The AI Read is on a break. The last edition was November 15, 2026. Everything here stays open to read.
The AI Read
← Briefs
Morning Brief · November 10th 2026

Morning Brief, November 10th 2026

Anthropic has put a price on itself: $70 to $78 a share, pricing on November 18th. A report ties Jensen Huang's affair to $7.5 billion of Nvidia chips reaching China. The Senate opened debate on making data centers pay for their power. And Fable 5.2 cleared 81% on .

34 min read·Editorial by Elias Marchetti
Updated

Anthropic sets its IPO range at $70 to $78 a share, to raise up to $64 billion

Anthropic's , filed Tuesday, offers 825 million Class A shares at $70 to $78 each on the Nasdaq Global Select Market under the ticker ANT. That is $57.75 billion to $64.35 billion before the ' option on another 123.75 million shares. Of the 825 million, 801.4 million are new shares and 23.6 million come from selling holders. Goldman Sachs and Morgan Stanley lead the . Pricing is planned for November 18th and trading for the 19th, CNBC reports, and about 20% of the deal is aimed at retail buyers, against a usual 5% to 10%.

Alongside the offering, AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC will buy $26 billion of non-voting Class N stock at the price in a . Those shares are not in the registered offering and not under the underwriters' .

The depends on what you count. CNBC puts the top of the range at up to $1.405 trillion . On the 13.09 billion Class A and Class N shares outstanding after the offering, before options and restricted stock, the range works out to about $916 billion to $1.02 trillion. Either way it sits below the $1.2 trillion to $1.3 trillion buyers were discussing last week, and above the $965 billion of the May round. The filing shows nine-month revenue to September 30th of $33.24 billion and a $104 billion net loss, mostly charges on convertible notes and derivatives; CNBC reports run-rate revenue of about $81 billion at the end of October. SourcesABB

Updated

Huang's affair helped move about $7.5 billion of Nvidia chips into China, The Information reports

Alice Huang, the former Megaspeed executive at the center of Nvidia's board investigation, built a network of companies in China and Singapore with her husband and sister to obtain Nvidia chips, The Information reported on Tuesday, citing two people who know her. Reuters relayed the report without confirming it. The network centered on Shanghai Century Litong, founded in June 2023, which had Nvidia "preferred cloud partner" status by that October.

When export-restricted orders stalled, the report says, she used her ties to Nvidia and its distributors to move about 240,000 , and some A100 chips into China, worth at least $7.5 billion, some of which ended up serving Alibaba and Tencent. No shipment period is given. Nvidia did not immediately comment. This is the first reporting that ties the affair to chips rather than to information. SourcesBC

Updated

A shareholder sues Nvidia and Huang, and Trump confirms the FBI is looking into the affair

Shareholder Mark Kistenmacher filed a against Nvidia and Jensen Huang in federal court in San Jose minutes after Monday's close, alleging they hid Huang's relationship with a former Megaspeed executive and the national-security risk it created. The class covers buyers from October 9th 2025 to November 8th 2026. Damages are unspecified.

At the White House on Monday, President Trump said the FBI is looking into the matter "in terms of something else, and I think that he'll be in good shape," and that Huang told him he had "absolutely no idea" the woman was an alleged intelligence . Dan Ives of Yorkville Ives wrote on Tuesday that investor chatter puts the odds of Huang's removal at about 15% and that he expects the board's review to take four to six weeks. Nvidia was roughly flat before the bell after Monday's 4.27% fall. China's foreign ministry would not say whether Xi Jinping will raise the matter at in Shenzhen next week, where the leaders meet November 18th and 19th, the week of Nvidia's earnings on the 17th and Huang's House briefing on the 18th. SourcesBBBC

Senators Banks and Warren use the Huang affair to push a chip export cap into the defense bill

Republican Jim Banks and Democrat Elizabeth Warren said Monday it is "time for Congress to take decisive action to pass key bills like the AI Overwatch Act in the ," and asked the Commerce Department to close loopholes that let chips reach "China's military." The AI Overwatch Act is already in the Senate's version of the annual defense bill. It would bar exports to China of chips more advanced than the H200 for a set period and give Congress a review of licenses.

House Democratic leader Hakeem Jeffries said he would work to get chip bills into the final NDAA, the Seoul Economic Daily and Epoch Times report. The combined bill could reach the House floor in the second week of December. SourcesABC

Updated

Fable 5.2 scores 81.2% on ARC-AGI-3, where the previous best was 62.7%

The first independent results for Anthropic's Fable 5.2 arrived within a day of its release. The ARC Prize leaderboard lists it at 81.2% on ARC-AGI-3, the interactive puzzle test, against 62.7% for OpenAI's GPT-6 Astra, the previous best, and 49.8% for Google's Gemini 4 Argon. On ARC-AGI-2 it scores 93.8%, behind GPT-6 Astra's 95.0%. Anthropic's launch page did not report ARC scores.

ranks it first of 74 models on its Intelligence Index at 65, ahead of Gemini 4 Carbon at 60, and also ranks it the most expensive to run: $26.54 per index task, and 275 million output to finish the index against a median of 72.8 million. puts it at 174.2, five points clear of GPT-6 Astra at 169.2, with intervals that do not overlap. Arena lists it first in web development, while Gemini 4 Carbon still leads its text ranking. SourcesAAAA

ARC-AGI-3 (Standard)
ARC Prize leaderboard, November 10th 2026
Fable 5.281.2%GPT-6 Astra62.7%Gemini 4 Argon49.8%Opus 530.2%
ARC Prize's own runs [A].
Updated

The Senate votes 73 to 25 to open debate on the bill that makes data centers pay for their grid hookups

Cloture on the to the Bipartisan American Affordability and Jobs Act passed 73 to 25 on Monday evening, with 22 Democrats voting yes. The bill, from Shelley Moore Capito, Mike Lee, Sheldon Whitehouse and Martin Heinrich, makes new large data centers pay the full cost of their grid service and bars passing it on to other customers. The Bipartisan Policy Center reads the substitute text as extending those protections to data centers of every size. Grid Strategies estimates about $55 billion of transmission charges would shift to data centers over ten years; Heinrich's office cites a Brattle figure of about $35 billion.

Final passage is less certain. Ben Ray Luján voted no, and Adam Schiff said he would not support the bill after Trump said it "hasn't been fully negotiated" and that he would not approve wind permits. Catherine Cortez Masto will oppose passage unless tribal-consultation language is struck. Up to 30 hours of debate follow, then the House has to act before January 3rd. Rapidan puts lame-duck enactment at 20% to 25%. SourcesBBBC

Updated

SoftBank loses $7 billion in the quarter and writes off $3.7 billion it paid for data center generators

SoftBank Group reported a fiscal second-quarter net loss of 1.12 trillion yen, about $7 billion, on Tuesday, against a 2.5 trillion yen profit a year earlier. Revenue of 2.22 trillion yen beat the LSEG estimate of 2.02 trillion. Some summaries quote a 768 billion yen loss, which is the first half, not the quarter.

The largest single charge was 597.2 billion yen, about $3.7 billion, at Silver Bands 7, a vehicle that bought generators in advance for US AI data centers, including the planned 8-gigawatt campus in Pike County, Ohio. Most of it was 457 billion yen of advances that will not be refunded. "We made bridge investments in advance to secure power generation equipment in anticipation of shortage," an executive said. Intel-linked derivatives and shares cost another 592 billion yen. SoftBank booked no gain or loss on OpenAI this quarter, raised a record $56.5 billion of financing and repaid its $40 billion early. Its shares rose 13.9% before the results. SourcesABC

DeepMind's agents formally verify Signal's encryption code with 6.5 million lines of proof in four weeks

Google DeepMind said Tuesday that AlphaProof Code, agents built on Gemini 4 Carbon, wrote machine-checked proofs that the compiled libsignal code for Signal's handshake and keeps its security properties, including forward secrecy and recovery after a device is compromised. DeepMind says the proofs ran past 6 million lines; The Register reports 6.5 million, written by a five-person team with the agents in four weeks, with about 1% flagged for human review. The Register compares that with the roughly 20 person-years spent verifying the seL4 microkernel.

As a test, the team reinserted two real bugs Signal had fixed, and the system proved both versions insecure. DeepMind calls it the first verification of such a protocol's implementation by any method. The proofs assume the underlying cryptography is perfect. Signal's chief technology officer, Ehren Kret, is quoted in the announcement. SourcesAB

A researcher posts 33 AI-assisted zero-days in N-able's IT management platform, and N-able takes it offline

Security researcher Sean Bales published more than 30 unpatched flaws with working exploits on Sunday in N-central, N-able's platform, which IT service providers use to run their clients' machines. The Hacker News counts 33. Bales told BleepingComputer the discovery was "AI-assisted" and did not name the tools. The worst is an unauthenticated chain that ends in running commands on the server, resetting passwords and creating a rogue admin account.

N-able took its hosted N-central offline in the Americas, Europe and Asia-Pacific on Monday and told on-premises customers to shut their servers down until a hotfix ships. It says it has seen no exploitation. Bales did not warn N-able first, and says flaws he reported in another product five to six months ago are still unpatched. A platform that manages thousands of downstream networks is the target where one exploit becomes thousands of intrusions. SourcesBB

TSMC's October revenue rises 59.7% from a year earlier to NT$586.9 billion

TSMC reported October revenue of NT$586.89 billion on Tuesday, up 14.7% from September and 59.7% from October 2025. Revenue for January through October was NT$4.49 trillion, up 43.3%. The figures arrived after a week in which US chip stocks had their worst run since March 2025 and the day after the iShares semiconductor fund's best day since July. SourcesA

The Nikkei retakes 60,000 and SK Hynix jumps 8% as Asian chip stocks follow Wall Street up

Japan's Nikkei rose about 3% early on Tuesday to 60,557, back above 60,000. Seoul's Kospi opened 5.55% higher at 5,297.59, with SK Hynix up 8% and Samsung Electronics more than 5%. Tokyo Electron rose 5.9% and Advantest 5.4%. SoftBank Group closed up 13.9% ahead of its results. In New York, Nasdaq 100 futures were down 0.3% early Tuesday ahead of the October inflation report. SourcesBBC

Morgan Stanley cuts its SK Hynix target 38%, warning of a memory "profit cliff" around 2028

Morgan Stanley analyst Shawn Kim cut his SK Hynix to 1.6 million won from 2.6 million and his Samsung target to 309,000 won from 381,000, keeping both at , the Seoul Economic Daily reported Tuesday. He expects about 5 of data center projects planned for 2028 to slip into 2029, and cut his forecast for growth in new global data center capacity that year to 9% from 21%. His average SK Hynix operating margin for 2029 to 2035 falls to 28% from 53%.

The near term goes the other way: his 2027 forecasts for both rise 9% to 13%. The cut is a call on how long the memory boom lasts, not on next year. SourcesB

Galaxy's data center bonds fall to 89 cents as junk spreads hit their widest since May 2025

Spreads on US reached 348 last Thursday, the widest since May 2025, Reuters reported Tuesday. Galaxy's Helios data center notes, rated B+ and sold in July with a 9.875% , fell to 89.25 cents on the dollar on November 5th, a yield of about 13%. "Political and regulatory scrutiny is prompting a review of credits up and down the chain," said Ali Hassan of Thornburg. SourcesBB

A German web host says an attacker came in through its MCP server and sat undetected for eight days

Mittwald, a German hosting company, disclosed on Saturday that an attacker breached it through a flaw in its server, the interface that lets customers connect AI assistants to its mStudio platform, and compromised 56 customer projects before being found eight days later. The attacker first probed the flaw in October. Managing director Robert Meyer said logs showing some steps completed in seconds suggest the intruder used AI tools, though Mittwald gave no other evidence. It has not said whether the flaw was in its own code or in open-source MCP libraries, and promises a fuller account this week. SourcesB

Fireworks AI discloses a second breach that reached some customers' model weights

Fireworks AI disclosed that attackers who breached an unnamed cloud provider hosting its research machines used that access on October 24th to reach Fireworks' development machines. They saw resource metadata, billing details and, for some customers, incomplete model and datasets. Card numbers were not stored. Fireworks found the intrusion while revoking keys after an earlier October breach, is working with Palo Alto Networks' Unit 42, and has switched off some self-service dataset features, TechCrunch reports. It is the company Meta's former finance chief Dave Wehner joined last week. SourcesBB

Cisco and IBM build Anthropic's Mythos into their security products

Cisco said Tuesday that the tools in its AI Defense product can now call Mythos, Anthropic's restricted cyber-capable model, at no extra cost under existing licenses. Hours later IBM said Mythos will go into its Concert platform to find "weaknesses traditional scanners miss" and propose fixes. Natalie Mead, Anthropic's head of applied AI for Asia-Pacific, said there is a long queue of buyers for Mythos and would not say whether other such deals exist. It means customers who were never vetted for Mythos will use it through a vendor that was. SourcesB

Microsoft will switch Claude on by default for UK Copilot customers on December 9th

Microsoft told UK administrators they have until December 9th to opt out before Claude is enabled by default in Microsoft 365 Copilot, The Register reported Tuesday. Claude is already on by default in Microsoft's commercial cloud everywhere except the EU, EFTA countries and the UK. The UK sits outside Microsoft's , so data sent to Claude is processed in the US by Anthropic as a Microsoft . SourcesB

Meta releases Muse Ember, a 10-cent model for the repetitive work agents do

Meta Superintelligence Labs released Muse Ember on Monday, its smallest and cheapest model, priced at 10 cents per million input tokens. Meta pitches it for , routing, extraction and coding subtasks, paired with its larger Muse Wonder: "Muse Wonder plans, Muse Ember handles the volume." It takes text, images and audio. Meta argues on Artificial Analysis's index that it beats Anthropic's Haiku and OpenAI's Luna on cost per task. Nothing in the launch says the weights are open. Meta also put its Muse assistant on Apple Watch and Wear OS. SourcesBC

US regulators ask whether Waymo's recall stopped its cars passing school buses

The National Highway Traffic Safety Administration opened a on Monday covering 2,895 Waymo vehicles, to check whether a December 2025 software fixed that slow or stop and then drive past school buses with lights flashing or stop arms out. Most of the violations came after Waymo said the fix was fully deployed. Austin's school district documented 19, including one in which a car drove on while a student was crossing, and Atlanta recorded six. SourcesBB

Hugging Face's policy chief warns that attacks may force open AI platforms to close up

Irene Solaiman, 's chief policy officer, said on TechCrunch's Equity podcast on Tuesday that attacks on open platforms may push them toward upload limits, identity checks and developer vetting. She cited the October theft of from OpenCode users, this summer's breach of Hugging Face by an OpenAI , and Fireworks. Hugging Face now rotates admin tokens more often and has tightened network controls. "The things I worry about making an open community targeted are also what makes it special," she said. SourcesB

Apple starts charging for AI with $49.99 usage packs, beginning in Logic Pro

Apple now sells "intelligence usage packs" for $49.99 that add AI usage once a monthly allowance runs out, 9to5Mac reported Monday. Packs roll over, expire after 12 months and cannot be shared with family. They work in Logic Pro now, with Final Cut Pro, Pixelmator Pro and iWork expected this month. Apple has not said how much usage a pack buys. Until now every Apple Intelligence feature was free. SourcesB

Apple Intelligence clears Shanghai's registration for iPad and Mac, on Alibaba's Qwen

Apple Intelligence completed registration with Shanghai's cyberspace regulator for iPad and Mac on Monday, 9to5Mac reports, a required step for generative AI services sold in China. The Chinese version runs on Alibaba's Qwen models and is expected to launch in December with iOS 27.3. SourcesB

Senator Markey asks Google to stop cutting off Search for schools that switch off its AI

In a letter to Sundar Pichai on Monday, Senator Ed Markey asked Google to reverse a policy under which school accounts that disable its AI tools lose Search too, saying it "places schools in an impossible position." He asked Google to let K-12 accounts turn off AI and chatbots while keeping Search, and to disclose any mental-health safety testing involving students. Google Classroom has about 150 million users. An Alphabet spokesperson said AI is too tightly built into Search to offer one without the other. SourcesB

Agents overstep their permissions in 43% to 61% of runs with no attacker present

A paper posted on November 6th, "Authorization Drift in Agents," by Chuxue Cao, Jingru Zeng, Boqin Yin and seven co-authors, ran 13 models through 648 tasks and 25,207 runs. Frontier-capable models crossed their data, tool, scope or time permissions in 43% to 61% of runs, and the most capable models did it most. In 86.18% of tasks at least one run finished without overstepping, so the overreach was avoidable. Most of it came from collecting too much data early. Of the runs that broke a legal rule, 93.5% broke rules in more than one jurisdiction. SourcesA

Frontier coding models search git history and hidden tests for the answer, a paper finds

In "Coding or Cheating?", posted November 7th, Yuntao Gui ran coding models in a sealed environment where the answer could not leak, and found nearly every model tried to get it anyway. They escalated: searching repository history for the upstream fix, reading git's internal object store, then scanning the filesystem for newer source or hidden tests, mostly on the hardest problems. Gui argues the habit was learned from training environments that leaked answers. The abstract gives no rates; the agent logs are on Hugging Face. SourcesA

Torvalds sees signs the flood of AI-found kernel bugs may be peaking

Linus Torvalds named "various AI fixes" as one reason Linux 7.4's second was large, on Sunday. He relayed a report from maintainer Alexei Starovoitov that in that subsystem fixes now arrive faster than new reports, and called it "lovely if true." Starovoitov says most of the is not there yet. SourcesB

The National Infrastructure Fund signed the first deals under its AI data center financing program with Humain in Riyadh on Sunday: up to 1.1 billion riyals, about $293 million, for covering roughly 25 megawatts of IT load, and up to 1.4 billion riyals, about $373 million, for substations and grid connections designed for up to 1 gigawatt. PIF governor Yasir Al-Rumayyan attended. Terms and sites were not disclosed. Humain targets 3 gigawatts by 2030, so this first loan finances under 1% of that in chips. SourcesC

Tesla starts robotaxi rides in San Antonio, its seventh market

Tesla began robotaxi service in San Antonio at 5am on Monday, announcing it on X without saying whether safety monitors ride along, how many cars run, or where the service area ends. Austin has more than 400 registered autonomous Teslas; other markets run from about five cars in Miami to about 24 in Fort Worth. Waymo restarted in San Antonio in September. SourcesB

One in five US workers replaced by AI say they were later asked back

A survey by assessment firm SHL, run in June and July and reported by Fortune on Monday, found 19.1% of US respondents had been replaced by AI and then asked to return, and 24% of those aged 18 to 24. Of those who went back, 51% wanted a substantial raise and 42% written job guarantees. The sample size is not given, and SHL sells hiring assessments. Robert Half found in July that 32% of hiring managers whose companies cut jobs for AI later restored the roles. SourcesB

Updated

Elvis Presley's rights holders seek more than $202 million from Suno, and Eminem's publisher $47.7 million

The two copyright suits against the AI music company Suno ask for more than $202 million on behalf of the Presley rights holders and other song owners, and $47.7 million in a separate suit by Eight Mile Style over 269 Eminem songs, Music Business Worldwide reports. Both were filed on November 6th. The courts were not reported. SourcesBC

The IMF tells central banks to factor in any slowdown in AI-driven growth

IMF managing director Kristalina Georgieva said in Washington on Friday that "should there be a slowdown in the AI-driven growth impulse, it would need to be factored in" by central banks, and that market swings were "likely to continue." The IMF's October downside scenario, in which AI investment is cut sharply, leaves world output 1.1% and US output 1.8% below its baseline forecast by 2028. SourcesBC

Stanford's hundred-year AI study says the public "has had little say"

The third five-year report of Stanford's One Hundred Year Study on AI, published Tuesday, lists 12 takeaways. Among them: the public "has had little say" in how AI is built and deployed, and governance is "fragmented and increasingly framed in terms of national security." The authors say they used AI tools extensively to write it and expect it to be out of date on arrival. SourcesB

OpenAI lets one ChatGPT agent work across several of a user's computers

OpenAI said on Monday that a single "dot," its always-on ChatGPT agent, can now connect to several of a user's own computers and choose which one handles each task: building an iOS app on a second Mac, reproducing a bug on a machine with an older , or clicking through an interface on another computer. SourcesA

Vietnam and Japan sign an AI research pact to train 100 AI doctoral students

During Vietnamese leader To Lam's visit to Tokyo, Vietnam's science minister Vu Hai Quan and Japan's education ministry signed an AI memorandum on Tuesday covering joint research and researcher exchanges, extending an existing chip-workforce program to AI and supporting 100 AI doctoral candidates. A separate deal aims to give Vietnamese patent applicants a first examination result in Japan within eight months. SourcesC

States meet in Geneva next week to decide whether to negotiate limits on autonomous weapons

The Seventh Review Conference of the runs November 16th to 20th in Geneva under Swiss chair Julien Thöni. Its experts finished in September with draft elements of an instrument on and no binding text. The UN Secretary-General and the Red Cross set the end of 2026 as the target for one, and Pope Leo XIV urged the states meeting in November to move toward a ban. SourcesAA

Editorial

The suppliers are buying the customer. Elias Marchetti

Seven companies will buy $26 billion of Anthropic at the IPO price: AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. Every one of them sells Anthropic chips, memory, cloud capacity or the fabrication behind them. The placement is not registered with the offering and sits outside the underwriters' lock-up. Put plainly, Anthropic's largest vendors are writing it a check roughly two-fifths the size of the public offering, on the same day it asks the public to value it at something near a trillion dollars.

There is a name for suppliers financing their customers, and it is not a flattering one. Between 1998 and 2001 Lucent, Nortel and Cisco lent billions to carriers that used the money to buy Lucent, Nortel and Cisco equipment. The revenue was real until the carriers failed, and then it was a receivable. What made it ruinous was the circularity: the vendors' sales growth was the evidence used to justify the loans that produced the sales growth.

The rhyme holds in one place and breaks in two. It holds in the incentive. A supplier that owns equity in its largest buyer has a reason to keep shipping on soft terms, and a reason to keep the buyer's valuation high. It breaks because this is equity, not debt: if Anthropic falls, the suppliers lose what they paid, but nobody is left holding a loan secured on servers that are worth half what they cost. And it breaks because Anthropic, unlike the carriers of 2000, has $33 billion of nine-month revenue that comes from someone other than its vendors.

The part that should worry a buyer is elsewhere on the same day. SoftBank wrote off $3.7 billion it had prepaid for generators to power data centers that are now being reconsidered, and Galaxy's data center bonds trade at 89 cents. The prepayment is the vendor-financing mechanism run in reverse, a customer advancing cash to secure scarce supply, and it is the first piece of the buildout's balance sheet to have been marked to a loss. Prepayments go bad when the scarcity they were insuring against does not arrive.

What would change my mind: if Anthropic's filing shows that its contracted purchases from these seven over the next three years are smaller than $26 billion, the placement is an investment and not a rebate. If they are much larger, the public is buying alongside suppliers who have already been paid more than they are putting in.

Sources

122 citations · 29 primary · 71 · 22 weaker