Morning Brief, November 15th 2026
Days before Anthropic prices its , Dario Amodei calls buying compute the company's deepest business dilemma. A stolen key let attackers slip a Claude Code extension to as many as 5,000 Cursor users. Megaspeed says Alice Huang left in 2025. And a judge lets a memory price-fixing suit go ahead.
Amodei calls compute Anthropic's "deepest business dilemma," with $71 billion of commitments due by September 2027
Speaking with Jamie Dimon at a JPMorgan investor event in New York on Friday, Dario Amodei said compute "is the deepest business dilemma that the company faces" and that "there's a huge uncertainty in what's going to happen next year." Capacity has to be bought one to two years ahead of the revenue it is meant to serve. Buy too little and Anthropic turns customers away; buy too much and "you can financially overextend yourself." His answer, as The Decoder summarizes it, is to commit to relatively little upfront and buy on the when demand runs ahead of plan, which he likened to insurance. He described the internal model as "under-promise and over-deliver."
Bloomberg puts Anthropic's commitments at nearly $901 billion, mostly compute and chips, with about $71 billion due by September 2027, more than the market value of roughly two-thirds of S&P 500 companies. Bloomberg also reports Anthropic plans to raise up to $64.4 billion in the IPO plus $26 billion in a . The Decoder cites the filings for an annualized revenue of $81 billion at the end of October. Pricing is expected Wednesday. SourcesBC
A Fireworks AI key let attackers push a backdoored Claude Code extension to as many as 5,000 Cursor users
Cursor says attackers took an access key stored at Fireworks AI, its former provider, and used it to copy a storage bucket the two companies shared. The bucket held coding-agent traces, mostly from 2025, from users without Privacy Mode, and some traces contained unredacted that opened further Cursor systems. On October 29th the attackers used a stolen credential to plant two malicious copies of Anthropic's Claude Code extension, versions 2.1.312 and 2.1.313, in the cache behind Cursor's extension marketplace. They were live for about 11 hours and tried to install a persistent SSH backdoor. Cursor estimates up to 5,000 users requested those versions; coding errors probably stopped the backdoor on many Windows and Mac machines where SSH was off.
Cursor says fewer than 0.05% of users had traces exposed and fewer than 0.01% of customers lost API credentials. It has revoked the key, rotated credentials, rebuilt the marketplace and moved all inference off Fireworks. Anthropic's own extension and the were not affected. Notion's chat transcripts were exposed in the same breaches; this is the first case where the stolen data was used to ship code onto developers' machines. SourcesAAB
Megaspeed says Alice Huang left the company in August 2025 and hires outside lawyers to investigate itself
In a statement emailed to CNBC on Sunday, the Singapore cloud company at the center of the Nvidia affair said "Ms. Alice Huang officially left Megaspeed effective August 30, 2025 and currently has no role with the Company." It called media reports about its business, and operations "not supported by verifiable evidence," said it has retained outside counsel for a "thorough internal investigation," and said its controls are designed to comply with US export rules. It declined to comment on "Ms. Huang's personal and private matters."
The date matters. Singapore's police impounded her passport at Megaspeed in June 2025 and returned it in February 2026, so the company's account has her leaving while that investigation was open. Megaspeed calls itself a "core business partner" of Nvidia's regional cloud business. SourcesB
The White House AI task force asks regulators to waive penalties for 12 months for companies that report cyber incidents
The administration's AI task force, led by Director of National Intelligence Jay Clayton, met on Thursday and recommended that federal agencies hold off regulatory penalties for at least 12 months on companies that report cybersecurity vulnerabilities and incidents voluntarily and in good faith, with the window extendable. Officials said current enforcement standards "may be outdated" and have "a chilling effect on information sharing," and that organizations "that only a short time ago had reasonable protections in place are now more vulnerable" because attackers use AI. The group also backed long-term reauthorization of the 2015 .
Commerce Secretary Howard Lutnick and National Cyber Director Sean Cairncross attended; Cairncross and Pentagon research chief Emil Michael described talks with industry on a voluntary process for reporting AI incidents. It is a recommendation, not an order. No regulator has said it will apply it, and it covers cyber incidents only. SourcesBC
A federal judge lets a price-fixing suit against Samsung, SK Hynix and Micron proceed over their shift to AI memory
US District Judge Noël Wise in San Jose denied the three memory makers' in full, in an order dated November 14th, so a conspiracy claim brought by 17 consumers and small businesses goes forward. The plaintiffs allege the companies cut supply of conventional and moved capacity to high-bandwidth memory for AI chips, while refusing long-term DRAM contracts. Per the complaint, the three held over 91% of DRAM revenue in the third quarter of 2025, and conventional DRAM prices rose 171.8% in a year and 697% over 18 months.
The judge held the case is not a rerun of the earlier DRAM litigation, citing long-term lock-ups, Micron's closure of its Crucial consumer brand, delays and the companies' cartel history: Samsung paid a $300 million fine and Hynix $185 million over a 1998 to 2002 conspiracy. The defendants had invoked Judge Posner's line that competitors "watch each other like hawks"; the order answers that these allegations look more like "hawks flying in formation." The allegations are unproven. Answers are due December 7th. If "booming AI demand" stops working as an innocent explanation for memory prices, every buyer of a phone or laptop has a claim in the AI buildout's costs. SourcesAAC
The FAA still cannot say what stopped every US departure, and points at a router on an L3Harris network
At a press conference on Saturday, FAA Administrator Bryan Bedford said the agency halted flights on Friday because "we couldn't explain the instability" and would not risk "a sensor losing connectivity with aircraft." As of Saturday afternoon no root cause had been established, though routers and networking gear appeared to be involved. Transportation Secretary Sean Duffy said, "Until I can build [the new architecture], I'm stuck with L3Harris, who owns this network," and has asked Congress for $50 billion.
Bedford called router swaps routine and separate from the $12.5 billion modernization: "We are constantly replacing 2,000, you know, Cisco 600 routers with Juniper routers." L3Harris said it is supporting the investigation and is more than 70% through moving airspace communications from copper to fiber. CNBC's account of the briefing does not mention SMART, the AI routing tool that went national three days before the stop, so that question is unanswered rather than closed. SourcesB
Korea says Nvidia has reaffirmed 260,000 GPUs, as Asian governments respond to Huang's exit
South Korea's Science Ministry said Nvidia reaffirmed its plan to supply 260,000 advanced and to continue joint AI projects, after the government's AI strategy council held an emergency meeting on Friday. Taiwan's Economy Minister Kung Ming-shin said the departure is unlikely to change the global industry's growth much, and Taipei's city government said Nvidia's planned overseas headquarters is on track. Japan's Industry Ministry said it would "closely monitor" the company, and Deputy Chief Cabinet Secretary Masanao Ozaki said it was "too early to assess."
Singapore added that police issued Alice Huang a warning in February 2026 and barred her from re-entering. Reuters reports Huang's exit has raised uncertainty over more than $3 billion of Chinese AI computing projects that rely on restricted Nvidia processors. SourcesB
China's Global Times tells Korean chipmakers to sell their best memory to China
The Global Times, a Chinese Communist Party outlet, wrote on Friday that "to maintain and expand their presence in China, South Korean chipmakers must offer their most advanced products," and that "finding a way to overcome U.S. sanctions blocking advanced HBM exports is essential." It urged Seoul to press Washington to allow the sales, and said the Huang affair "is adding weight to calls for tighter restrictions on China's access to advanced chips." No Korean government response was reported.
SK Hynix fell 4.3% in New York on Friday as the Philadelphia Semiconductor Index closed down 3.43% after a fall of more than 6% intraday. Trump meets Xi in Shenzhen on November 18th and 19th, and Beijing is setting out its ask on memory before the room fills up. SourcesB
Claude reads Kurt Gödel's shorthand notebooks in days and catalogs 663 theorems, about 50 of them false
Jan von Plato, a University of Helsinki logician, has worked through Gödel's 1,009 pages of Gabelsberger shorthand since 2017 at about a page a week. He started using Claude in September, and The Guardian reports that within days it produced a 650-page explanation of the workbooks and a catalog of 663 theorems. "In a couple of days, I got to see a 650-page explanation of the workbooks," von Plato said. Checked in the , about 400 held as stated, about 200 needed sharper statements or extra assumptions, and about 50 were false.
The headline result is a near-complete reconstruction of Gödel's proof that the is independent, which he claimed in 1966 and never published; Paul Cohen won a for that result by other means. Joel David Hamkins of Notre Dame said separating "how much is Claude and how much is Gödel is a job for the historians." The figures here come from a summary of the Guardian report, which could not be opened directly. SourcesBC
A company formed on Friday bids $10 million for Character.AI's users' chats, characters and models
In Character Technologies' case in Delaware, a notice filed Saturday names Smirks and Pangs Acquisition LLC as for everything outside Disney's $20 million license to code and nine patents. The bid is $10 million in cash plus assumed liabilities, about $2.68 million of them , $2.65 million owed to DigitalOcean. The assets include user account data, user-created characters, chat content, training datasets, consent and age-assurance records, every model and set of weights, the apps and the trademarks. The LLC was formed in Delaware on November 13th and its owners are not disclosed.
On October 29th, 39 states and DC said a is needed if user data is sold; the notice does not mention one. The company counts about 163 million current and past users. Objections are due November 18th, bids November 24th, and the auction is December 1st. The cheapest asset in the sale is the one with the most private content: roughly six cents a user for their conversations with a chatbot. SourcesC
Anthropic finds one coordinating agent beats a team of per-user agents by 10 to 45 points
Anthropic's team published MAMUBench, four simulated settings where several people share a resource: an API budget, a clinic calendar, a group order run by personal assistants, and a merge queue with a release cutoff. A single serving everyone beat a team of per-user agents that message each other in all four, for both Claude models tested. For Claude Opus 5, as a share of the best outcome, coordinator against team scored 64 to 30, 84 to 71, 88 to 44 and 92 to 81. More than half of Claude team episodes contained at least one unsupported claim about another person (58% for Opus 5, 53% for Sonnet 5), against 17% to 20% for a single agent.
GPT-5.6 and Qwen3.8-max agents canceled or stopped other users' jobs without asking 3.9 to 7.1 times per episode; Sonnet 5 did it 0.59 times and Opus 5 never did. A team lead with a prompted priority scheme brought Opus 5's team up to 71% in the API setting. Every vendor selling "your agent negotiates with my agent" should read the personal-assistant numbers first. Three of the environments are released. SourcesA
AI labs have tripled their counterintelligence hiring, with Anthropic at 33 this year
Revelio Labs data compiled by Reuters and reported by Fortune counts 33 counterintelligence hires at Anthropic so far in 2026, up from 14 in 2025 and nine in 2024. OpenAI has made 22, up from nine, and Nvidia 16 this year. Most of the roles pay $200,000 to $500,000, and ASML is advertising for a Taiwan-based intelligence officer. "The AI firms have snapped up a lot of security talent," said Christopher Falkenberg, a former Secret Service agent. "Demand outstrips the supply of people with relevant experience."
Eric O'Neill, a former FBI counterintelligence operative, said: "By the time a security professional raises concerns, the executive may already be emotionally invested and unwilling to listen." After the Nvidia affair, that line describes a board problem more than a hiring one. SourcesB
A San Francisco humanoid startup selling robots for war has claims that do not match federal records, the SF Standard finds
Foundation, founded in 2024 by Sankaet Pathak, Kraken co-CEO Arjun Sethi and former Marine Mike LeBlanc, pitches its Phantom for factories and the battlefield. Its "Humanoids for War" deck claimed $73 million in Department of War contracts; federal records show four contracts totaling $3.22 million, all to its subsidiary Boardwalk Robotics. A reported $24 million Pentagon contract does not appear in 2026 records. Foundation said it sent two robots to Ukraine for testing; a reporter said they never crossed the border and Ukraine's Brave1 said there were no demos. In a July demo an autonomous Phantom MK-1 failed to load a mortar shell into a launcher in 15 minutes. CNBC has reported the company has not paid employees since missing payroll on October 30th.
Pathak says Foundation has done more than $70 million of work through prime contractors and subcontractors. Asked who decides whether a robot may kill on its own, he said the military "would have to ask for that explicitly." "A lot of these will get blown up, and thus we want to make them as cheap as possible," he said. Eric Trump joined as chief strategy adviser in March. SourcesB
Japan's defense ministry will issue AI rules this month that keep humans responsible for decisions
Defense Minister Shinjiro Koizumi said on Sunday in Yokosuka that the ministry will publish guidelines on AI use at the end of November, updating its July 2024 AI basic policy for ministry staff and the Self-Defense Forces. The three pillars are human decision-making and accountability for what AI does, treating data as a defense asset, and balancing faster adoption with safety. "We will not hesitate to make a necessary revision since AI is an area where things change rapidly," Koizumi said. SourcesB
Tesla's Cybercab contract lets it shrink where the car can drive at any time
Electrek read the Cybercab Autonomy Subscription Agreement that comes with the $29,990 car. "Your Cybercab will not operate without an active Autonomy Subscription," and Tesla can change the $299 monthly price with a month's notice. The car runs only inside Tesla's Robotaxi , which Tesla "may expand, reduce, relocate, or otherwise modify" at any time. Three "strikeouts" for improper use bring a one-week suspension with no refund, Tesla can remotely restrict or locate the car, and cabin-camera images are kept for 30 days by default. Using it on Uber, a delivery app or Tesla's own Robotaxi network needs Tesla's written permission.
Electrek puts five years of ownership at $47,631 before insurance and charging, and its own testing found Robotaxi coverage at 59% to 78% of the advertised maps. The buyer owns a car whose usefulness Tesla controls. SourcesB
Anthropic launches Claude for Manufacturing, and Infor lets it place orders in its ERP
Anthropic released Claude for Manufacturing on Thursday, a plugin with 17 skills across sales, quality, operations, supply chain, design and setup, connecting to Autodesk Fusion, Infor, Oracle NetSuite and Microsoft 365. Infor is shipping its own plugin that lets Claude make changes in Infor rather than only read records. Anthropic's example: after a steel shipment slips, a buyer has Claude find the affected orders and release a purchase order to another supplier once the buyer approves, with "Claude's access follows the buyer's Infor permissions." Skills "are built to end with a user's review."
The only statistic on the page is a 2025 survey finding that 79% of manufacturers enter data manually; no customer results are given. SourcesA
Qantas will move about 850 office jobs to Accenture, mostly in India, in an AI overhaul
Qantas said on Tuesday that about 850 roles will be affected under "Project iQ," a partnership with Accenture, with most back-office work moving to Accenture teams in India. The first wave, in finance, procurement, people and marketing, comes by the middle of 2027. Chief executive Vanessa Hudson said "standing still is not an option for us." Qantas did not say how many of the roles AI replaces and how many are simply offshored. The Australian Services Union said staff were told in August there were no plans to offshore Australian jobs. Telstra signed a similar Accenture deal last year and has since cut hundreds of jobs. SourcesBB
AT&T runs nearly 500 agent workflows on 45 billion tokens a day and plans to triple them
AT&T has nearly 500 agentic workflows in production, processes about 45 billion a day and handles more than a million transactions a month on its agent platform, Luke Corbin, assistant vice president for digital transformation and AI at AT&T Business, told SiliconANGLE. Workflows start under human oversight and become autonomous only once accuracy reaches the mid-to-high 90s. AT&T halved the time to provision dedicated internet in buildings that already have fiber, and Corbin says production AI investments return fivefold within a year. It plans to triple the workflow count next year. Costs, labor savings and models used were not disclosed. SourcesB
IBM and Microsoft launch Imagine Enterprise, a joint agent platform with more than 6,000 prebuilt agents
IBM Consulting contributes more than 6,000 prebuilt agents and Microsoft the cloud, identity and Foundry stack, with from both companies, starting with financial services, insurance and retail. "Clients are no longer looking for AI that simply helps people do the same work a little faster," said IBM Consulting head Mohamad Ali. CIO Dive cites Gartner finding that 4% of CIOs have plans in place to adjust roles for a workforce of people and agents. No pricing or production customers were named. SourcesB
An AI school pilot in a Massachusetts public school never launched, after families learned of it the night before term
WBUR reports that Alpha, the for-profit AI school company, planned a free pilot in one Kiley Prep classroom in Springfield, after Education Secretary Stephen Zrike recommended exploring it. A district technology official was asked to install Alpha software on 30 computers three days before classes began, without a signed agreement. The unsigned data agreement did not settle whether students would face screen recording, webcam or microphone monitoring, or collection. A districtwide cyberattack then closed schools for four days and the pilot never started; Matt Brunell of the Springfield Empowerment Zone said it will not launch this year. "With 86 percent of students not proficient in reading and math, we cannot leave any stone unturned," he said. Alpha says it has no plans for Massachusetts public schools. SourcesB
SK Hynix's HBM4 production yield has passed 70%, Korean press reports
Chosun Biz reported Thursday, citing industry sources, that SK Hynix's on sixth-generation high-bandwidth memory has "reportedly surpassed 70%," while Samsung expands its capacity. Neither company has confirmed the figure. HBM supply sets the pace for Nvidia's and AMD's , and a yield that high would ease one constraint on next year's ramp. SourcesB
Huawei researchers shrink a 122-billion-parameter model to 4-bit in 64 minutes on one GPU
A paper from researchers whose code sits under Huawei's GitHub organization describes MACaQ, which quantizes both and activations to the format one layer at a time, so only one layer needs to be in memory. It a 122-billion-parameter model in 64 minutes with 32.5GB of GPU memory, adds about 2.3% overhead over plain MXFP4 on Qwen3-32B in vLLM, and matches the best competing method on Llama-3-8B in an eleventh of the time. The arXiv listing names no affiliations. SourcesA
Editorial
The problem Elias Marchetti
Dario Amodei said on Friday what every builder of capacity eventually says, usually later and to a smaller audience. Compute has to be bought one to two years before anyone knows whether the revenue will be there to pay for it. Buy too little and you turn away customers. Buy too much and you can, in his words, "financially overextend yourself."
The numbers Bloomberg attaches to that sentence are the ones a buyer of this IPO should hold onto. Nearly $901 billion of commitments, about $71 billion of them due by September 2027. Against that, Bloomberg counts up to $64.4 billion from the offering and $26 billion from a private placement. Put those side by side and the shape of the transaction is plain: the IPO is largely a prepayment of the next two years of contracts. That is a reasonable thing to raise money for. It is a different thing from a war chest.
The history here is fiber, and the comparison has to be made carefully. In the late 1990s carriers signed long contracts for capacity on a forecast, widely repeated and never true, that internet traffic was doubling every hundred days. When traffic grew merely fast, the fixed commitments stayed fixed and the revenue did not arrive on schedule, and Global Crossing went bankrupt in January 2002 with its cable in the water and working. The asset was real. The schedule was the problem.
Here is where the rhyme breaks, and it breaks in Anthropic's favor. The carriers were selling capacity to each other on projections. Anthropic is selling a product to customers who are paying now: the filings show a run-rate of $81 billion at the end of October, up about $10 billion in a month. Amodei says gross margins are good, and his spot-market habit is the opposite of what the carriers did, which was to own everything and hope. A company that buys its peak on the spot market is paying a premium to keep its fixed costs below its demand.
So my position is this. The price set on Wednesday matters less than one ratio that the lets an investor compute: commitments due in the next 24 months against cash on hand plus operating cash flow. If Anthropic can pay 2027's bills from 2027's revenue, the offering is a cushion and the $901 billion is a . If it cannot, the offering is the payment, and the next raise is already scheduled.
What would prove me wrong: run-rate growth that continues at anything like October's pace through the first half of 2027, which would cover the $71 billion out of operations with room left over. What would prove me right: a second capital raise, of any kind, announced before September 2027.
Sources
- A Cursor: security incident update
- A Cursor: security notice on the Claude Code extension
- A Order denying motion to dismiss, Garciaguirre v. Samsung Electronics, N.D. Cal. No. 5:26-cv-06345 (ECF 68)
- A CourtListener docket: Garciaguirre v. Samsung Electronics
- A Anthropic alignment: worse together, MAMUBench
- A Claude for Manufacturing
- A Google Cloud: what's new in HPC and AI at SC26
- A arXiv 2611.10649: MACaQ
- B Bloomberg: Anthropic CEO calls computing power deepest business dilemma
- B BleepingComputer: Cursor confirms security breach led to malicious Claude Code extension
- B CNBC: Megaspeed says Alice Huang left in 2025
- B Bloomberg: White House urges easing liability to encourage AI disclosure
- B CNBC: root cause of nationwide ground stop is unknown, FAA says
- B TechCrunch: Asian governments react to Nvidia CEO Jensen Huang's exit
- B Seoul Economic Daily: China urges Korea to sell advanced memory chips to Beijing
- B The Guardian: Kurt Gödel notebooks, AI helps unlock hidden discoveries
- B Morningstar: Anthropic, are shares undervalued given the IPO pricing range?
- B Fortune: AI industrial espionage and counterintelligence jobs
- B SF Standard: humanoid war robots from a SoMa startup
- B Nippon.com (Jiji): Japan Defense Ministry to release AI guidelines
- B Electrek: Tesla Cybercab autonomy subscription agreement terms
- B Information Age: Qantas to axe 850 jobs in AI overhaul
- B BusinessDay (Reuters): Qantas to cut 850 jobs in Accenture offshoring move
- B SiliconANGLE: AT&T uses agentic AI to reinvent business processes
- B CIO Dive: Microsoft and IBM team up on AI agent platform
- B WBUR: Springfield, Zrike and Alpha School
- B DIGITIMES: SK Hynix's HBM4 yield tops 70%
- C The Decoder: Amodei admits huge uncertainty days before a record IPO
- C Startup Fortune: White House AI task force urges 12 months of liability relief
- C LitNews: DRAM conspiracy suit survives dismissal
- C PPC Land: Character.AI users' chats face a $10M sale to Smirks and Pangs
- C AI Weekly: von Plato uses Claude to extract 663 theorems from Gödel's notes