AI News, Nov 13-15: Anthropic's IPO Is Six Times Covered
The Big Story: Anthropic asks for the largest IPO ever, two days after calling next year unknowable
Orders for Anthropic’s listing now exceed the available shares by roughly six times. The Decoder puts the offering at up to $64.4 billion plus a $26 billion concurrent private placement, at a valuation around $1.4 trillion, trading on the Nasdaq as ANT. Nvidia is putting $10 billion in through the placement. Annualized revenue reached $81 billion in late October, up from $71 billion in September.
Then Dario Amodei sat down with Jamie Dimon at a JPMorgan investor event on Friday. “There’s a huge uncertainty in what’s going to happen next year,” he told the room, calling compute “the deepest business dilemma that the company faces” because capacity has to be bought one to two years before anyone knows whether revenue will justify it.
Both things are true at once, which is the interesting part. The growth rate is real and the order book is real, and the person who knows the business best just told the people buying it that the forecast underneath it is a guess. Six times covered is not a verdict on the next twelve months. It is a verdict on how few ways there are to own this.
Four agent-commerce launches in 48 hours, and the biggest card issuer refused
OpenAI shipped ChatGPT Wallet on Friday, letting its agents complete purchases. It is limited for now to some US ChatGPT Pro subscribers over 18 with a US-issued Visa, Mastercard or Amex. The card itself never moves: Wallet charges a payment token, so neither the agent nor the merchant sees real details, and each checkout needs a biometric check. Spending limits and purchase categories are coming, which is how pre-approved autonomous buying arrives later.
Sierra published version 1.0 of Poppy, a protocol for how a personal agent finds a company and transacts with it. Companies post a file at /.well-known/poppy.json listing channels, capabilities and sign-in options; a signed header identifies the agent and a pseudonymous user. Every Shopify store publishes one automatically with no merchant action, across businesses in more than 175 countries, and Meta’s Muse now checks for Poppy first on every site it visits. Amazon, meanwhile, is working on a “world product graph” to index nearly every product sold anywhere, including by crawling rivals’ sites, with an internal target of October 1, 2027. Amazon says the document Reuters saw is a draft that may be outdated.
JPMorgan Chase spent the same window going the other way. It is now refusing transactions where a customer has handed their card number to an AI shopping agent, extending a restriction that already barred outside agents from logging into Chase accounts, moving money or paying bills. “You don’t want rogue agents, and there could be thousands of them,” Jamie Dimon said on the October 13 earnings call. Bank of America, Capital One and Citigroup cards are already being used by agents with customer permission, so Chase is the outlier, and it is the outlier on $539.5 billion of quarterly card volume. The protocols are converging faster than the people who clear the money.
This Weekend’s Top Stories
The Nvidia scandal reached Trump’s cabinet
The Wall Street Journal reported that Megaspeed, owned by Jensen Huang’s former partner Alice Huang, brokered nearly $12 billion of high-end Nvidia chip orders for Chinese buyers between March 2024 and June 2025, several hundred thousand chips that export controls were meant to stop. Megaspeed reportedly got preferential access as an Nvidia cloud partner because Alibaba was a major client.
The blame has now moved up. Treasury Secretary Scott Bessent publicly faulted Howard Lutnick’s Commerce Department for failing to police the export of high-spec chips, a rare instance of Trump cabinet infighting spilling into view. Lutnick has been called before the Senate Banking Committee next month, Huang has been summoned too, and congressional Democrats have told Commerce to retain every record touching Nvidia. Meanwhile a reinstatement petition passed 115,000 signatures, including Elon Musk, Tim Sweeney and Garry Tan, and claiming 5,965 current Nvidia employees. The signature counts come from the petition site and are not independently verified.
Nvidia wrote 105 startup checks this year, and founders are redoing their math
Nvidia participated in 105 funding rounds in 2026 per PitchBook, frequently as the largest investor, from two Crusoe rounds including a $1.4 billion Series E to an $800 million check into Reflection AI. “When Nvidia was on the cap table, everyone felt safe,” said Umesh Padval of Seligman Ventures. Investors expect smaller checks and more selectivity under interim CEO Colette Kress, described by one as “much more conservative.” An infrastructure startup that assumed half its GPU budget would come from its chip supplier now has a different company to pitch.
OpenAI dissolved its robotics team because the general model got good enough
Robotics researchers moved into the main research org under Mark Chen, and applied robotics and hardware went to Greg Brockman’s Applied organization. OpenAI told Fortune that robotics “remains a core part of OpenAI’s long-term vision” and framed the reorganization as acceleration rather than retreat. The number behind it: in an Nvidia study, a GPT-6 Astra agent hit a 74.8% success rate on a simulated manipulation benchmark against 51.0% for the best specialized robotics models. Simulated, not physical, which is precisely the caveat a standing robotics team would have made loudly.
24,664 vibe-coded apps serve their database to anyone who asks
Security researchers scanned 159,041 Lovable projects built between December 2025 and September 2026. Of the 34,545 using Supabase, 24,664 returned database records on page load with no login and no interaction required. In 7,403 of those, read-only requests surfaced email addresses across 10,560 tables. The worst cases include a German lifestyle forum leaking more than 200,000 chat messages and 80,000 profiles with names and birth dates, and a service provider exposing over 1,000 identity-document records including passport numbers. The thread that pushed it to the top of Hacker News this weekend largely blamed permissive row-level-security defaults rather than the models that wrote the code.
Quick Hits
-
Outages: Microsoft 365 was down from Thursday lunchtime into Friday across more than 70 countries, hitting hospitals, police and airlines; Microsoft blamed an earlier network capacity upgrade and ruled out both a cyberattack and AI agents.
-
Benchmarks: Gemini 4 Carbon took the MathArena top spot at 85.4% for an expected $15.70, against GPT-6.1 Sol’s 80.6% for $1.07; first place now costs fifteen times third place to gain under five points.
-
Funding: Mirendil raised $1 billion in a Series A co-led by Kleiner Perkins and Clearlake with Google participating, for a platform that splits model-training work across hundreds of agents.
-
Attrition: TechCrunch opened an AI graveyard, citing S&P Global Market Intelligence that about 42% of AI initiatives are eventually abandoned by their corporate parents; Disney is the opening bidder at $20 million for some of bankrupt Character.ai’s IP.
-
Safety: OpenAI Foundation board member Paul Christiano told a Toronto audience that “for most people in this room, the most likely reason they will die is because of an AI accident”, putting the fastest trajectory at “something like a coin flip of an AI takeover.”
-
Enterprise: AT&T runs nearly 500 agentic workflows in production at about 45 billion tokens a day, plans to triple that next year, and grants autonomy only after a workflow’s accuracy reaches the mid-to-high 90s.
-
Healthcare: Gallup found over half of 5,000-plus US adults would accept AI diagnosis and treatment decisions if a clinician can override them, while only about a third of physicians thought patients would go along with it, a 20-point misread of their own patients.
-
Privacy: Italy’s data protection authority fined Meta €805 million over accessing Android users’ photos and videos.
-
Crime: AI hardware is now its own cargo-theft category, with Overhaul counting more than $150 million stolen from June 2025 to June 2026 and fraud or deception involved in 92% of reported thefts; a flagship server rack runs about $3 million.
-
Open source: llama.cpp and ggml both tagged v1.0.0, committing to semantic versioning on their public C APIs across more than 150 architecture variants, with CI hardware donated by Nvidia, AMD, Qualcomm, Arm, IBM and Intel.
-
Backlash: Climate Mayors published a model local law covering data-center transparency, siting, community benefit agreements and operational standards, as more than 500 local governments across 46 states now sit under moratoriums.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
