AI News: November 6, 2026

1. The Nasdaq 100 Loses a Record $2.3 Trillion in a Day as the AI Selloff Spreads to Asia

Markets. The Nasdaq 100 fell 6% on November 5, losing a record $2.3 trillion in market value, and the Philadelphia Semiconductor Index dropped about 12%, as investors sold AI stocks on expectations that the new Democratic majorities in Congress will bring tighter AI rules and slower data center construction, Bloomberg reports. On Friday, South Korea’s KOSPI opened 8.55% lower and set off a 20-minute trading halt as Samsung Electronics fell 10% and SK Hynix 11.7%, Japan’s Nikkei fell 6.7%, and Taiwan’s Taiex posted its largest intraday point drop on record, according to Investing.com and Nikkei Asia. The Wall Street Journal separately reported that executives at an election event hosted by Jensen Huang voiced concern about the anti-data center results, and that OpenAI has discussed resetting its 2027 capacity plans. Source Investing.com Nikkei Asia via Techmeme WSJ via Techmeme

2. METR Finds Anthropic’s Models Are Improving About Twice as Fast as Before, Mostly Because of Its Own AI Use

METR. After a high-access evaluation inside Anthropic from August 12 to October 29, the evaluation nonprofit estimates that Anthropic is improving its models on well-specified tasks at about twice the March 2024 to October 2025 rate (it would not be surprised by anything from 1.3x to 3x), and that scores on AECI, Anthropic’s version of the Epoch Capabilities Index, have grown 1.9x faster since October 2025, about two years of historical progress in one. METR attributes most of the acceleration to Anthropic’s use of AI in its own R&D rather than to faster growth in training compute, saw no bottlenecks likely to stop further acceleration over the next 12 months, and says Anthropic redacted a majority of the evidence it found most informative. Business Insider adds that the report counts three models Anthropic finished training in the past six months that are not public, and that staff Slack polls put the average self-reported uplift from the most capable internal model at about 12x, which METR treats as weak evidence. Source Business Insider

3. Thinking Machines Lab Releases Inkling 2 Small, a 594B MoE With Apache 2.0 Weights Promised Next Week

Thinking Machines Lab. The lab’s second-generation model has 594B total and 17B active parameters, accepts image input, supports variable reasoning effort, and was trained with RL on hundreds of thousands of long-horizon coding tasks across several agent harnesses; it is available now on Tinker with closed weights, and the lab plans to release full weights under Apache 2.0 next week after external safety testing and red-teaming. Artificial Analysis scores it 31 on its Intelligence Index v5.0.1, 18 points above Inkling Small and level with Gemini 3.8 Flash (high), which would make it the highest-scoring open-weight model developed outside China once the weights ship; it scores 25% on FrontierCode 2.0 against 11% for Claude Sonnet 5.5 (medium), but 13% on Terminal-Bench 4.0 against 30%. API pricing is $0.30 per million input tokens and $1.00 per million output tokens, although the model uses about 150K output tokens per index task, roughly four times its predecessor. Source Artificial Analysis

4. Thinking Machines Lab Is Reportedly Raising $5 Billion at a $40 Billion Pre-Money Valuation

Thinking Machines Lab. Accel is leading the round, which could close as soon as this week, Business Insider reports, citing multiple sources; it would more than triple the valuation of Mira Murati’s lab, which was valued at more than $10 billion after a $2 billion seed round last year. Talks last year over a round at a $50 billion valuation were abandoned, and the company declined to comment. Business Insider notes that about a third of the founding team has left, including Barret Zoph, Luke Metz and Andrew Tulloch. Source

5. Baidu Releases ERNIE 5.5 Flash Base, a 280B-Parameter Base Model Shipped Without Post-Training

Baidu. ERNIE-5.5-Flash-Base is a sparse mixture-of-experts model with 280B total and 14B active parameters, pre-trained from scratch on 33.5 trillion tokens with a 128K training context and support for 1M tokens during fine-tuning and inference. Baidu released the checkpoint from the end of pre-training, before any mid- or post-training, as a common starting point for domain adaptation, instruction tuning and RL research, and says hybrid sparse attention keeps KV cache and long-context costs down. The weights are on Hugging Face in FP8 and BF16, Baidu reports results comparable to DeepSeek V4 Flash and MiMo V2.5 on reasoning, code and long-context benchmarks, and a post-trained version is planned, according to The Decoder. Source The Decoder

6. Investors Reportedly Doubt Anthropic’s IPO Can Reach $1.5 Trillion After the Tech Selloff

Anthropic IPO. Three major investors planning to buy into the offering told the Wall Street Journal that the valuation would likely have to come down, with one saying it could end up closer to $1 trillion, and that shrinking or postponing the listing are also possible. Anthropic had been aiming to list the week of November 16 and raise up to $100 billion, which would be the largest IPO on record, and Reuters reported that investors are questioning whether the roadshow can start next week as planned, Mint reports. Earlier this week investors were reported to be pushing for a valuation around $1.5 trillion; the new reports follow Thursday’s selloff. Source Mint

7. Shield AI Raises $2.5 Billion at a $28.8 Billion Valuation to Buy Radar Maker Echodyne for $5.3 Billion

Shield AI. The defense autonomy company is raising $1 billion in Series H equity at a $28.8 billion post-money valuation, $1 billion in structured equity from Blackstone-managed funds and $500 million in debt led by JPMorgan Chase to fund the deal, of which $4.8 billion is paid at closing and up to $500 million depends on Echodyne meeting financial targets. Echodyne builds solid-state MESA radars used in counter-drone and security systems in more than 40 countries and won a contract with a $250 million ceiling from the Army’s Joint Interagency Task Force 401 in September. CEO Gary Steele says the combined company will run autonomy software, autonomous systems and mission systems units, and the deal is expected to close in early 2027. Source

8. Microsoft Reportedly Pulls Back From Riskier Neocloud Deals as IREN Says More Than $4 Billion in Microsoft Revenue Slips

Neoclouds. Microsoft is tightening oversight of some data center projects, stepping away from those it considers risky and debating whether to sign new neocloud contracts, Bloomberg reports; Nebius fell about 14%, IREN about 17% and CoreWeave about 10% after the report, according to Finimize. In its earnings report the same day, IREN said more than $4 billion in annual recurring revenue expected to start in 2026 has moved to at least January because of “a customer-requested change,” and co-CEO Daniel Roberts named Microsoft as the customer, which asked to reconfigure network infrastructure for 200 megawatts of capacity. IREN’s revenue rose 66% from the prior quarter to $228 million and AI cloud revenue more than doubled to $200 million, after its shares fell 55% over the past month. Source CNBC Finimize

9. About 600,000 Nvidia H200s Are Reportedly Stuck in Hong Kong, While Moonshot AI Looks to Europe for Blackwell Capacity

China AI compute. Roughly 600,000 Nvidia H200 chips are stored in Hong Kong awaiting shipment to mainland China because Beijing is withholding approval for large-scale imports, Reuters reports, citing people familiar with the matter; several major Chinese internet companies have each obtained about 10,000 units in recent months, and an eight-card H200 server was offered in the mainland for more than 6 million yuan at the end of October, about three times its price a year earlier. Separately, CNBC reports that Kimi developer Moonshot AI has held talks with at least one European data center to lease clusters of 8,000 or more Blackwell GPUs, mainly for inference but also for training. Renting overseas capacity is the route targeted by the Remote Access Security Act, which has passed the House but still needs the Senate and the White House. Source Tom’s Hardware

10. House Republican Chairs Ask Jensen Huang for a Closed-Door Briefing on His Ties to a Reported Chinese Asset

U.S. House. Foreign Affairs Committee chair Brian Mast, Select Committee on China chair John Moolenaar and Intelligence Committee chair Rick Crawford invited Nvidia’s CEO to an in-person briefing on November 18, the day after Nvidia reports third-quarter earnings, about his interactions with Megaspeed executive Alice Huang and Nvidia’s “security protocols against foreign adversary espionage.” The New York Times reported last week that U.S. intelligence determined Alice Huang, a Chinese citizen, was passing information to Beijing and had developed a close relationship with Jensen Huang. Nvidia did not say whether he would attend and repeated that its regional cloud customers receive “ordinary business information,” not sensitive trade secrets. Source

11. Zipline Raises $1.5 Billion at a $21.5 Billion Valuation as Walmart Adds 300 Drone Delivery Sites

Zipline. Paradigm, WCM Investment Management and Layer Global led the round in the autonomous drone delivery company, with Valor, Fidelity, Uber, Tiger Global, J.P. Morgan, ARK Invest and Salesforce Ventures participating; the Wall Street Journal notes the valuation is up from $7.85 billion in January. Walmart signed a multiyear national agreement for 300 more delivery sites, and Uber is targeting one million deliveries a day through Zipline’s network by 2029. Zipline says its median order in October arrived 19 minutes after checkout and that it completed more than one million deliveries in Dallas-Fort Worth this year; the money goes to U.S. manufacturing, hiring and a nationwide network. Source WSJ via Techmeme

12. Cyera Is Reportedly Raising About $800 Million at a $20 Billion Valuation

Cyera. The five-year-old cloud data security startup plans to raise around $800 million in a round led by Cyberstarts at a $20 billion valuation, up from $12 billion in June, The Information reports. The round comes in the same week as CrowdStrike’s agreed purchases of Fiddler and Dreadnode and Palo Alto Networks’ reported talks to buy agent credential startup Clutch Security. Source

13. Dropbox Buys Tigris Data for $350 Million to Build Storage for AI Agents

Dropbox. The deal, Dropbox’s largest acquisition, supports DBX Agent Cloud, a new storage platform for companies building AI models and agents that will first offer S3-compatible object storage as a standalone product, followed by support for documents, images, audio and video and filesystem features that let agents keep continuity across tasks and sessions. Tigris, founded by former Uber storage engineers, automatically places data near the users and compute that need it, and the deal is expected to close in the fourth quarter. Dropbox is also letting users run AI tasks on their files, pick models and connect content to ChatGPT, Claude and Muse, an update it says has reached more than 70,000 teams. Source

14. Humble Robotics Raises a $51 Million Series A for Cabless Autonomous Electric Trucks

Humble Robotics. Obvious Ventures led the round, with insiders Eclipse and Energy Impact Partners participating, bringing total funding to $75 million less than a year after the company was founded. Humble designs its electric Humble Hauler and its autonomy software as one system for dock-to-dock routes under 250 miles between ports, rail terminals, warehouses and distribution centers, a segment it says accounts for 87% of U.S. truck tonnage. The money goes to customer pilots, testing and validation, early manufacturing and a wider family of vehicles. Source

15. SB 53 Author Scott Wiener Wins Pelosi’s House Seat After Campaigning on AI Regulation

Scott Wiener. The California state senator won Tuesday’s election to succeed Nancy Pelosi in the U.S. House after campaigning on AI regulation, The Information reports. Wiener led the effort that passed SB 53, California’s AI safety law, through the state legislature last year, and his race was still being counted when the first tallies of AI PAC results came out on Wednesday. Source

16. Capgemini Will Lead an EU AI Office Consortium Assessing Cyber-Offense Risks From General-Purpose Models

European Commission. The AI Office picked a consortium led by Capgemini, with Wavestone and France’s national metrology and testing laboratory LNE, to support cyber threat modeling and risk assessment of general-purpose AI models under the AI Act, under Lot 2 of procurement EC-CNECT-2025-OP-0032. The work covers modeling cyber-offense risks such as automated vulnerability discovery, exploit generation and large-scale attack orchestration, building evaluation tools and technical compliance methods, and running on-demand technical assessments to help enforce the Act’s obligations for GPAI providers. Source

“Who Pays for AI?” Using de-identified records from 14.2 million UK bank and card accounts, economists Nikhil Datta and Peter John Lambert of the University of Warwick and Patrick Schneider of Imperial estimate that UK adults spend about £1.07 billion a year on AI assistants at July’s rate, with 5.1% of adults paying in June, up from 2.3% a year earlier. Claude’s share of spending rose from 7.4% in December 2025 to 29.4% in July while ChatGPT’s fell from 73.5% to 52.5%, although ChatGPT still had 71.2% of paying subscribers, and 6.2% of Claude subscribers pay for plans of £80 or more a month against 0.3% for ChatGPT. The study counts only subscriptions people pay for themselves, and Claude’s July share falls to 22.8% once estimated app store payments are included. Source

18. Lawsuit Alleges ChatGPT Helped a Woman Plan to Poison Her Husband

Gates v. OpenAI. A lawsuit filed by James Gates in San Francisco County Superior Court and made public this week alleges that OpenAI and Sam Altman failed to shut down chats in which his wife Cheryl, indicted in October on charges of attempted first-degree murder and food contamination, obtained instructions on preparing, dosing and administering oleander over at least 25 exchanges, including whether it would show up in an autopsy. The complaint says ChatGPT repeatedly calculated lethal doses of medications in the home for someone weighing about 200 pounds and said Celsius energy drinks would mask the taste better than beer or soda, instead of referring her to law enforcement or ending the conversation. Source

19. Airbnb Commits to Launching Its Own AI Travel Agent in the First Half of 2027

Airbnb. In its third-quarter shareholder letter, Airbnb said the agent will assemble responses from a library of hundreds of travel-specific interface components, such as maps, photos, comparisons, calendars and itineraries, rather than replying in text, and will handle in-trip tasks like changing reservations, payment problems, refunds, rebooking and 24-hour safety support that “an outside agent can’t easily do on its own.” CEO Brian Chesky told analysts that general-purpose agents are not the best way to plan travel. A filing also disclosed that Airbnb holds an option to invest in and license technology from Chesky’s stealth AI lab, which a source told The Information he has funded himself so far; Airbnb’s quarterly revenue rose 19% to $4.87 billion. Source The Information via Techmeme

20. Agility Robotics Becomes the First Dedicated Humanoid Robot Maker to List in the US

Agility Robotics. The maker of the Digit humanoid began trading on the Nasdaq under the ticker AGLT after a SPAC merger that raises $465 million, Bloomberg and TechCrunch report. Agility had revenue of $174,000, mostly from related parties, and a $362 million loss in the first half of 2026, and says it has $300 million in planned contracts from 30 customers for Digit 5, a version designed to work safely around people, contingent on throughput, uptime and placement accuracy targets. CEO Peggy Johnson cites 65,000 hours of operational data from customer facilities, and the company says its factory can produce 10,000 robots a year. Source Bloomberg via Techmeme

21. Box Builds Nvidia’s Nemotron Embeddings Into a Content Retrieval Service for Agents

Box. Shown at Box’s developer conference, Box Content Retrieval combines NVIDIA’s open-weight Nemotron Embed models with Box’s existing metadata and permission controls so that agents receive the most relevant document chunks instead of reading whole files. NVIDIA’s Kari Briski told TechCrunch she expects queries to be answered four times faster with 20% fewer errors and 90% fewer tokens. The service rolls out to Box Enterprise Advanced customers early next year. Source

22. Axiom Releases a 16-Million-Line Lean Formalization of the Entire Stacks Project

Axiom. The math AI startup, valued at $1.6 billion, formalized the roughly 7,500-page open-source algebraic geometry reference that Columbia’s Johan de Jong started in 2005 as a single machine-checked Lean library of about 16 million lines, which it calls the largest integrated library of formalized mathematics. Coordinated swarms of agents translated, proved and integrated theorems in parallel, picking many of their own tasks within a loose hierarchy, and permanent Stacks tags link each Lean declaration to the original text. Axiom says researchers and agents can now import the definitions and theorems into new formal proofs in a field that previously had little Lean infrastructure; it raised a $200 million Series A led by Menlo Ventures in March. Source TechCrunch

23. Discovered Materials Puts $1 Million in Prizes Behind Five Unsolved Materials Science Problems

Discovered Materials. The Y Combinator spring 2026 startup, which uses AI to find materials for advanced chips, worked with academic advisers including Stanford’s Felipe Jornada to define five open problems, among them a theory of cuprate superconductivity, a numerical method for predicting silicon and silicon dioxide interface structures, and an improvement to density functional theory. Before announcing the prizes, the company ran frontier models as autonomous agents with internet access for several days on each problem and saw little progress. Co-founder Advaith Sridhar says the problems are meant as goalposts for tracking whether models are improving in the field, whoever solves them. Source

24. PulseAI Gets FDA 510(k) Clearance for an Arrhythmia Model That Cuts Inconclusive Smartwatch ECG Readings

PulseAI. The Belfast company received FDA 510(k) clearance for its AI arrhythmia analysis model for wearables, which analyzes recordings in the cloud and returns structured results to connected apps for clinical review. In a prospective study of 296 participants published in Heart Rhythm, the model improved atrial fibrillation detection sensitivity and cut inconclusive interpretations by 92% compared with the Apple Watch’s built-in algorithm. PulseAI plans to integrate it into compatible ECG devices and software platforms. Source

25. FrontierPhysics: No Agent Setup Passes More Than 46 of 168 End-to-End Physics Research Runs

FrontierPhysics. BenchFlow and collaborators built 56 tasks from 42 physicists’ own research across nine theoretical, computational and experimental domains; each requires an agent to review the literature, plan and carry out research and experiments, and write a manuscript, graded with deterministic verifiers and expert rubrics for both correctness and publishability. Across 22 models in 26 configurations, no configuration passed more than 46 of its 168 rollouts. Manual review of the trajectories found the main weaknesses were insufficiently rigorous reasoning and limited expert intuition, not gaps in physics knowledge or execution errors. Source

26. D3D Decouples Diffusion Drafting From the Target Model for Asynchronous Speculative Decoding

Fudan University. D3D is a standalone diffusion drafter that does not need the target model’s hidden states, so it can draft the next block while the current one is being verified, in a single packed pass. Pruning 5 of its 36 layers keeps more than 99% of mean acceptance length, and across seven benchmarks it averages 5.00x and 5.71x decoding speedups with Qwen3-8B and Qwen3-32B as targets, against 4.03x and 3.67x for DFlash, reaching 7.48x on GSM8K with Qwen3-32B. The code is on GitHub. Source

27. Nace.AI’s Open Drex Decision Models Return Option Probabilities in 183 ms Without Decoding

Nace.AI. Drex is a family of four open-source decision models (8B, 9B, a 35B mixture-of-experts with 3B active, and the 27B Max) that take a state such as a document, record or agent context and return a probability distribution over user-supplied options, a yes or no judgment, or a score, without autoregressive generation. Trained with supervised fine-tuning and reinforcement learning from adjusted feedback, Max scores 67.52 on the public Decision Index 0.4.1 suite against 60.70 for Jev 1.14.0, ties d3-turbo and GPT-6 Sol on Decision Bench v4, and trails Claude Opus 5.5 by 1.6 points. Max answers a typical request in 183 ms, against about 2.0 seconds for GPT-6 Sol and 3.5 seconds for Claude Opus 5.5, and its answers stay stable when the options are reordered. Source