airlines blocks AI agents from booking flights
Medium Policy Safety
The airline cites cybersecurity risks as the reason AI agents may no longer book on customers' behalf.
OpenAI customer story: Holiday Extras rebuilt its app with Codex in two weeks, versus six months before
Certain Products Business
According to OpenAI, the app's error rate fell from 19% to 9% and conversion from opening to booking rose 32.5%. Releases went from roughly once a year to weekly. These are OpenAI's own figures.
Huawei open-sources MACaQ: quantizes a 122-billion-parameter model in 64 minutes
Medium Research Models
The tool (Apache 2.0) reportedly runs on 32.5 GB of device memory and outputs 2- to 4-bit GGUF files, useful for running large models locally.
Update: alongside Gemini Robotics 2.5, DeepMind releases the open ASIMOV-Jailbreak safety benchmark
Medium Safety Research
Alongside the robotics model, which hit 96% in a Mercedes trial, comes a public benchmark for how easily robots can be talked into breaking safety rules. Details of the test design are limited in the source.
Next up: 17 Nov: Nvidia earnings; 18 Nov: Anthropic IPO pricing and Huang hearing; 19 Nov: ANT starts trading