AI Weekly Roundup: Nvidia CEO Exit, Anthropic IPO, Wallets

I read the primary posts for this AI weekly roundup on Sunday, 15 November. A board statement, an 8-K, a roadshow notice, two security advisories, a Chrome release note and a few research pages. I did not run a model, try a wallet or let Claude into a meeting. The one thing I measured myself was a count of names in that Chrome release note.

The week has one shape: the founder left, the agents got a wallet. Jensen Huang resigned as Nvidia’s CEO on Thursday, in a week when the stock fell 13% according to CNBC. Anthropic put a price range on its IPO and, per CNBC, seven chip and cloud suppliers will buy $26 billion of its stock at that price. OpenAI gave its agents a card that needs a fingerprint for every purchase. Google made its fast model smarter and pricier at once. Security was louder: an inference provider breached by people who resell stolen LLM access, Notion AI chats leaked through that vendor, and Cloudflare counting 30 times more AI pentest traffic than on 1 October. Builders got llama.cpp 1.0 and a robot model that learns by watching people. The rest lives on the AI hub.

🪑 Nvidia without its founder

NVIDIA’s board announced on Thursday 12 November that Jensen Huang resigned as President, CEO and board member, “effective immediately”. The 8-K shows the board accepted it a day earlier, on 11 November.

Colette Kress, EVP and CFO since September 2013, is interim CEO and keeps the CFO job until a permanent successor is named. The search “will consider internal and external candidates”. Harvey C. Jones, former CEO and Executive Chairman of Synopsys and, per the 8-K, a director since 1993, becomes the board’s “first-ever Chair”. Jones says the board expects “the company’s operations, customer commitments and product roadmap to continue without interruption.”

How it got here, in the board’s words only. On 8 November the board opened an independent investigation “in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort”, saying it had “no reason to believe that either national security or NVIDIA secrets were compromised”. NBC News reported on 13 November, citing sources, that the independent directors plan to continue that investigation. The rest of the coverage is about a person. This bulletin is about a company.

The market was not patient. CNBC says shares fell 13% this week, the most since April 2025, a sixth straight weekly decline and the longest losing streak since 2012. By CNBC’s tally Nvidia has lost about 30% and roughly $1.7 trillion in market value since early October, and slipped below Apple. CNBC cites LSEG: analysts expect Q3 revenue of $108.9 billion, up 91% year on year.

If you build on CUDA, the roadmap is now a board promise rather than a founder’s habit. The first test is close. Q3 fiscal 2027 results are still due on Tuesday 17 November, around 00:20 Istanbul time on Wednesday, with the call at 01:00. Kress will read the numbers and run the company on the same night. Caveat: nothing in the filing moves a product date, and nothing confirms one either. “Without interruption” is a sentence, not a schedule.

💵 What is Anthropic worth on paper?

Anthropic launched its roadshow on Tuesday 10 November. Its post lists 825,000,000 Class A shares: 801,351,354 sold by Anthropic and 23,648,646 by existing stockholders, which bring Anthropic nothing. Expected price is $70.00 to $78.00, on the Nasdaq Global Select Market as ANT. Nasdaq’s notice says the shares “will begin publicly trading on November 19, 2026”.

CNBC, citing people familiar, says pricing is planned for 18 November with about 20% aimed at retail investors, where a typical IPO gives them 5% to 10%. At the top of the range CNBC puts proceeds at up to $64.35 billion and the fully diluted valuation at up to $1.405 trillion, against $965 billion in the May funding round.

The line I kept rereading is from the amended filing, via CNBC. AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC agreed to buy $26 billion of Class N common stock in a private placement at the IPO price. Those shares are not registered in the offering and sit outside the underwriters’ lock-up. Yes, Nvidia, in the same week it lost its CEO.

The financials, also via CNBC: revenue of $33.24 billion for the nine months to 30 September, up about 1,200% year on year, and a $104 billion net loss, mostly from fair value changes on convertible notes and derivative liabilities. Run-rate revenue was about $81 billion at the end of October, up $10 billion in a month. Last week’s weekly went through the S-1/A’s RSI and margins, and the day before the roadshow Anthropic shipped Claude Fable 5.2 and Mythos 5.2 (my deep dive).

As someone who buys tokens, here is what I will remember. The chip and cloud suppliers are now shareholders at the IPO price, so read every “partnership” announcement between them and Anthropic with that in mind. Caveat: the price is a range until 18 November, the registration statement is not yet effective, and $1.405 trillion is a top-of-range scenario, not a final valuation.

💳 Agents get a wallet

OpenAI introduced ChatGPT Wallet on Friday 13 November, built with American Express, Mastercard and Visa. It arrives first through dots, OpenAI’s always-on agents, for ChatGPT Pro subscribers in the US. OpenAI calls it “the first wallet that lets agents pay without access to users’ actual card details and that requires a security check for every purchase”.

You add a card once. Before each purchase you review the items, the total and the card, then approve with Face ID or a fingerprint. The payment network issues a payment token (a substitute card number, security code and expiry), so neither the business nor the agent sees the real card. Rewards and eligible dispute protections stay. The Help Center adds the fine print: 18 or older, US billing address, eligible US-issued American Express, Chase, Mastercard or Visa cards, and only in conversations with a dot, not in Chat, Work or standalone Codex threads.

OpenAI says 1.2 billion people use ChatGPT each week, which makes this pilot a small door into a very large building. Spending limits and pre-approved purchases are things OpenAI expects to add, not things that shipped (The Verge has more).

Three days earlier OpenAI opened Businesses in ChatGPT in beta. Business Hub lets eligible businesses claim their profile and add product feeds with pricing and availability, Shopify merchants included. They get aggregated insights: mentions, website clicks and “discovery topics” where their site is cited. OpenAI says it does not share users’ conversations with businesses. Eligible US businesses can also build a labeled Business Agent and sponsor it as the destination for ChatGPT Ads.

For site owners this is SEO’s next storefront. ChatGPT now tells you how often it mentions you and lets you pay to be the agent people talk to. OpenAI also wants businesses to make sure agents are “not inadvertently blocked”. In plain terms, if your checkout treats every bot as fraud, it may soon be turning away paying customers. Caveat: Pro only, US only, cards only. It is a pilot, and a fingerprint per purchase is the opposite of autonomous.

Glass robotic hand offering a shopping cart while a human thumb approves on a glowing fingerprint pad — ChatGPT Wallet asks for a fingerprint or Face ID before every purchase an agent makes. Illustration: Higgsfield.

⚡ Gemini 4 Hydrogen: faster, smarter, not cheaper

Google introduced Gemini 4 Hydrogen on Thursday 12 November for “rapid execution across diverse agentic workflows”, next to the frontier Gemini 4 Carbon and Gemini 4 Argon. The name retires the Flash branding.

Google’s developer post lists $1.50 per million input tokens and $7.50 per million output tokens, flat across the whole 1 million token context window, with four thinking levels. Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high level, against 50 for Argon, and 54 at “xhigh”. VentureBeat reports Google’s claim of tasks finished 2.5 times as fast as Gemini 3.8 Flash, which Artificial Analysis had at 31 at high reasoning. VentureBeat calls the new model pricier than Flash.

It is in the Gemini API, AI Studio, the Gemini app (free users via auto model selection), Google Search for everyone and Antigravity for paid subscribers. The Search post calls it “the biggest model upgrade to Search yet”, with more query fan-out and “a broader range of helpful links”. If you run a website, watch that last phrase.

Gemini 4 Carbon opened to paying developers two days earlier (our Carbon and Muse Ember digest). Alibaba’s Model Studio doc describes Qwen4-Max-Preview as a temporary preview, live 11 to 25 November (UTC+8), with a 1M token context and Token Plan access only.

The fast tier is no longer automatically the cheap tier. A flat price over 1M tokens is easy to budget, so put it next to your current bill before switching. Caveat: the index numbers are Google quoting Artificial Analysis.

Prices per million tokens, input / output, from each vendor’s page or announcement.

Model	Company	Date	Price per 1M tokens (in / out)	Context	Access
Gemini 4 Hydrogen	Google	12 Nov	$1.50 / $7.50 flat	1M	API, AI Studio, Gemini app, Search
Gemini 4 Carbon	Google	10 Nov	$2 / $10 to 30 Nov, then $4 / $20	n/a	Paid API, AI Ultra
Claude Fable 5.2	Anthropic	9 Nov	$8 / $40	1M	API, Claude apps
Muse Ember	Meta	9 Nov	$0.10 / $0.50 (up to 200K input)	n/a	API
Qwen4-Max-Preview	Alibaba	11 Nov	No pay-as-you-go price, Token Plan only	1M	Model Studio, until 25 Nov

🧯 Who was holding your AI chats?

This one starts at Fireworks AI, an inference provider, and ends inside Notion. The Hacker News recaps the first disclosure on 10 October: stolen internal credentials gave access to environment variables, including API keys and tokens stored as secrets, in some customers’ evaluation jobs. Fireworks’ new bulletin, published Monday 9 November, says attackers came in on 23 October through an unnamed third-party cloud provider hosting its R&D machines, reached development machines on 24 October and ran an unauthorized workload. For a “limited set of accounts” that exposed resource metadata, billing information and a subset of incomplete model weights and datasets. Fireworks says it does not store card numbers.

The motive is the interesting part. Fireworks ties it to a financially motivated group “aimed at selling tokens from unauthorized access to large language models (LLMs) to illicit routing gateways”. That is LLMjacking. Someone steals your model access, resells it, and you get the invoice.

Notion’s advisory says fewer than 1% of workspaces were affected. Notion disconnected from Fireworks on 3 November, a day after hearing about access to synthetic test data, and contacted affected customers on 9 November. The dataset included Notion AI chat transcripts used for the Insights feature, “unattributable” for the vast majority of affected workspaces, with no author or workspace label. Notion disabled Insights, stopped serving open-weight models through Fireworks, revoked exposed Notion API keys on 11 November and reported exposed third-party keys to their issuers on 12 November. Its advice: rotate any password or key you pasted into Notion AI. TechCrunch covered the Notion side.

The vendor behind the vendor is now part of your threat model. The AI feature you use runs on a model host you never signed a contract with, people paste keys into chat boxes, and those chats become someone’s training or classification data. Caveat: Fireworks calls the impact on its inference platform customers limited, and nothing published says inference traffic was exposed.

🕷️ Fifty attack frameworks in three days

ARTEX, an open-source agentic pentesting system, was taken down by its creator on 8 October. Last week I wrote about the roughly 815 ARTEX consoles left open. Deleting the repo did not delete the idea. Cloudflare’s Cloudforce One counts over 12,000 repositories since then, several hundred unique forks and hundreds of similar orchestration frameworks, with 204 public repositories carrying their own code changes.

Then the sentence that made me sit up: “In just the last three days, we have seen over 50 framework families attempt to scan Cloudflare customers globally.” The IPs number “in the thousands”, and daily requests from these tools have “grown by 30x since October 1.” Some attacks chain stolen customer credentials for lateral movement. Cloudflare’s explanation is the best line of the week: the tools “turn compromises into a hill-climbable problem, which agents are specifically good at.”

Its fix is a six-pillar “Modern Security Model”: WAF with rate limits, anomaly detection, bot defense, identity-gated access for high-risk apps, validated schemas and threat intelligence. Good advice. Also, by remarkable coincidence, a Cloudflare product list.

Defenders were busy too. Chrome 157 shipped on 11 November (157.0.8103.78/.79 for Windows and Mac, .78 for Linux) with 83 security fixes, five of them Critical, all use-after-free. By my count of Google’s release note, 8 entries credit Anthropic researchers, and 5 of those say “assisted by Claude”. Another 3 credit “OpenAI Codex Security”. The note does not say any of these bugs were exploited in the wild.

CISA’s Known Exploited Vulnerabilities catalog grew all week:

9 Nov: four, including CVE-2026-123158 in OpenBSD OpenSSH and CVE-2026-102895 (improper authentication) in GPUStack, a GPU cluster manager for running AI models.
10 Nov: six, including CVE-2026-124194 in Microsoft Office Outlook, CVE-2026-125350 (use-after-free) in Google Chromium and two Barracuda flaws.
13 Nov: two, UnrealIRCd CVE-2026-114713 and Check Point CVE-2026-123039.

The same week attackers had 50 agent frameworks, defenders’ models were filing Chrome bugs. The race is symmetric. The patching is not. Update Chrome. Caveat: Cloudflare’s counts come from its own network and nobody outside can check them.

Broken root sprouting dozens of glowing forked branches tipped with scanning drones aimed at server towers — Cloudflare counts over 12,000 repositories descended from or inspired by ARTEX. Illustration: Higgsfield.

🛠️ Builders’ corner: llama.cpp 1.0 and a robot that watches people

On 13 November the project tagged llama.cpp v1.0.0 with ggml v1.0.0, and the Hugging Face post explains the numbers. A major bump now means a breaking change to the public C API in ./include. Minor means backward-compatible features, backend updates, model support or API additions. Patch means a bug fix. After 3.5 years the project supports more than 150 architecture variants and “all mainstream speculative decoding methods”, and ships through Docker, Homebrew, Winget and more. If you ship on top of libllama, you finally have a contract. Pin to 1.x. Local model fans also have the Edge0-145B deep dive.

At CoRL (The Robot Report has the context), NVIDIA’s GEAR lab showed GR00T 2, a “world-action model” where one diffusion transformer predicts future video and robot actions together, across grippers, dexterous hands and humanoids. NVIDIA’s numbers: 40% zero-shot success on 50 tasks never seen on the target robot, rising to 95% with five demonstrations per task (under 3 minutes of robot data) and a 30-minute fine-tune on a single RTX 5090. The pretraining mix explains the “watches people” part: 402,730 hours of egocentric human video, 10,184 hours of handheld UMI gripper demos across 2,897 tasks and 23,500 hours of teleoperation. 100,000 hours of the human video are open on Hugging Face as Human-1M-100K. GR00T 2 Turbo runs on the Unitree G1’s own Jetson Thor at 124 ms per prediction. Caveat: the lab page does not list a weights download yet.

World Labs released Atlas, “our next generation world model”, to the public on 12 November, with tools and developer APIs to generate and reconstruct 3D worlds. The company joined AMD, as its blog announced on 28 September.

Robot hands assembling a wooden toy truck while ghostly human hands show the same step above them — GR00T 2 learns from human video, then needs only a few robot demos per task. Illustration: Higgsfield.

📌 The rest of this AI weekly roundup

Claude Meetings beta (10 Nov). Claude joins Google Meet calls as a visible participant, writes notes and action items and builds docs and whiteboards live, on Team and Enterprise plans. It stays out of calls with outside guests by default and saves transcript, summary and screen descriptions, not audio. Related: our Claude agent memory digest.
Adobe MAX (10 Nov). Creative Agent across Premiere, Illustrator, InDesign, After Effects and Frame.io, plus a preview of local MCP integrations that let third-party agents edit native Photoshop, Illustrator, Premiere and After Effects files. Private beta in early 2027.
EU AI Act (12 Nov). The Commission published Article 50 guidelines: providers mark AI content in machine-readable form, deployers disclose deepfakes and AI text on public-interest matters. Article 50 applies from 2 August 2026.
Already covered here: Claude Fable 5.2 and Mythos 5.2, Gemini 4 Carbon and Muse Ember, Edge0-145B.

🗓️ What to watch next week

Nvidia reports Q3 on Tuesday 17 November, with Kress on the call at 01:00 Istanbul time on Wednesday. Anthropic plans to price on 18 November and trade as ANT on 19 November. Qwen4-Max-Preview retires on 25 November, and Gemini 4 Carbon’s introductory price ends on 30 November.

If an agent asks for your fingerprint next week, at least you will know which wallet it means.

📚 Sources

Oğuzhan Koçaklı

Oğuzhan Koçaklı writes and advises on AI engineering, agents, GenAI products, and applied ML. Daily digests and deep dives in EN + TR at oguzhan.co.

All posts

AI weekly: Nvidia without Jensen Huang, Anthropic’s IPO price, and agents with a wallet