Silicon Valley Daily
About 240,000 banned Nvidia chips reportedly reached China through Alice Huang's network. On the third day of the Huang affair, Axios reported Commerce Secretary Lutnick received a classified memo, Commerce again left Megaspeed off the Entity List, Schumer and Schiff called for investigations, and Nvidia fell another 5.5%. Anthropic's bankers are weighing a higher IPO price that could push its valuation past $1.5 trillion, per the FT. OpenAI told investors Q3 revenue reached about $10 billion, CoreWeave raised its outlook, and Mercor raised $750 million at $20 billion. New York opened registration under its AI safety law, and researchers found 35,000 Grok Build apps exposed.
Anthropic weighs raising IPO price above $70–$78 range after heavy demand
Anthropic launched its roadshow offering 825 million Class A shares at $70 to $78, implying up to $1.404 trillion fully diluted, plus $26 billion of non-voting stock sold at the IPO price to AMD, Broadcom, Google, Nvidia, TSMC and others, with Google and Nvidia taking $10 billion each. The Financial Times reports bankers are weighing a higher price after heavy demand, which could lift the valuation target to $1.5 trillion or more. An Australian retail offer opened Wednesday and closes November 17; shares are expected to trade on Nasdaq as ANT around November 19.
Nvidia board hires Wachtell Lipton, ex-acting US attorney Joon Kim, for Huang probe
Nvidia's independent directors retained Wachtell, Lipton, Rosen & Katz and former acting SDNY U.S. Attorney Joon H. Kim to investigate CEO Jensen Huang and whether company secrets or national security were compromised. Huang admitted on Sunday to an affair with Alice Huang, a Megaspeed executive U.S. officials have described as a Chinese intelligence asset. Critics including Casey Newton noted the board had already said it had no reason to believe secrets were compromised. CNBC reports Huang returned to headquarters Monday, apologized at the start of his regular executive meeting, took no questions and moved on to business; employees described the mood as sad. Nvidia reports quarterly earnings in a week.
Axios: Lutnick got a classified memo on Huang; Commerce leaves Megaspeed off Entity List
Axios reports Commerce Secretary Howard Lutnick was sent a highly classified memo warning about Jensen Huang's relationship with Alice Huang; per The New York Times, Secretary of State Marco Rubio oversaw its drafting and it circulated before Trump's May trip to China. The Financial Times reports Commerce delayed new Entity List additions by two months and left off Megaspeed despite four other agencies pushing for it. Former NSC official Chris McGuire called it a dereliction of duty and urged license requirements for large orders across Southeast Asia.
Updated S-1: $81B run-rate, $104B nine-month loss, founders selling $1.75B
Anthropic's amended S-1 puts run-rate revenue at $81 billion at the end of October, up about $10 billion from September; third-quarter revenue was $17 billion and the nine-month net loss $104 billion, against $900.9 billion in compute and infrastructure commitments. Each of the seven co-founders will sell nearly 3.4 million shares, about $250 million apiece at the midpoint. The Information reports the filing adds roughly $10 billion in new Microsoft and Google cloud commitments and lets Anthropic cancel a $34.9 billion Lambda deal after a delay, which it says it does not plan to do.
The Information: Alice Huang helped move about 240,000 banned Nvidia chips into China
The Information reports that Alice Huang, a former Megaspeed executive, helped move about 240,000 banned Nvidia chips into China, mostly H100 and H200 systems. Some reached data centers serving Alibaba and Tencent through Century Litong, an Nvidia preferred cloud partner with ties to her family. D.A. Davidson analyst Gil Luria said the biggest risk is that Nvidia could be barred from selling across broad regions of Asia if investigators find it improperly supplied companies serving China.
Google opens Gemini 4 Carbon to paid developers and AI Ultra users
Google made Gemini 4 Carbon, its strongest model, available via API to paying developers and enterprises and to AI Ultra subscribers, two weeks after releasing it to cyber defenders. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next best external model. It also previewed /swarm in Antigravity, which turns one prompt into a coordinated team of agents.
Nvidia falls 5.5%, briefly dips below $4 trillion; AMD and Broadcom rise
Nvidia closed down 5.47% on Tuesday after a more than 4% drop on Monday, briefly dipping below $4 trillion in market value before recovering above it by the close. AMD gained 1.3% and Broadcom nearly 4%, after rising 13% and just over 10% on Monday. Dan Ives of Yorkville Ives & Co. wrote that the Street assigns roughly a 15% chance Huang is removed, and expects the board probe to take four to six weeks.
Schumer and Schiff join Moolenaar in pressing on Huang and chip exports
Senate Minority Leader Chuck Schumer said on the floor that Trump was reportedly briefed more than a year ago and did nothing, and urged passage of the AI Overwatch Act, the MATCH Act and the Chip Security Act. Sen. Adam Schiff called on X for a full congressional investigation and immediate export controls on advanced chips to China. Republican House China committee chair John Moolenaar told CNBC he wants to know whether Huang shared privileged information and that the internal probe must not become damage control. Democrats take control of both chambers in January, giving them power to investigate Nvidia.
Anthropic launches Claude Meetings, Whiteboards, Sheets and Data
Anthropic released workplace tools that put Claude into video calls as a visible participant, drafting plans and answering questions while people talk; Claude Meetings is in beta for Team and Enterprise on Google Meet and leaves if an external guest joins. Claude Whiteboards and Claude Sheets let teams plan and build spreadsheets with formulas alongside Claude, and Claude Data is for telling stories with data. Bloomberg framed the launch as a challenge to Microsoft, and The Information reported Microsoft shares fell on the news. Investor Jason Calacanis wrote on X that Anthropic is folding Office, Google, Canva and Figma features into Claude and questioned whether free bundling raises antitrust issues.
OpenAI tells investors Q3 revenue rose about 50% to $10 billion
OpenAI told investors Tuesday that third-quarter revenue rose nearly 50% from the prior quarter to about $10 billion, after 17% growth in the second quarter, The Information reports. Leaked documents seen by The Wall Street Journal show its adjusted operating margin improved from -150% to -109% over the same period, still far from profitability.
CoreWeave beats with $3.66B revenue, $132.5B backlog, and lifts 2026 outlook
CoreWeave reported third-quarter revenue of $3.66 billion, up 168% and above the $3.55 billion estimate, with a net loss of $709 million. It raised 2026 revenue guidance to $13 billion to $13.3 billion, and shares rose more than 5% after hours following a roughly 40% slide over the past month. CFO Nitin Agrawal said that at current yields it will not issue more parent-level unsecured debt through 2027, relying mainly on customer prepayments and hardware-partner financing.
OpenAI launches Businesses in ChatGPT, with verified Business Agents
OpenAI opened a beta that lets companies claim and manage a profile in ChatGPT and see how users engage with them. Eligible US e-commerce and online service businesses can build a Business Agent that answers product questions and handles comparisons when the business appears in a ChatGPT answer. TechCrunch described the move as making ChatGPT more like Google as a discovery engine.
Mercor raises $750M Series D at a $20 billion valuation
Mercor, which supplies human experts for evals and training data to frontier labs, raised a $750 million Series D at a $20 billion valuation from investors including Benchmark, General Catalyst, Nvidia, Salesforce and Zoom. CEO Brendan Foody said run-rate revenue grew from $1 million to $3 billion in 27 months, accelerating over the past six months as enterprises began training their own models.
Nvidia's 11 directors, most longtime Huang associates, face the CEO decision
The Information reports that most of Nvidia's 11 directors have known Huang for decades and none is a former CEO of a major company. CFO Colette Kress could serve as interim chief but is not seen as a lasting successor. Separately, the Financial Times reports Alice Huang built her career alongside Hexi Capital founder Kang Wei, who faced stock market probes, and that sources believe she is back in China.
New York opens RAISE Act registration for frontier AI labs, launches testing center
New York opened a registration portal Tuesday under the RAISE Act: frontier developers with more than $500 million in annual revenue must register by January 1, 2027, publish safety frameworks and report critical incidents within 72 hours. Fines reach $1 million for a first violation and $3 million for later ones. Governor Kathy Hochul also launched a Center for AI Responsibility and Research at SUNY Binghamton to independently evaluate advanced systems, led per Politico by Jeannette Wing and Greg Morrisett. California's SB 53 allows 15 days to report incidents.
TSMC board approves up to $500M each for Anthropic and OpenAI stock
TSMC's board approved buying up to $500 million of Anthropic shares at the IPO price and up to $500 million of OpenAI preferred stock, Reuters and Nikkei reported. TSMC is among the seven companies in Anthropic's $26 billion private placement. The world's largest contract chipmaker is now taking equity in two of its biggest end customers.
SoftBank writes off $3.7B in data center advances; AI equity raises top $300B
SoftBank wrote off $3.7 billion of non-refundable advance payments for data center equipment after reviewing its plans, out of $9.5 billion committed, The Information reports. The publication also counts more than $300 billion raised by AI companies through stock sales this year, with about $100 billion more possible, as falling share prices make the build-out harder to fund. CoreWeave's 2032 bond yield has climbed above 16%.
Grok Build share links exposed 35,000 apps, including SpaceXSI internal files
Security firm Claroc found that share links for Grok Build, SpaceXSI's coding agent, could be enumerated, leaving about 35,000 apps open to anyone. Exposed data included medical exams for 3,784 workers at a Peruvian construction firm, a Tesla service-center scorecard naming 22 technicians, and a sales territory map and confidential onboarding plan from 24 of SpaceXSI's own apps. SpaceXSI pulled the 24 apps within hours of Ars Technica's inquiry; other users' data remained exposed at publication.
Fireworks discloses second October breach; Notion says some AI chats exposed
Fireworks AI said a third-party cloud provider hosting its R&D machines was compromised on October 23, and attackers reached its development systems the next day, accessing billing records, resource metadata and, for some customers, incomplete model weights and datasets. Notion said fewer than 1% of its workspaces were affected, with unattributed Notion AI chat transcripts exposed, and stopped serving open-weight models through Fireworks. Bloomberg reports Cursor and Harvey are also assessing the impact.
Stripe completes its acquisition of OpenRouter
Stripe said OpenRouter, the API gateway that routes developer requests across AI models, is now part of the company. Stripe said the combination will help businesses try new models and optimize token routing and usage. Terms were not disclosed.
Oura to buy AI primary care startup Counsel Health for $625M
Smart ring maker Oura agreed to acquire Counsel Health, which pairs medical AI chat with licensed physicians and was integrated into the Oura app in June, for $625 million. Oura said quarterly revenue grew 153% to $533 million and fiscal 2026 revenue reached about $1.75 billion. The deal comes weeks after Oura postponed an IPO of up to $2.2 billion.
Google adds TPU 8tx for confidential training and TPU 8i Max
Google unveiled two eighth-generation TPU variants. TPU 8tx, designed with AMD, builds security into the hardware so data and weights stay isolated during training and inference. TPU 8i Max combines two inference-focused TPU 8i chips in one package, doubling memory, compute and bandwidth.
DeepMind's Gemini Robotics 2.5 hits 96% sorting parts at Mercedes plants
Google DeepMind released Gemini Robotics 2.5, which it says achieves 30% to 100% higher success rates than the prior version and runs 20% to 30% faster. In a pilot on Mercedes-Benz lines in Hungary and Germany, Apptronik's Apollo 2 humanoid sorted metal parts with a reported 96% success rate. DeepMind also open-sourced ASIMOV-Jailbreak, a benchmark for robot safety.
AMD details Zen 6: 27% average IPC gain, up to 256 cores; Verano targets Vera Rubin
AMD said its Zen 6 cores deliver an average 27% IPC uplift across AI and server workloads, with clocks from 4.1 to 5.3 GHz and Epyc parts of up to 256 cores. Its Verano CPU is built for integrated CPU-plus-GPU racks, which AMD positions against Nvidia's Vera Rubin. AMD also revived its Epyc X line after three years with Venice-X for EDA and HPC.
Railway says October hackers stole AI provider keys, running up about $740K
Cloud host Railway says attackers exploited flaws in its infrastructure to steal customer secrets in October, apparently to resell AI provider API keys, BleepingComputer reports. Unauthorized spend on the stolen keys was estimated at about $740,000.
Trump nominates Palantir security chief Michael LoSapio to lead CISA
President Trump nominated Michael LoSapio, Palantir's CISO since September 2024, to run the Cybersecurity and Infrastructure Security Agency. If confirmed, he would end a nearly two-year wait for a Senate-confirmed director since Jen Easterly left in January 2025. The agency's workforce has been sharply reduced.
PSI researcher used Fable 5.2 early access to settle 18-year-old quantum conjecture
David Turturean of startup lab Physical Superintelligence used early access to Anthropic's Fable 5.2 to prove the Aaronson-Ambainis conjecture on when quantum computers can outpace classical ones. Scott Aaronson wrote that he believes the proof is correct because it was autoformalized in Lean. The negative result closes off one route to quantum speedups.
Microsoft patches 1,093 flaws in November, including an exploited Outlook bug
Microsoft's November Patch Tuesday fixes 1,093 vulnerabilities, 63 of them critical and 535 allowing privilege escalation. The only one under active exploitation is an Outlook spoofing bug, CVE-2026-124194. Microsoft has shipped more than 5,100 security patches this year, including over 1,400 last month, versus about 100 a month a year ago, and credits AI-assisted bug finding for the increase.
VulnCheck: Glasswing has 37,966 findings but just 217 marked fixed
VulnCheck counts 37,966 findings reported in seven months of Anthropic's Project Glasswing, 7,349 made public and only 217 marked fixed; since September nearly 12,000 findings were added against 15 new fixes. It attributes 442 CVEs to Anthropic, two exploited in the wild, and notes curl maintainers rejected 19 of 34 reports.
Nvidia's GR00T 2 learns robot skills from 400,000 hours of human video
Nvidia's GR00T 2 is a world-action model that predicts both the next movement and a short video of what the camera will see, aiming to transfer skills across robot bodies. It was trained on more than a thousand GB300 chips with about 400,000 hours of head-camera footage of people doing tasks. Success rates still range from 10% to 80% depending on the task.
OpenAI acqui-hires Odessia, the AI travel agent from Sonder founder Francis Davidson
Sonder founder Francis Davidson said OpenAI acquired Odessia, his year-old AI travel agent, and that he has joined OpenAI to work on travel; terms were not disclosed. Odessia had raised a $6 million pre-seed led by Sequoia, and the product has been shut down. An OpenAI spokesperson said 1.2 billion people use ChatGPT weekly, with trip planning among common uses.
Amazon swaps Claude Sonnet 4.6 for Strands Decider, cutting ticket-labeling cost 98%
Two Amazon engineers replaced Claude Sonnet 4.6 with Strands Decider, a decision model that returns probabilities over fixed options, for labeling tickets in one of the largest internal ticketing systems. A Gemma 4 E4B build matched or beat Sonnet on their test sets at 98% lower cost, which the team says will save millions of dollars a year.
Point72 unit Cubist sues ex-analyst, alleging he used Claude and Codex to take code
Point72's quantitative unit Cubist alleges in a lawsuit that a fired lead research analyst used Anthropic's Claude and OpenAI's Codex to steal source code before joining a competitor, Bloomberg reports.
N-able takes N-central offline after researcher dumps dozens of zero-days
N-able took its N-central remote access service offline and urged self-hosted customers to shut down servers after researcher Ben Bales published a set of zero-day bugs he calls Able2Shell. Bales says they allow credential theft, creation of admin accounts and root access. N-able says it has no evidence of exploitation in customer environments.
Apple is testing Harbor, a Mac app that pools Macs into a local AI server
Apple is privately previewing Harbor, a Mac app under its Claris brand that runs open-weight models locally on Apple silicon, 9to5Mac reports citing leaker Aaron Perris. Macs linked over Thunderbolt 5 can pool unified memory to run models too large for one machine.
Nvidia donates OpenShell agent runtime to CNCF, launches Kubernetes Engine
At KubeCon, Nvidia contributed OpenShell, an open-source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation as a sandbox project. It also announced NVIDIA Kubernetes Engine and Kubernetes Recipes for running managed Kubernetes on GPU infrastructure, and is contributing to Google-led Agent Substrate.
Daylight, Coframe, Graphon, General Translation and Garage raise new rounds
AI managed-security firm Daylight raised an $80 million Series B led by Goldman Sachs Alternatives at about a $400 million valuation, the WSJ reports. a16z led a $27 million Series A for localization startup General Translation, whose customers include Cursor and Ramp. Business Insider reports website personalization firm Coframe raised a $25 million Series A and agent memory startup Graphon a $35 million Series A led by RRE; YC-backed heavy equipment marketplace Garage raised a $41 million Series B.
SpaceX's Texas AI data centers face power permit delays
SpaceX is pursuing large AI data center projects across Texas, but delayed power permits could leave it paying for leased chips before facilities are ready, The Information reports.
a16z: HubSpot's Dharmesh Shah on where startups can build off the labs' path
In an a16z podcast recorded September 30, HubSpot co-founder and CTO Dharmesh Shah argued that AI should be treated as infrastructure, as AWS was 15 years ago, and that frontier labs' business is selling tokens to recoup their spending. He said most business apps derive value from the work done, not the interface, so being reached through ChatGPT or Claude does not erase it. HubSpot has launched uSpot, a CRM for one-person businesses.
All-In: Axon CEO Rick Smith says Taser-drone R&D continues despite 2022 board exits
On All-In, Axon CEO Rick Smith said the company has become the largest investor in drone maker Skydio, letting police launch drones from fixed docks that sometimes reach a scene before dispatchers finish logging the call. He said the Taser-armed drone project that prompted ethics advisers to quit in 2022 is still in R&D, arguing non-lethal force delivered remotely beats gunfights, and that the strongest demand comes from poorer, high-crime neighborhoods.
Meta turns to consulting firms after enterprise AI pushback; Muse reaches watches
Meta is building consulting partnerships to sell AI products to large companies after resistance from retailers including Amazon and Booking.com over customer data and control, The Information reports. Meta's Muse assistant also became available on Apple Watch and Wear OS via voice, text and dictation.
Trending on X
What X was arguing about today that the news hasn't covered. Views are the posters' own.
Redwood's Buck Shlegeris puts one-in-three odds on superintelligence within a year
Buck Shlegeris wrote that there is about a one in three chance AI companies build superintelligence within the next year, and that such systems would probably take over human institutions. He cited new METR research finding current AIs increase the rate of AI progress at Anthropic by about 70%. Nate Soares said the rate at which private models automate AI research is what sets timelines. Ed Zitron replied that none of the people saying this live as if it were true.
Anthropic's roadshow says Mythos got Claude running on AMD GPUs in three days
Andrew Curran quoted Anthropic's roadshow video: bringing up a new chip platform used to take dozens of engineers months, but Mythos got Claude running on AMD in three days with basically no human intervention, porting or writing almost 100% of the kernels. Investor Marcelo Lima responded that the Nvidia CUDA moat is gone because Anthropic treats all compute as fungible.
Andrew Curran: Anthropic's S-1 says it has met its RSP threshold for AI-accelerated R&D
Andrew Curran posted that Anthropic's latest S-1 amendment says it has met its Responsible Scaling Policy threshold, based on internal estimates that it “may experience a further increase in the rate of progress.” The tracker account MTS described the threshold as the one for AI-accelerated R&D. The post circulated alongside the METR report on Anthropic's research speed.
Mitchell Hashimoto says everyone at Superlogical pushes straight to main
HashiCorp co-founder Mitchell Hashimoto described how his new company Superlogical works: everyone can push to main with machine-aided review before and human review after, with few PRs. Anyone can cut a release, which starts a 15-minute countdown that anyone can pause but not override. He said cross-platform feature parity arrives within 24 to 48 hours and that the system works only on mutual trust.
A user asks for "Alive Maps" that route for beauty instead of speed
Maja wrote that a walking route two minutes longer than the maps default takes her past a park, architecture and a street of cafés, and asked for a version of Google Maps that optimizes for beauty and aliveness. Replying, @aliceisplaying said AI makes it buildable now, but recalled a Google employee explaining it doesn't exist because it would routinely route people away from poor neighborhoods.