Daily AI News — 13 Nov 2026

Nvidia said on November 12 that Jensen Huang has resigned as chief executive of the chip company he co-founded 33 years ago and will also leave the board. CFO Colette Kress was named interim CEO. The board moved after revelations of extramarital affairs, including one with a businesswoman tied to a Chinese company and Nvidia customer, tipped the company into crisis. Huang, 63, had become the most influential dealmaker in AI, shaping chip supply, export policy toward China and relationships with nearly every major lab and cloud. He was Nvidia's fourth-largest shareholder with a 3.61% stake. The exit sent AI stocks lower and left the industry guessing at what comes next.

Nvidia's chips sit under almost every frontier model, and Huang personally brokered the deals that steer where compute flows. Losing the person investors trusted to hold that complex system together injects real uncertainty into chip supply, export policy and the confidence propping up the whole AI build-out.

A startup founder planning a GPU purchase or a multi-year cloud commitment now has to price in leadership risk at their most critical supplier, hedging by lining up a second chip source or locking terms before any strategy shift under new management.

Anthropic said in an updated filing on November 10 that it will offer 825 million Class A shares at $70 to $78 each, raising roughly $57.75 billion to $64.35 billion. At the top of the range the company would be worth about $1.404 trillion on a fully diluted basis, up from a $965 billion private valuation in May. Separately, Anthropic agreed to privately sell $26 billion of non-voting stock at the IPO price to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC, with Google and Nvidia putting in $10 billion each. The company reported $17 billion in third-quarter revenue and an $81 billion run rate by the end of October. It plans to trade on Nasdaq under 'ANT' around November 19.

A trillion-dollar listing for a six-year-old AI lab tells you how much capital markets now believe frontier models are worth, and the chipmakers and cloud giants buying in signal the whole supply chain is doubling down together. It also sets a public benchmark every rival will be measured against.

A builder choosing which model provider to standardize on can read Anthropic's disclosed revenue, losses and compute commitments to judge how durable the company and its pricing are before wiring a product's core to Claude for years.

Google released Gemini 4 Hydrogen on November 12, pitched at developers who need a quick model for writing code, working through documents and running multi-step agents. Google says it finishes tasks 2.5 times as fast as Gemini 3.8 Flash at its default reasoning setting, and independent testing from Artificial Analysis scored it 51 on their Intelligence Index, up from 31 for Flash. On Google's own terminal benchmark it completes 57% of tasks versus 18.2% for Flash, though it still trails Claude Sonnet 5.5. The catch is cost: at $1.50 per million input and $7.50 per million output tokens it runs about 15 times the cheapest rates from rivals like Claude Haiku, GPT-6 Luna and Muse Ember, so teams pay for the extra capability.

A lot of production AI is routine agent work, where speed and reliability matter more than raw genius. A model that recovers better from failed tool calls and nearly triples terminal-task success can quietly replace slower, pricier reasoning models for the bulk of that work, reshaping which model teams default to.

A team running a coding assistant that keeps stalling on multi-step tasks swaps in Hydrogen, gets faster completions and fewer dead-ends when a search returns empty, and reserves a heavier reasoning model only for the genuinely hard problems.

Google DeepMind introduced AlphaProof Code on November 10, an agentic framework that used Gemini Carbon-powered agents to mathematically prove key security properties of the actual compiled code behind the Signal protocol, which protects billions of private conversations. The agents generated over six million lines of Lean proofs in four weeks, then flagged just one percent of the codebase for human review while Lean automatically checked the rest. Unlike most protocol verification, which stops at the design level, the work proved the real implementation holds its guarantees down to the running code. To confirm the system can catch bugs, researchers reintroduced two previously fixed flaws and it proved both versions insecure. DeepMind is open-sourcing the verification artifacts.

Most real-world security holes hide in the gap between a clean design and the messy code that ships. Proving production cryptography correct at this scale, in weeks instead of years, points to AI making the highest standard of security assurance practical for critical systems that could never afford it before.

A team shipping firmware for a medical device or payment system could, in time, point AI agents at their compiled code to generate machine-checked proofs that security properties actually hold, catching implementation flaws that design-level audits and testing routinely miss.

United Airlines said on November 11 that it will restrict unauthorized third-party AI agents and assistants from searching, booking or managing flights on its platforms, directly and indirectly through travel agencies and aggregators. The carrier argued that handing an agent your MileagePlus credentials and payment details risks exposing sensitive information and triggering unintended purchases or account changes, and that letting a bot handle check-in could violate FAA and DOT rules, since a human must confirm their bag holds no banned items. Bookings agents made before November 11 will not be canceled. United said it does not yet work with any third-party AI agent but is open to standardized, secure integrations later. Fares stay bookable through aggregators like Kayak and Google Flights.

This is the clash everyone building agents saw coming: autonomous assistants that act on your behalf versus platforms that want to own the customer and control security. How big brands like United, Amazon and Walmart split between blocking and partnering will shape whether consumer agents can actually transact or stay stuck at planning.

A startup building a travel-booking agent now has to design around platforms that refuse unauthenticated bots, routing United fares through permitted aggregators or pursuing official integrations rather than screen-scraping the airline's site with a user's login.

Anthropic announced a slate of workplace tools on November 10, led by Claude Meetings, which lets teams add Claude to a video call as a visible participant. As people talk, Claude follows the conversation and gets to work: drafting the plan being discussed, mapping a timeline on a whiteboard, answering questions in the meeting chat and pulling data from connected work apps. Afterward the notes, action items and anything Claude built are available in claude.ai. Claude identifies itself and posts a notice in the chat, requires admin approval to stay in calls with external guests, and otherwise leaves automatically. Anthropic also launched Claude Whiteboards and Claude Sheets, and made Claude Data generally available. Meetings is in beta on Team and Enterprise plans, starting with Google Meet.

Most of a meeting's real output starts only after everyone hangs up. Pushing an AI into the live conversation to produce plans, timelines and pulled-up figures in real time is a bet that the next productivity battleground is the meeting itself, where Anthropic now takes on Copilot, Otter and Granola.

A product team on a planning call adds Claude, which drafts the roadmap and timeline as they debate it and surfaces last quarter's numbers on request, so the group walks out with a reviewed draft instead of a to-do list of notes to write up later.

Runway released Praxis-1 in November, an open-weight world-action model that repurposes its 27-billion-parameter video foundation model to control real robots. A small action expert reads the video model's internal state as it imagines what comes next, so each time the robot acts, Praxis-1 predicts both how the scene will look and the motion to get there. On NVIDIA's RoboLab-120 simulation benchmark, Praxis-1 fine-tuned on DROID solved 46.2% of episodes, the top score among world-action models and open-source robot foundation models, ahead of every open-weight rival on the board. On a real Franka arm it succeeded in 88 of 100 trials across everyday tasks like placing a mug and clearing bottles into a bin, computing each 1.6-second chunk of motion in about half a second on one H100. Runway is releasing three checkpoints under Apache 2.0.

Robotics has been bottlenecked by scarce data and heavy per-task engineering. Showing that the huge investment already poured into video generation can transfer into fast, capable robot control, and shipping it as open weights, gives the whole field a head start instead of forcing every lab to collect robot data from scratch.

A robotics team building a tabletop manipulation demo downloads Praxis-1, fine-tunes it on their own arm with a small dataset, and gets usable real-time control in everyday pick-and-place tasks without training a policy model from the ground up.

ElevenLabs launched two speech-to-text models on November 10: Scribe v3 for recorded audio and Scribe v3 Live for real-time. The company reports Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages and has the most accurate speaker labels of any model it tested, cutting diarization errors 17% on recordings with three or more speakers. Scribe v3 Live makes 43% fewer errors than the previous real-time model, adds speaker diarization for the first time, and uses conversational context rather than mere silence to detect when a speaker has actually finished, so voice agents stop cutting callers off mid-sentence. The models run at a median 150 ms latency, support a keyterm library and a multilingual code-switching mode, and switching from v2 is a one-line change to the model ID. A healthcare-tuned version is available in ElevenAgents.

Every audio product, from meeting notes to voice agents to clinical documentation, starts with a transcript, and every error there cascades downstream where it costs more to fix. Sharper real-time transcription with turn detection and speaker labels directly fixes the two things that make voice agents feel broken: wrong words and talking over people.

A team building a phone support agent swaps in Scribe v3 Live so the bot waits for callers to finish a thought instead of interrupting during a pause, and correctly labels who said what in a three-way call, making the whole interaction feel less robotic.

Norwest said on November 12 that it has closed a $4 billion fund, its eighteenth and largest in 65 years, up from the $3 billion vehicle it closed in 2024. The firm, backed solely by Wells Fargo, will invest across the United States, India and Israel and now manages more than $18 billion. Senior managing partner Jeff Crowe told TechCrunch the firm is 'taking a broad approach to AI because its impact is broad,' naming early-stage companies building AI-native applications and infrastructure, physical AI, domain-specific models and cybersecurity as priorities, with further expansion into voice AI, agentic cybersecurity and aerospace. Recent AI bets include Coval, which tests and evaluates AI agents, and Generalist AI, which builds foundation models for robots. Norwest's portfolio also includes Plaid, Spotify, Uber and Vuori.

Even with a broader pullback from AI investing, one of the steadier multi-stage firms is putting its biggest-ever fund to work and explicitly spreading it across applications, physical AI, domain models and agent security. Where patient capital like this lands signals which corners of AI investors think will still matter after the hype cools.

A seed-stage founder building a domain-specific model, a voice-AI tool or security for AI agents now has a concrete, well-capitalized firm whose stated thesis matches their category to target for a raise, rather than pitching generalist funds chasing the same few frontier labs.

VentureBeat reported on November 12 that CrowdStrike is retiring 'breakout time,' the long-standing metric for how long defenders have before an intruder spreads, because AI agents have collapsed it. In one case inside an Argentine healthcare organization's network, 29 seconds after firewalls blocked its first attempt, an AI agent switched tactics and targeted the domain controller. CrowdStrike's reported eCrime breakout average fell from nearly two hours in 2018 to 29 minutes in its 2026 report, with the fastest recorded breakout at 27 seconds. AWS said one of its honeypots caught an AI agent completing a full cyberattack in 12 minutes and 42 seconds. Security leaders now argue the window for human response has effectively closed, but there is no agreed replacement metric for measuring defense in the agent era.

Security planning has assumed humans have minutes to hours to contain a breach. If AI-driven attacks move in seconds, every playbook built on human reaction time breaks, forcing defenders toward automated, AI-speed response, and the lack of a replacement metric means teams are flying partly blind on how fast they really need to be.

A security team that still measures itself on human response times starts deploying automated containment that can isolate a compromised host the instant anomalous agent behavior appears, rather than paging an on-call engineer who cannot possibly react inside a 27-second window.