AI News: November 7, 2026

1. OpenAI Closes $11 Billion of a New Round at a $1.25 Trillion Valuation, Well Short of Its Original $30 Billion Goal

OpenAI. The company has closed an initial $11 billion of a round that values it at $1.25 trillion and could grow to about $20 billion in the coming weeks, Bloomberg reported on Friday, and CNBC confirmed the figure; in late September Bloomberg had reported that OpenAI was looking to raise $30 billion, at a $1.4 trillion valuation according to The Verge. CFO Sarah Friar told CNBC the round is ongoing, that “a good CFO always raises money when she can,” that OpenAI will keep investing in data centers despite the midterm results, and that the company is still looking at a 2027 listing. OpenAI raised $122 billion in March at an $852 billion valuation. Separately, the Financial Times reports that Anthropic still plans to list before Thanksgiving and expects to raise $80 billion to $100 billion at a valuation of $1.25 trillion to $1.75 trillion, while Bloomberg says some prospective IPO buyers see a valuation as low as $1.1 trillion. Source CNBC TechCrunch The Verge FT via Techmeme

2. PSI Says Its Research System, Running a Non-Public Anthropic Model, Proved the 18-Year-Old Aaronson-Ambainis Conjecture

Physical Superintelligence (PSI). The startup says a research system it built internally, running instances of an unreleased Anthropic model on PSI’s own infrastructure, had a complete proof of the 2008 conjecture on the third day of a run and the Lean proof assistant accepted it; the system then improved the result twice, and Lean accepted the final version. Aaronson and Ambainis had shown that if bounded low-degree polynomials always have an influential variable, quantum computers are at most polynomially faster than classical ones on yes-or-no questions about random inputs, so large speedups like Shor’s require structure in the input. PSI’s proof gives some variable an influence of at least c·v^(5/2)/d^6 (c is about 0.0025, v is the polynomial’s variance and d its degree), following a plan by Shachar Lovett and Jiapeng Zhang and replacing the inequality they left as a conjecture with a new one. Stanford’s Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan, who settled a special case in August, are reviewing the proof and writing a shorter exposition for experts, and called it “a huge result.” Source

3. Democrats Plan Bills on AI Testing, Liability and Data Center Power Costs, and Put Nvidia on Their Oversight List

U.S. Congress. When they take control of Congress in January, Democrats plan to introduce legislation to shield consumers from data-center-driven electricity costs, require independent testing of advanced AI systems and clarify when AI companies are liable for harm, lawmakers and aides told Reuters. Rep. Lori Trahan plans bills requiring independent experts to evaluate leading models against safety standards and making companies responsible when AI systems act harmfully, expected majority leader Chuck Schumer named data center energy costs as a priority, and Michigan Senator-elect Abdul El-Sayed campaigned on an independent AI regulator; any bill would face President Trump’s opposition, and overriding a veto takes two-thirds of both chambers. Bloomberg reports that Democrats, with Elizabeth Warren a leading contender to chair Senate Banking, are poised to step up oversight of Nvidia over chip exports and circular financing, and The Washington Post reports that Apple, Meta and Elon Musk’s companies are among early investigation targets. Source Bloomberg via Techmeme Washington Post via Techmeme

4. ING Puts About $500 Million of Data Center Loans Up for Sale as 40% of Planned US Capacity Sits in States Whose New Governors Want a Pause

Data centers. The Dutch bank is taking bids on 12 data center loans, including $25 million of debt behind Project Jupiter, the potentially 2.45 GW New Mexico campus Stack Infrastructure is developing for the OpenAI, Oracle and SoftBank Stargate venture, plus debt tied to Vantage and CyrusOne projects, Bisnow reports, citing Bloomberg; it is among the first auctions of a portfolio made up only of data center debt. Market research firm Cleanview estimates that about 40% of planned US data center capacity is in states whose newly elected governors have called for a construction pause, including Texas, where incoming governor Gina Hinojosa has pledged to halt new projects until protections are in place, and Senator Bernie Sanders has called for an immediate national moratorium, The Guardian reports. In Amsterdam on Friday, about 100 protesters entered the construction site of three 85-meter data center towers that Microsoft will occupy, a project planned in 2016 that is proceeding despite the city’s ban on new data centers until 2030. Source The Guardian The Verge

5. Microsoft Is Reportedly Switching Off Rented Oracle AI Capacity After Declining to Renew a 2023 Deal Built on Older Nvidia Chips

Microsoft and Oracle. Microsoft has been turning off computing capacity it rented from Oracle under a 2023 agreement to supply servers for its AI-powered Bing search, after deciding not to renew because the older Nvidia hardware no longer fit its specifications, Bloomberg reports, citing people familiar with the matter; Bloomberg could not determine the contract’s size, and Microsoft declined to comment. Oracle said demand for its AI infrastructure remains strong and that GPU capacity that came up for renewal in its fiscal first quarter, mostly at least four years old, was renewed or resold under long-term contracts at prices 20% above the previous agreements. Bloomberg says Microsoft has committed more than $80 billion to renting capacity from outside providers including CoreWeave, Nscale and Nebius, and has tightened oversight after frustration with their costs and delays. Source crypto.news

6. Crusoe Sells Its Wyoming Data Center Site to Google, Which Takes Over the Project Behind Its $17 Billion Campus

Crusoe. The Denver data center developer, last valued at $30.9 billion, said Google has acquired its site in southeast Wyoming, ending Crusoe’s role after site selection, government approvals and a hybrid power plan developed with Tallgrass Energy that combines fuel cells, grid power and natural gas turbines with carbon capture. Crusoe had tried and failed to sign Google as a customer for the project, and Google had raised concerns about its cost and construction timeline, TechCrunch reports, citing Bloomberg; the price was not disclosed, and Google has said it will invest $17 billion in the campus. Crusoe originally planned to scale the site to 10 GW, says it has more than 6 GW contracted and over 40 GW in its pipeline, and remains the developer of a Google campus near Amarillo, Texas. Source TechCrunch

7. Ramp Data Shows Business AI Spending Grew Just 1.5% Over August and September While Token Volumes Kept Rising

Ramp. Token spend on the corporate card platform rose only 1.5% across August and September after a 42% month-over-month jump in July, then picked up to 8.9% from September 23 to October 23, lead economist Ara Kharazian told TechCrunch. He attributes the slowdown to OpenAI and Anthropic price cuts, enterprise discounts and more cost-conscious usage, such as budgets, defaults that avoid expensive models and routers that move work to cheaper models or decision models like Jev, rather than to weaker demand. The top 1% of firms dipped slightly below the roughly $10,000 per employee per month they reached at their summer peak, while median firms spend $12 to $15 and the top 10% spend $600 to $1,000; TechCrunch notes Anthropic said in a filing that its price cuts for Fable 5.1 and Opus 5.5 reduced revenue growth by about $600 million. Source

8. China’s Ministry of State Security Recruited the Megaspeed Executive Close to Jensen Huang, The New York Times Reports

Nvidia. U.S. intelligence concluded that Alice Huang, a Megaspeed executive and Chinese citizen, was recruited by the MSS to gather information about American AI chips and potentially influence Nvidia to push for its most advanced chips to be made available to China, according to the Times; Megaspeed is an Nvidia customer, and she also has relationships with top executives in the AI chip industry, including at Gigabyte. The recruitment finding is new since the paper’s report last week on her relationship with Jensen Huang, which President Trump was briefed on last year without taking action, The Verge notes. The report follows Thursday’s request from three House committee chairs for Huang to attend a closed-door briefing on November 18. Source NYT via Techmeme

9. Axon Invests $1.2 Billion in Skydio and Becomes Its Exclusive Public Safety Distributor in the US and Canada

Axon. The cash-and-stock investment, a mix of primary capital and shares bought from existing holders, adds about 20% to Axon’s fully diluted stake in the autonomous drone maker at a $5.2 billion pre-money valuation, and is expected to close in the fourth quarter of 2026 and first quarter of 2027. Skydio serves more than 1,500 public safety customers, expects about $1 billion in total bookings this fiscal year, and says agencies using its drone-as-first-responder docks have cut priority response times by up to 65%, with drones arriving before officers on about 80% of calls. Axon puts the drone-as-first-responder market at about $13 billion in the US and Canada. Source

10. Chinese Video Model Startup Sand.ai Raises a $200 Million Series A+ Ahead of the Full MAGI-2 Release

Sand.ai. Cornerstone Capital and Wangsu Science & Technology co-led the Beijing company’s round, announced November 5, which values it above $1 billion and brings its fundraising this year to $300 million across three rounds, TechNode reports. The money goes to video model research, training and inference infrastructure, product development and international expansion, and Sand.ai plans to release the full version of MAGI-2, a larger mixture-of-experts video generation model, soon. Source

11. Artificial Analysis Launches VIALS-AA, a Scientific Image Reasoning Benchmark Led by Gemini 4 Carbon at 64%

Artificial Analysis. VIALS-AA is its implementation of Handshake AI’s VIALS benchmark: 161 tasks that ask models to read evidence in scientific figures such as gels and plasmid digests, each run three times in the open-source Stirrup harness with code execution and image tools, and counted only when all three repeats are correct. Gemini 4 Carbon (high) leads at 64%, answering all three repeats correctly on 103 tasks, two more than GPT-6 Astra (max), while Claude Sonnet 5.5 (max) scores 60% with 3.6 times as many tool calls per task as GPT-6.1 Sol (max) at 59%. Kimi K3.1 Pro (max) and GPT-6 Astra (max) each solve about three quarters of tasks at least once but score 45% and 63% under the all-three rule, and cost per task ranges from $0.02 for GPT-6 Luna (max) to $1.77 for Grok 4.7 (xhigh). Source

12. Nanbeige4.3-33B Matches Qwen3.8 27B on the Intelligence Index as an Apache 2.0 Model Small Enough for a Laptop

Nanbeige. The foundation model team at Chinese recruitment platform BOSS Zhipin released a dense 33B text-only reasoning model with open weights under Apache 2.0 and a 524K-token context window, which Artificial Analysis scores 20 on its Intelligence Index v5, level with Qwen3.8 27B (xhigh) and a new Pareto point for score against total parameters. It beats Qwen on Terminal-Bench 4.0 (30% against 6%) and the CritPt v2 hard subset (16% against 4%) but trails on FrontierCode 2.0 (11% against 15%) and HLE-Diamond (7% against 15%). It declines 71% of Omniscience questions, for an 18% hallucination rate against Qwen’s 29%, uses 158K output tokens per index task, 44% more than Qwen, and has no confirmed public API yet. Source

13. Tesla Will End Cybertruck Production in 2027 to Make Room for More Cybercab Robotaxis

Tesla. Elon Musk said Tesla will wind down Cybertruck production at its Texas Gigafactory around mid-2027 and use the space to build more two-seat Cybercabs, TechCrunch and the Financial Times report. Cybercabs have run in Austin for two months without major incidents, Tesla plans to expand them to Dallas later this month, and it also runs robotaxi Model Ys in Austin, Houston, Dallas and three Florida cities. The Cybertruck is the third vehicle Tesla has dropped for its AI and robotics push, after it ended the Model S and Model X earlier this year to free space for Optimus humanoid robots; it sold tens of thousands of units in total against Musk’s late-2023 prediction of about 250,000 a year. Source FT via Techmeme

14. Prosecutors Charge Three Foreign Nationals With Using AI Clones of CEOs and Lawyers to Steal More Than $129 Million

U.S. Attorney’s Office, SDNY. An indictment in the Southern District of New York charges Jacob Soall, Sharon Hazan and Guy Sheetrit, arrested in Spain in June and August and in the UK in September, over a scheme that ran from July 2024 to February 2026 and used AI to mimic executives’ voices and images to get accounting staff to wire money for supposedly confidential transactions, NBC News reports. Co-conspirators posed as attorneys at large international law firms, and the proceeds were converted to cryptocurrency. Victims included an investment bank in New York, hospital systems in Missouri and California, a Rhode Island bank, a UK fintech company, a Massachusetts mutual fund company and a Florida pet food company. Source

15. TSMC Is Weighing a Move Into 3D DRAM at Nvidia’s Request, The Information Reports

TSMC. The chipmaker is considering expanding into manufacturing advanced 3D DRAM at the request of Nvidia, its biggest customer, as the AI buildout drives memory chip shortages, The Information reports; Bloomberg and Reuters followed the report. TSMC already manufactures Nvidia’s GPUs and packages them with high-bandwidth memory bought from memory makers, so producing DRAM would be a new business for the foundry. Source

Brown University. Qile Jiang and George Em Karniadakis replace the long text chains of thought in LLM-guided evolutionary search with swarms of hundreds of low-cost agents per round that reason without decoding to text and share information by passing KV caches; a novelty filter keeps the most distinct ideas by measuring distances between the agents’ hidden states, and a summarizer agent combines them by receiving all the swarm’s caches. On 21 mathematical discovery and scientific computing problems, SwarmEvolve posts the highest average score and needs a median of 3 to 8 minutes to reach the best scores other frameworks reach in 60 minutes. On several problems, running on Qwen3 8B, it improves on solutions AlphaEvolve first found with frontier models. Source

17. ByteDance Seed Uses Sparse Autoencoders to Show Qwen3 MoE Experts Specialize Below the Domain Level

ByteDance Seed. Instead of counting how often a router picks each expert on domain-labeled data, the researchers train sparse autoencoders on a mixture-of-experts layer’s output, which lets each latent’s activation be split exactly across the experts. On layers 4, 30 and 46 of Qwen3-30B-A3B, contribution is far more concentrated than gating, most latents are driven by one or two experts, and auto-interpretation finds separate experts for Java and for front-end web code, or for geometry and for arithmetic and polynomials, while coarse domains take very unequal numbers of experts. The SAEs and code are on GitHub. Source

18. WeChat’s WeLM-Image Generates and Edits Images With About 5.4B Active Parameters Per Token on a Single GPU

WeLM-Image. Designed for single-GPU deployment in WeChat’s consumer apps, the unified generation, editing and multi-reference model processes text, reference images and target latents as one sequence on a mixture-of-transformers backbone with mixture-of-experts modules, storing 31.7B parameters and activating about 5.4B per target latent token. Trained up to native 1.5K resolution with rubric- and checklist-based rewards, the 1536-resolution model ranks first among the open-source models the authors compared on PRISM-Bench, Qwen-Image-Bench and UnicBench. The distilled WeLM-Image-Turbo produces a 1024x1024 image in 4.98 seconds on one NVIDIA H20 in eight steps. Source

19. Amazon Lets Shoppers Text Alexa to Search, Compare and Buy, After Blocking Meta’s Muse Agent in September

Amazon. Select U.S. customers are being invited to text Alexa for Shopping in natural language, or send a photo of an item or outfit, to search Amazon, add items to the cart and complete purchases, with list management, price tracking, web-wide search, product comparisons and order status carried over from earlier Alexa upgrades. Others can join a waitlist, and Amazon says the feature will reach more users in the coming weeks, ahead of the holiday shopping season. Source

20. Underdog Releases Woof, a 31-Million-Parameter Open Speech and Music Model That Runs Offline in a Browser

Underdog. The startup’s Woof model is 31.7 MB, generates both speech and music, runs on most devices without an internet connection and is open for anyone to download; Underdog demonstrated it with a browser game in which typed text is spoken back and a radio tuner generates music. It follows Bark, Underdog’s 167 MB on-device speech-to-text model, and a 27B model sized for laptops, and founder Sigil Wen says the goal is American-made open models that let people keep their data on their own devices. Source