November 15, 2026

Jensen Huang quits Nvidia; its startups expect smaller cheques

The price of the AI you use this year was set by bets placed a year or two ago, and this week the people placing them changed. Jensen Huang, who ran the AI-chip maker Nvidia, resigned on Thursday, and investors expect his stand-in to write smaller cheques to the cloud startups and model labs Nvidia funded. Anthropic, days from selling its shares on the stock market for the first time, called buying computing power its deepest dilemma, because it has to be ordered one to two years ahead. AI assistants also learned to spend: ChatGPT and Replit now pay for things once you approve the total, which makes that one tap the main safety check, in a week when two publish-and-forget defaults left passwords and patient records open. Nvidia's results on Tuesday will show whether the cheques shrink.

The Big Story

Jensen Huang resigns from Nvidia, and the startups it funded brace for smaller cheques

What happened

Nvidia's board announced on November 12 that Jensen Huang had resigned as president, chief executive and board member, effective immediately, after 33 years running the company he co-founded. Colette Kress, the finance chief since 2013, is interim chief executive and keeps the finance job; Harvey Jones becomes the board's first-ever chair. The exit followed reports of Huang's affair with Alice Huang, an executive at Megaspeed, a reseller suspected of passing Nvidia chips into China; even inside the US government, according to CNN, the evidence that she worked for Beijing is disputed. Separately, researchers reported traces of export-banned Nvidia chips in models from DeepSeek, Moonshot, MiniMax and Ant Group. The Wall Street Journal points to a second cause: under Huang, Nvidia promised to stand behind AI cloud companies and data-centre builders so they could afford its chips, in exchange for a cut of their revenue. The biggest such promise under discussion was a $250 billion guarantee for an OpenAI data centre in Ohio; Kress, then finance chief, cut it in half, and the smaller deal was signed in August. According to PitchBook, Nvidia took part in 105 startup funding rounds this year, often as the largest investor. Nvidia reports quarterly results on November 17.

Why it matters

Nvidia's guarantees let AI clouds buy chips they could not otherwise afford, and its money backed the companies that rent out AI chips, the services that run open models for you, and labs such as Mistral and Black Forest Labs, and investors expect Kress to be more conservative. If your app rents chips from one of those clouds or runs on a model one of those labs gives away, your supplier is not failing, but the investor who made the others comfortable may step back; this is a good month to know what moving to a second provider would take. Nothing changes yet in the AI apps most people use, though their cheap plans are priced off that same chip-rental market. Rivals such as Google, which announced two new chips this week, could step in, though investors doubt Big Tech will fund companies with no revenue the way Nvidia did.

Read the original →

Who Pays for the Computers

  • Anthropic sets its share price at $70 to $78, listing on Nasdaq as ANT

    Anthropic began pitching its first share sale to investors, offering 825 million shares on Nasdaq. Bloomberg reports it is seeking up to $64.4 billion from the public sale plus $26 billion sold directly to chosen investors, which together would make it the largest listing ever. Its chief executive, Dario Amodei, told a JPMorgan investor event that computing has to be bought one to two years before revenue shows whether it was needed; Bloomberg counts nearly $901 billion in commitments, about $71 billion of them due by September 2027. The listing money is his cushion, and it decides whether next year's Claude capacity is ordered generously or rationed.

    Read more →
  • Anthropic buys $3.4 billion of gas-turbine generators to power its data centres

    Anthropic is buying Caterpillar generators able to produce up to 1.7 gigawatts, the Financial Times reports. The company has preferred to rent computing rather than own buildings, but the queue for new turbines now runs into the early 2030s, and the wait to plug a new site into the power grid can stretch just as far. That puts the power order ahead of the chip order in Anthropic's queue.

    Read more →
  • OpenAI's quarterly revenue reaches $10 billion while its loss widens to $10.9 billion

    OpenAI grew revenue about 50% from the previous quarter, according to the Wall Street Journal and The Information, while its adjusted operating loss rose from $10.1 billion to $10.9 billion. Anthropic brought in $17 billion over the same period with an operating loss of $567 million. To catch up with Anthropic among business customers, OpenAI is cutting its prices, so part of a cheaper OpenAI bill is being paid for by a company losing money every quarter.

    Read more →

Agents at the Checkout

  • ChatGPT Wallet lets an assistant pay without ever seeing your card number

    OpenAI launched ChatGPT Wallet with American Express, Mastercard and Visa, starting with Pro subscribers in the US. You add a card once, then approve each purchase with Face ID or a fingerprint after reviewing the items and the total; the card network then issues a stand-in card number for that payment, so neither the assistant nor the shop sees the real one. Your card's usual refund protections stay in place. To accept these payments, OpenAI says, shops must stop accidentally blocking agents, the AI assistants that click and buy the way a person would, and teach their fraud checks to recognise purchases made by one.

    Read more →
  • Replit Agent buys supplies for your business and runs your Google Ads

    Ask Replit's agent for "50 shipping boxes that fit this product, under $75 including delivery" and it finds options, fills in the shop's checkout through a browser, and waits for you to approve the total before anything is charged, using Stripe's Link wallet and Visa. A one-person business that built its app on Replit can hand the ordering and the ad buying to the same agent, the two jobs that usually come next.

    Read more →
  • United Airlines moves to keep AI agents off its booking and check-in systems

    United said it will restrict AI assistants' access to its apps and site, and through travel agencies, citing the risk of unintended purchases, account changes and misused frequent-flyer miles. It also flagged check-in, which asks a human to confirm that they carry no banned items such as explosives. Expedia and Booking, by contrast, are working with OpenAI, and Walmart partnered with Meta's assistant while Amazon blocked it. That splits travel in two: agents work where a company signed a deal with an AI lab, and fail where they have to drive the website themselves, the method Replit's agent relies on.

    Read more →
  • Two OpenAI engineers propose letting agents tell websites your country and time zone

    A draft shared with the web standards body that writes the rules for HTTP would let an agent attach six optional details to each request — your network, country, a masked IP address, language, time zone and browser. Sites block agent traffic because thousands of users arrive from the same few data-centre addresses, and these details would make the agent look like an ordinary visitor again. The critic Joost de Valk objects that today an agent tells a site almost nothing about you, and this would change that. The company running the agent, not you, would choose what to share, and because a verifiable signature is optional, anyone could fake the details.

    Read more →

Public by Default

  • Grok's app builder left 94.7% of the apps still online open with no sign-in

    The security firm Claroc tried 43,551 addresses of apps people had built by describing them to Grok; 37,571 were still online and 35,578 opened with no sign-in. The publish button suggested "Anyone with the link", and each app's address lands in public records of web addresses that can be searched, so strangers found them without being sent anything. Among them were an ICU patient's lab results, a timesheet listing employees' passwords, and 24 apps from the staff of SpaceXSI, the company that makes Grok. Nobody reviewed these apps before they went live; the only check was a publish button with the wrong default.

    Read more →
  • Claude Code's "don't ask again" saved passwords into files that ended up on GitHub

    When you approve a command in Claude Code and choose "Yes, and don't ask again", it saves the full command text to a settings file in your project, including any password typed into it. Apiiro, a security firm, found 104,185 public copies of that file on GitHub, 1,122 of them holding a real secret. The file is meant to stay out of version control, but that protection can lapse when a project is cloned elsewhere or a folder becomes a repository later. If you use Claude Code, check whether .claude/settings.local.json was ever committed, and change any key you find in it.

    Read more →
  • Cursor traces its breach to a key leaked by a partner, then a fake extension

    Cursor, the AI coding app, had given the model provider Fireworks a key to some of its data, and Fireworks exposed it. An attacker used the key to copy coding-session records Cursor had shared for training, and a few still held passwords nobody had removed. From there the attacker reached Cursor's internal tools and placed a malicious copy of Anthropic's Claude Code extension in Cursor's marketplace for about 11 hours, where at most 5,000 users may have downloaded it.

    Read more →

Bigger Jobs, Smaller Machines

  • Google's Gemini 4 Hydrogen matches its previous top tier at $1.50 per million tokens

    Google released Gemini 4 Hydrogen at $1.50 per million tokens of input and $7.50 per million of output. A token is roughly three-quarters of a word, and the price holds across the model's whole million-token memory. It matches Gemini 4 Argon, one of Google's two top models, on an independent ranking, and finishes tasks 2.5 times faster than the model it replaces. In Google's coding tool, a new mode lets the bigger Carbon model plan while Hydrogen writes and tests the code, coming close to Carbon working alone on 24% less of your usage allowance.

    Read more →
  • Claude Fable 5.2 completes 41.3% of real freelance jobs to a client's standard

    On the Remote Labor Index, a test built from 240 completed Upwork projects, Anthropic's new model delivered work a paying client would accept 41.3% of the time; a year ago the best system managed 2.5%. The next best are Opus 5.5 at 21.25% and GPT-6 Astra at 20.83%. These are whole jobs that took a human freelancer just under 29 hours on average, and failures usually come back amateur or broken. For someone who commissions that kind of work, the job shifts from doing it to judging it, with roughly a two-in-five chance the result is usable.

    Read more →
  • Edge0 releases a 145-billion-parameter open model that runs on a laptop

    Edge0 published Edge0-145B and a smaller 120B model, free to download. Parameters are the internal settings a model learns, and a model this size usually needs a server; Edge0 reports under 30GB of peak memory and more than 80 tokens a second on a compatible laptop. Running locally means no per-use fee and files that never leave your machine, though the comparisons with cloud models come from Edge0 itself.

    Read more →

Tools & Launches

  • Nod by n8n▲ 304

    A personal assistant from the makers of the automation tool n8n that works inside a sealed-off computer of its own on your Mac or PC, so it can use your apps and ordinary websites without touching your desktop. It takes orders only from you, asks before doing anything you did not request, and runs on whichever AI provider you choose. It is free and open source. If you want something to book and file things but would rather not hand a cloud service your logins, start here.

    Visit site →
  • CALL-E▲ 260

    Lets ChatGPT, Claude or Muse make phone calls to US and Canadian numbers, to book a table or move an appointment. It asks you for anything missing and dials only once you have confirmed the plan, then gets through phone menus, hold music and voicemail and brings back a summary and the transcript. Meta's Muse added its own calling for paying subscribers this week. New users get 100 free credits. It is for anyone who keeps putting off the calls where a phone menu turns two minutes into twenty.

    Visit site →
  • Plurai Simulation▲ 418

    Builds a pretend copy of your agent's world from one command in Claude Code, Codex or GitHub Copilot: invented customers and documents, plus fake versions of the tools, APIs and databases it calls. The agent then runs in a sealed test area that never touches your live systems, so its worst mistakes show up before a user meets them. It is for builders who currently test an agent by watching it work on live accounts and hoping.

    Visit site →
  • Omnigraph▲ 294

    An open-source database where several AI agents share one store of facts, with every change made on a separate branch that can be reviewed and merged, the way code is in Git. That means one agent's wrong note can be checked before the others start relying on it. It runs on your own cloud file storage and searches by meaning and keyword, and follows links between facts, in one query. Teams whose agents share memory through a pile of unchecked text files get a review step before anything is merged.

    Visit site →

From the Blog

On Women of the Island: Minimalism and Making Things in the Age of AI

A conversation from a Chinese-language podcast about women living in New Zealand, on owning less and on whether it is still worth making things yourself when AI can write the code and edit the video.

Read on chanmeng.org →

In Brief

On Tuesday I will be listening to Nvidia's results call for whether anyone says what happens to its startup investments. If you rent AI chips from a smaller cloud or build on a downloadable model, reply and tell me which, so I know who to follow up on.

Keep building — Chan


← All issues