
A Farewell to Chips
Week 46, 2026: Ernest Hemingway walks the week. Jensen Huang resigns and his staff want him back, stolen keys open doors all over the software world, and the agents learn to pay and to make phone calls.
In the voice of Hemingway - Sparse, hard, declarative. Short sentences. The iceberg under the surface
Week 46, November 9 to 15, 2026. Told in the voice of Ernest Hemingway.
That week the big man left the company he built, and the doors stayed open everywhere else. Nobody had locked them. Some had been open a long time.
The man who built it
Jensen Huang started Nvidia in 1993 and he ran it for thirty-three years. On Thursday he resigned as chief executive, as president and from the board, and it was effective at once. Colette Kress, the finance chief, became interim CEO and kept the old job too. Harvey C. Jones became the first independent chair in the company's history.
It went fast after the letter of the week before. On Wednesday the Wall Street Journal reported that the board was getting ready to push him out. The same day the New York Times reported a second affair, with an Nvidia executive who reported to him. On Thursday Reuters reported a third. The first woman, Alice Huang, had worked at Megaspeed, a Singapore company US officials suspect of moving Nvidia chips into China. The Journal later reported from documents that Megaspeed arranged nearly $12 billion in Nvidia chip orders for Chinese companies between March 2024 and June 2025. Megaspeed says she left in August 2025 and that its business was "legitimate and lawful." The spy allegations against her are not confirmed.
The market did not wait for confirmation. The Decoder counted about $1.4 trillion of value gone since the first spy report on October 29. On Friday Nvidia fell below $4 trillion for the first time since July 2025. It was the sixth week down in a row.
Then a strange thing happened. The people who worked for him wanted him back. Wired saw an internal post asking the board to listen to the staff, and it drew more than 1,600 comments, most of them for his return. "If you believe in the mission, bring back THE boss," one of them wrote. A public petition passed 115,000 signatures, later 150,000, though Business Insider found names on it like "Tim Cruise, Hollywood actor." The Financial Times reported that Huang had talked about coming back with close allies.
There was no succession plan. More than sixty people reported straight to him. The board told CNBC it was still choosing a search firm, and the FT reported it had approached Arm's Rene Haas, who spent seven years at Nvidia. In Washington, both parties want tighter chip export rules now, and Huang agreed to testify.
A former OpenAI researcher, Will DePue, added a quiet fact to the noise. Model weights carry a faint mark of the chips that trained them, and six Chinese models released since April carry the mark of Blackwell, the chips the US banned from China. The mark shows the hardware. It does not show where the chips were or who owned them. And on Sunday Meituan released LongCat-2.5, a 1.6-trillion-parameter open model trained on Chinese hardware, with a hosting guide for Huawei chips and none for Nvidia.
For builders, nothing in your GPU stack changed this week. Watch who the board picks. Watch the export rules more.
The open doors
It was a bad week for keys.
Starting Tuesday, someone used stolen credentials for 4,837 GitHub accounts to push a secret stealer into every repository those accounts could write to. It ran in at least 1,806 repositories and took what the CI workflows could reach: Actions secrets, tokens, cloud credentials, AI service keys. It came in quietly, as a pull request about "dependency audit notes." In some repos it could rewrite an AWS role's trust policy so that rotating the keys would not be enough. Accounts hit in October's wave were about a thousand times more likely to be used again.
Cursor told the rest of a story it had begun in October. The poisoned extension on its marketplace started with a key it had given Fireworks AI. The attacker took the key in the Fireworks breach, read old coding agent traces, found credentials left in them, and walked into Cursor's tools. The bad build was live for about eleven hours and reached at most 5,000 users. Cursor moved all its inference off Fireworks.
At Railway, keys that customers stored ran up an estimated $740,000 in AI charges after an October breach. Arctic Wolf tied AntiFox, an open-source AI pentesting tool with a few hundred stars on GitHub, to real intrusions that stole AWS credentials. Nobody needed a clever trick. The keys were lying where someone could pick them up.
Then there were the doors nobody had meant to open. Claroc counted 37,571 apps built with Grok Build and found 94.7% of them open without a sign-in. The publish dialog had "Anyone with the link" picked by default. Twenty-four of the open apps belonged to staff at the company that makes Grok, and one held a sales map of 594 accounts. It is hard to write a better sentence about defaults than that one.
The lesson is old and it is not interesting and it is true. A self-hosted runner is a saving, not a wall. A key you gave a vendor is your key still. Look at the sharing setting before you look at the code.
The agents go out
The agents went out into the world this week and the world did not always open the door.
OpenAI gave its agents a wallet. ChatGPT Wallet lets the dots, its personal agents, find a thing and buy it. You approve each purchase with your face or your finger, and the card network makes a one-time stand-in number, so neither the agent nor the store sees your card. For now it is for Pro subscribers in the US, eighteen and older. OpenAI tried once before and dropped it.
Meta's Muse can now phone businesses for paying users, and many businesses hang up when they hear it is not a person. United Airlines went further and said no to booking by agents at all. It gave three reasons: your data, the federal rule that a human must say the bag holds no explosives, and the chief executive's own plain one. "To be honest, I don't want to create another middleman," Scott Kirby said. The demand is small so far. In Halloween week Uber handled 320 million trips and deliveries, and agents booked 500 food orders and four rides.
OpenAI also opened a Business Hub where companies claim a profile and put their own agent in ChatGPT to answer questions. Everyone wants to be the place where the buying starts.
If your agent buys things for people, plan for the sites that will refuse it. Show the choices and give the last click back to the human.
The price and the bill
The models came cheaper this week, and the bill did not always follow the price.
Meta released Muse Ember, its smallest model, built for the repeated work of agents, at a third of the cost per task of the model it was aimed at. On real coding it lost, 27.2% to 34.9%. Google's Gemini 4 Hydrogen, the new fast model, beat the bigger Argon on the Artificial Analysis index while Argon was less than four weeks old. Gemini 4 Carbon took the top of MathArena from GPT-6.1 Sol and cost about fifteen times as much at the tested settings. The model in second place cost about forty times as much. One of Carbon's scripts ran more than eleven hours with nobody watching.
The number that matters is the cost of a finished task, on your tasks. Run three of your own before you switch. The leaderboard does not pay your bill.
The money
The money kept moving in large amounts. Anthropic's IPO was about six times oversubscribed, Bloomberg reported, aiming at up to $64.4 billion, which would be the largest offering ever, with pricing set for November 18. Banks turned away buyers from mainland China and Hong Kong. OpenAI's quarterly revenue crossed $10 billion for the first time and it still lost money on operations. LangChain raised $300 million at a $3 billion valuation, and Mercor raised $750 million to sell evaluations to companies.
And the Federal Reserve asked the people who watch the financial system what worried them most. 87% said AI, up from 50% in the spring. They named the valuations, the cyberattacks and the tangled financing. They did not name the man at the top of Nvidia, but they did not have to.
What to do
Go through every repository you can write to and look for commits since November 10 that you did not make. Then rotate every secret those workflows could read, and do it today, not when you have time. The man who comes through an open door does not get tired, and he does not take weekends off.
Sources
25 events · 148 source articles
- Is bad news for Jensen Huang also bad news for the AI industry?"Jensen is Nvidia, right? And so much of this economy is riding on Nvidia."📰 News:techcrunch.comtheverge.comtheverge.comthe-decoder.comtechcrunch.comtechcrunch.comtechcrunch.comtheverge.comtechcrunch.comthe-decoder.comria.rusiliconangle.comtheverge.comtechcrunch.comtheverge.comwired.comsiliconangle.comtechcrunch.comtechcrunch.comtechcrunch.comtechcrunch.comtheverge.comtechcrunch.comtechcrunch.comtheverge.comthe-decoder.comtheverge.comapnews.comtechcrunch.comtechcrunch.comtechcrunch.comtechcrunch.comtheverge.comcnbc.comtechcrunch.comwired.comtechcrunch.comtheverge.comtechcrunch.comtheverge.comtechcrunch.comtechcrunch.comtechcrunch.comtheverge.comarstechnica.comthe-decoder.comtechcrunch.com
- Anthropic's Claude Fable 5.2 doesn't generate images, it paints themAnthropic's Claude Fable 5.2 doesn't generate images the way other AI models do. Instead of computing pixels, the language model writes Python code that places hundreds of thousands of simulated...📰 News:the-decoder.comthenewstack.ioinfoq.comthe-decoder.comthenewstack.iothe-decoder.comthenewstack.iothenewstack.iothe-decoder.comthenewstack.iothe-decoder.comthe-decoder.comthe-decoder.comthe-decoder.comsiliconangle.comthenewstack.iocnbc.comsiliconangle.comthenewstack.iotechcrunch.comthe-decoder.comengadget.comtechcrunch.comthe-decoder.comtheverge.comthe-decoder.comtechcrunch.comsiliconangle.comthe-decoder.comartificialanalysis.aitechcrunch.comthenewstack.iotheverge.comanthropic.comthenewstack.io
- Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 SolFor the first time in a while, no OpenAI model leads the MathArena AI math benchmark. Google's Gemini 4 Carbon and Anthropic's Fable 5.2 have knocked GPT-6.1 Sol from the top, though both cost...📰 News:the-decoder.comsiliconangle.comtheverge.comthenewstack.iothenewstack.iothe-decoder.comtechcrunch.comarstechnica.comarstechnica.comantigravity.googleblog.googlearstechnica.comsiliconangle.comthe-decoder.comsiliconangle.comengadget.comthe-decoder.comthe-decoder.comtheverge.comgithub.cominfoq.com
- Meta's Muse now calls businesses for you, but many companies might just hang up on the AIMeta's AI agent Muse can now call businesses on behalf of paying US subscribers, handling everything from cable providers to restaurants. The agent already has over 10 million users, according to...
- Google retools Kubernetes for AI agents instead of building a new stackAgentic workloads are pushing Kubernetes into territory its creators never planned for. Agents generate untrusted code on the fly, sit idle for long stretches and churn through pods at a pace that...
- ChatGPT Wallet lets OpenAI’s Dots agents make purchases for youAfter launching its Dots AI agents platform earlier this year, OpenAI is launching a wallet for the agents. ChatGPT Wallet is rolling out starting today for ChatGPT Pro subscribers in the US aged 18...
- TSMC employee data breach linked to cloud providerTaiwan Semiconductor Manufacturing Company (TSMC) confirmed that personal employee data has been leaked online after a threat actor put 6.9GB of information allegedly stolen from the company up for...
- Meta infrastructure chief Santosh Janardhan to step downMeta's longtime head of infrastructure and co-head of engineering, Santosh Janardhan, is leaving the company, Meta confirmed to TechCrunch. Dave Brown, who joined Meta as vice president of...
- OpenAI turns ChatGPT into a business directory with AI agents that handle customer questionsOpenAI is opening up ChatGPT as a business platform: through a new Business Hub, companies can claim their profile, add product data, and track how users interact with them. "Business Agents" let...
- Stolen Fireworks AI Key Behind Cursor Breach and Trojanized Claude Code ExtensionA stolen access key exposed in the Fireworks AI breach opened the door to the compromise of Anthropic's official Claude Code extension on Cursor's marketplace, Cursor said in a Friday update. "An...
- Goal EngineeringThe future of agents is them taking on open ended goals and driving towards completion of those goals. Learn more.🏢 Companies:LangChain
- AntiFox AI Pentesting Tool Tied to AWS Credential Theft in Confirmed IntrusionsSecurity researchers have linked an AI-driven penetration testing platform called AntiFox to confirmed malicious intrusions that stole Amazon Web Services credentials and moved laterally through...
- Apache Spark vs Apache Flink: which should you use?For most teams, Apache Spark™ is the better choice. Spark runs batch ETL, SQL analytics,...🏢 Companies:Databricks
- Production MLOps at GS Caltex with Amazon SageMaker Canvas, part 1GS Caltex connects Amazon SageMaker Canvas to Amazon SageMaker Pipelines so process engineers' no-code models run in production. This post shows how ML engineers reuse validated preprocessing flows...🏢 Companies:AWS AI/ML
- Route coding agent requests to the right model on Amazon Bedrock with NVIDIA NeMo SwitchyardCoding agents work in tasks, and the number that matters is cost per solved task, not price per token. In this post, you use the AWS CDK to deploy LiteLLM with the NVIDIA NeMo Switchyard routing...🏢 Companies:AWS AI/ML
- Daemons can now send traces to Vercel TracingVercel Daemons can now send OpenTelemetry traces to Vercel Tracing, allowing you and your agents to inspect background worker operations alongside the rest of your project's traces.Vercel provides...🏢 Companies:Vercel AI
- Architecture patterns for fast session startup on Amazon Bedrock AgentCore runtimeAmazon Bedrock AgentCore runtime V2 snapshots your initialized environment and restores it for each new session, so session start no longer grows with image size. This post explains how V2 starts a...🏢 Companies:AWS AI/ML
- OpenFold3 on Amazon SageMaker JumpStart: Predict biomolecular structuresDeploy OpenFold3-preview from Amazon SageMaker JumpStart to predict 3D structures of proteins, DNA, RNA, and ligand complexes in your own AWS account. The post covers asynchronous inference,...🏢 Companies:AWS AI/ML
- Snowflake a Leader in Gartner® Data Lakehouse Magic Quadrant™Snowflake is named a Leader in the inaugural Magic Quadrant for Data Lakehouse Platforms report by Gartner. Explore its approach to modernization, governance and AI.🏢 Companies:Snowflake
- Advancing science & engineering with Antigravity’s /teamworkEducation Innovation🏢 Companies:Google Research
- Cost-optimizing AssemblyAI speech-to-text inference on AWS with Amazon EC2 SpotLearn how AssemblyAI runs cost-optimized speech-to-text inference at scale on AWS, using Amazon ECS, Amazon EC2 Spot GPU instances, and an Amazon SQS-based retry architecture to achieve 70% cost...🏢 Companies:AWS AI/ML
- Token Pass now includes $150 in tokens for $30/monthVercel Token Pass now includes $150 of tokens per month (up from $80) for the same $30 per team per month.The credit applies to featured open-weight models served through AI Gateway, including Kimi...🏢 Companies:Vercel AI
- Build a Robot Pick-and-Place Simulation with SimReady Assets🏢 Companies:Hugging Face
- LangChain Raises $300M at $3B Valuation to Grow AI Agent Management PlatformLangChain, the startup behind AI agent creation and management software, has closed a $300 million Series C round at a $3 billion valuation, the company announced Friday. CapitalG led the round, with...
- XYBench: Helping with the problem behind the question🏢 Companies:Hugging Face