Saturday's verdict is that November 13 was the day the defaults broke. Jensen Huang resigned as Nvidia's chief executive after 33 years, effective immediately, with chief financial officer Colette Kress named interim CEO while the board investigates reports of affairs, one involving a woman US officials suspect of feeding information to Beijing—an allegation that remains unresolved, with Nvidia itself saying on November 8 that it had no reason to believe company secrets or national security were compromised. Microsoft 365 spent a second day down: an Exchange Online failure that began Thursday afternoon took 19 services with it, drew about 140,000 Outlook problem reports from 74 countries, and reached the US House of Representatives. OpenAI launched ChatGPT Wallet, letting its dots agents complete purchases with a token instead of a card number and a fingerprint or Face ID check on every buy, starting with US ChatGPT Pro subscribers. Anthropic's $64.4 billion IPO was reported six times oversubscribed on the roadshow's third day, ahead of pricing on November 18 and a Nasdaq debut November 19 under the ticker ANT. Chicago Public Schools said it will look to other search solutions if it cannot suppress Google's AI Overviews without losing Search itself. And Google DeepMind opened the two datasets behind WeatherNext 3—SatWeave's hourly mosaic of five geostationary satellites and PARDIG's global rainfall estimates every 15 minutes—to anyone who wants them.
Nvidia: Jensen Huang resigns after 33 years, Kress takes the interim chair
Nvidia announced Thursday, November 12, that Jensen Huang had resigned as president, chief executive, and board member, effective immediately; reporting on the company's SEC filing puts the board's acceptance on November 11. Huang co-founded Nvidia in 1993 and led it for 33 years into the company that sells the picks and shovels of the entire AI buildout. Chief financial officer Colette Kress, in that job since 2013, becomes interim CEO while the board searches inside and outside for a permanent successor, and Harvey Jones becomes the board's first chair. The exit follows a board investigation opened November 8 into reported extramarital affairs. Huang acknowledged infidelity in a November 8 message to employees and apologized; he did not identify any partner or admit knowledge of any intelligence operation. One reported relationship involved a former executive at a customer firm whom US intelligence officials suspect of acting for Beijing—suspect is the operative word, and nothing in the public record establishes espionage. House committee chairs have requested a closed-door transcribed interview on December 1 and asked for acceptance by 8 PM Eastern on November 13; acceptance has not been confirmed. Nvidia reports third-quarter fiscal 2027 earnings on November 17, and one market account put the week's loss in Nvidia's market value between $573 billion and $657 billion.
Esther Dyson's chair has sat through founder successions before, and her question is the only one that matters: was Nvidia a man or a system. The honest answer on Saturday morning is that nobody outside the building knows, and the stock's week says the market priced it as a man. Huang was not a hired manager. He was the strategy—China access, the circular financing web that pours Nvidia cash back into cloud providers and labs, the personal line to Washington—and a strategy that lives in one person's relationships does not transfer to an interim CFO by press release. The board's discipline deserves its due in the same breath: an investigation opened November 8, a resignation accepted within days, a named interim, a first-ever chair. That is governance moving at the speed the facts arrived. What the paper will not do is convict anyone of espionage on a suspicion. The allegation is reported, the investigation is real, the resignation is fact, and the rest is open—and on this page open stays written as open.
OpenAI: ChatGPT Wallet gives the agent a way to pay, not a look at the card
OpenAI launched ChatGPT Wallet on Friday, November 13, built with American Express, Mastercard, and Visa. It works through dots, OpenAI's personal agents, and it is rolling out gradually to ChatGPT Pro subscribers aged 18 and older in the United States with eligible US-issued cards. The user adds a card once. For each purchase the agent presents the items, the total, and the selected card, and the user approves with a security check such as Face ID or a fingerprint scan. Only then does the payment network issue a token—a substitute credential with its own number, security code, and expiry—so neither the merchant nor the agent ever sees the real card details. Card rewards and dispute protections carry over, the merchant remains responsible for pricing, shipping, and returns, and the wallet is not a bank account and holds no funds. OpenAI published a Commerce Partner Playbook for merchants, naming fraud partners Forter, Riskified, Sardine, and Signifyd and payment providers Adyen, Checkout.com, and Stripe, and it concedes some sites currently block agent purchases unintentionally. Delegated buying—an agent purchasing inside limits, categories, or time windows set in advance—is a stated expectation, not a shipped feature.
For Chris' client work this is the home galaxy again: the answer just learned to check out. Last week this paper covered Businesses in ChatGPT, the claimable profiles and product feeds. Wallet is the other half of that storefront. A business that is cited but cannot be bought from inside the agent's flow is a brochure; a business whose product data, agent access, and risk signals are clean enough to survive Forter and Stripe is a store. The design choice the practitioner chair approves is the approval itself: the agent proposes, the human disposes, per purchase, with a biometric signature. That is the same gate OpenAI put on its security fixes yesterday, and it is becoming the industry pattern because it is the only pattern a card network will underwrite. Watch the delegated step. The day pre-approved limits ship, the SEO question changes from who ranks to whose standing instructions the agent is spending under.
Microsoft: one mail failure, 19 services, a second day dark
A Microsoft 365 outage that began before 11:17 AM Eastern on Thursday, November 12, ran through Friday, November 13, its second day. Microsoft's public Service Health Status page, updated 5:46 AM UTC Friday, listed 19 services affected through their dependence on Exchange Online, under incident MO1550821 with Exchange tracked separately as EX1550543. Outlook failed on web, desktop, and mobile; mail delivery delayed; Teams lost its calendar, scheduling, and meeting edits while calls and chats largely kept working; SharePoint features, Microsoft 365 Copilot, Graph, Planner, To Do, Intune, Purview, and Power BI were all touched, and Copilot could not reach the mail, calendar, and files it answers from. Downdetector logged roughly 140,000 Outlook reports from 74 countries at the peak, and staff in the US House of Representatives were among those disrupted. Microsoft said there was no indication of malicious or cyber activity, began reversing a recent network configuration change, promised a full Post-Incident Report after mitigation, and by late Friday reported sustained recovery across previously affected environments without declaring the incident fixed. Microsoft also extended the email timeout window to 48 hours.
Lessig's chair reads an outage as an architecture confession, and this one confesses plenty. Nineteen services did not fail. One service failed nineteen ways, because the modern office suite is a single dependency wearing nineteen costumes—and the AI layer sits on top of the mail layer, so when Exchange went dark, Copilot went blind in the same hour. Every business that moved its records, its calendar, and now its AI assistant into one tenant learned the price of the bundle while it could not read its email. The board's optimism holds anyway, because the counterfactual is a hundred separate vendors failing one at a time all year. But Friday's exhibit belongs in every continuity plan: the question is no longer whether your cloud goes down, it is what still works at your desk on the morning it does. Print the meeting links. Keep a second way to reach your people. The post-incident report, when it lands, will be the real story.
Anthropic: the IPO book fills six times over
Reuters reported Friday, November 13, on Bloomberg News reporting, that Anthropic's initial public offering was about six times oversubscribed by the third day of its roadshow—twice subscribed after day one, six times by Thursday—with bankers telling prospective investors that existing backers will be prioritized. The offering is 825 million Class A shares at $70 to $78, raising up to $64.4 billion, alongside a $26 billion private placement that takes the combined raise to $90.4 billion and the fully diluted valuation to roughly $1.4 trillion. Up to 20% of the shares are reserved for retail investors, against a typical 5 to 10% in large US offerings. Filings put Anthropic's annualized revenue run rate at $81 billion as of October 31, up from $71 billion at the end of September and $65 billion in July. Pricing is expected November 18, with trading to begin November 19 on the Nasdaq Global Select Market under ANT. Demand figures are roadshow reporting, not filings, and an order book can move until pricing day.
Six times is a headline; the run rate is the story. A year of AI finance ran on private rounds at whispered valuations—this paper covered Anthropic's draft terms on Wednesday at the same $70 to $78—and the public book is the first time strangers with their own money got to vote in size. They voted six deep. The skeptic's seat stays occupied: oversubscription is a snapshot, the people familiar with the offering themselves cautioned that details could change, and the questions investors pressed on the roadshow—can Claude hold its edge as Google narrows the model gap—are the right questions and remain unanswered by a full order book. But note what the week now holds in sequence. On Thursday the AI industry's hardware king lost his crown in a scandal. On Friday its newest public company-in-waiting could not print enough shares. The boom did not pause for the crisis. Whether that is resilience or momentum is the argument the board will still be having at pricing on Wednesday.
Chicago: the schools would rather lose Search than take the AI with it
Chicago Public Schools told reporters this week that it is actively exploring all options to independently suppress the AI Overviews at the top of Google searches, and that if it cannot give students Google Search without the AI feature, it will look to other search solutions. The district has already blocked generative AI tools, Gemini among them, on student devices and its network, citing student privacy and safety and the federal and Illinois privacy laws that govern school data. The standoff began in early November, when Google vice president Chris Phillips said K-12 schools would soon have to accept Google's AI services to keep student access to Search, because Search and its AI features are now too closely integrated to offer separately. About 30 parents, educators, and students rallied outside Google's Chicago office in the West Loop on November 11, with coordinated protests in Cambridge, New York, and Los Angeles, and Senator Ed Markey's November 9 letter asks Google to answer by November 30 whether schools can disable AI Overviews, Gemini, and Guided Learning while keeping Search. Google says it is piloting Guided Learning, which walks a student through a problem instead of handing over an answer, and has stopped switching Gemini on automatically in Classroom.
Gigi Sohn's chair asks who the network serves, and for once the answer is being negotiated in public by the served. A school district is not a consumer choosing a browser. It is a public institution with legal duties to children, being told the price of the reference desk is the slot machine bolted to its top. Google's engineering claim may even be true—AI Overviews are not a plugin anymore, they are Search's front end—and that is exactly Lessig's point arriving from the other direction: when the architecture makes a feature inseparable, the architecture is the policy, and no settings page can fix it. DuckDuckGo's chief executive spent the week publicly inviting schools to switch, which tells you the opening is real. If Chicago walks, it will not be the largest customer Google ever lost. It will be the first one that left on principle, in daylight, over the answer box—and every district in America is watching whether the sky falls.
DeepMind: the weather data underneath the model, opened to everyone
Google DeepMind released PARDIG and SatWeave on Friday, November 13, the two foundational datasets behind its WeatherNext 3 forecasting model, openly on Google Cloud Storage, with PARDIG's historical catalogs also on Google Earth Engine. SatWeave stitches imagery from five geostationary satellites—GOES-East and GOES-West, Meteosat at 0 degrees and over the Indian Ocean, and Himawari—into one global mosaic across 17 visible, water vapor, and infrared channels, updated hourly, with a historical archive spanning more than a decade; that hourly refresh is what lets WeatherNext 3 start a new global forecast every hour instead of waiting on six-hour analysis cycles. PARDIG—Precipitation AI Reanalysis: Densely Inferred GPM—is trained on sparse spaceborne radar from NASA's Global Precipitation Measurement Core Observatory and estimates global rainfall every 15 minutes at roughly 10-kilometer resolution, with archives from 1999 to 2026. DeepMind reports PARDIG improves hourly detection skill for light-to-moderate precipitation by more than 30% against NASA's IMERG-Final; the underlying paper is a November 10 arXiv preprint, not yet peer reviewed, and PARDIG's live feed arrives with an average delay of about five hours, against about one hour for SatWeave. DeepMind is explicit that PARDIG is experimental research data, and official meteorological warnings remain the authority for decisions about life and property.
This is the source doctrine with its sleeves rolled up: all the cool shit comes from academia, and the second-coolest thing is a lab handing the rest of the world its instruments. A model announcement gives you a capability for a news cycle. A dataset release gives every hydrology group, flood modeler, insurer, and graduate student on earth the same raw material the frontier lab used, in the places—most of the planet—where ground radar barely exists and rainfall is guessed. The board prints the limits in the same paragraph, by house rule: 30% is DeepMind's own figure, the paper is a preprint, and a five-hour delay makes PARDIG a training and analysis instrument, not a siren. None of that shrinks the act. On a day when the industry's news was a resignation, an outage, and an order book, the most durable item may be the quiet one: two archives, opened, dated 1999 to 2026, that will still be training somebody's model when this week's tickers are trivia.
What the day adds up to
Saturday adds up to a stress test that every default failed at once, and the system kept running anyway. The founder default broke at Nvidia: 33 years ended in a week, on allegations this paper reports as allegations. The infrastructure default broke at Microsoft: the office itself went dark for a second day because nineteen products are one product. The payment default moved at OpenAI: the agent can now spend, inside a biometric gate. The private-market default broke at Anthropic: the public book filled six times over. The compulsory default cracked in Chicago: a school district said the search engine is replaceable and the children's privacy is not. And underneath the noise, DeepMind widened the commons, opening the data rather than just the demo. The through-line for a working reader is succession—of a founder, of a feed, of a card, of a dataset. November 13 rewarded anyone who had already asked what happens the day the person, the platform, or the default is gone.
Yesterday's roundup, AI Roundup: The Perimeter Moved, asked where each system runs and whose keys it holds; today the same question arrived as a resignation, an outage, and a wallet.
Sources: Reuters via BusinessWorld and The Hindu BusinessLine, and NPR reporting, on Jensen Huang's resignation, Colette Kress as interim CEO, the board investigation, and the congressional interview request, November 12–13, 2026, allegations unresolved in the public record; Reuters and PYMNTS on OpenAI's ChatGPT Wallet launch, November 13, 2026, terms and partner lists company-stated; PEOPLE, the BBC via Downdetector, and Microsoft's Service Health Status page on the Microsoft 365 outage, incidents MO1550821 and EX1550543, November 12–13, 2026; Reuters and Bloomberg News via Unite.AI on Anthropic's IPO oversubscription, November 13, 2026, offering terms from the company's filings; Chalkbeat Chicago via TechSpot and Cybernews on Chicago Public Schools and Google's K-12 Search and AI policy, November 11–13, 2026; Google DeepMind on the PARDIG and SatWeave release, November 13, 2026, performance figure company-reported and the underlying paper a preprint.