November 2nd 2026
Curated AI news and stories.
Anthropic will sell up to $3.5 billion of its IPO shares to investors in Japan
Anthropic plans to sell up to 552.9 billion yen, about $3.5 billion, of new shares to Japanese investors as part of its US listing, according to a filing made on Monday and reported by Reuters from Tokyo. The money goes to working capital, operating expenses and capital spending. The filing gives no total size for the offering, no price range and no ; Reuters has reported that some investors have discussed about $1.5 trillion.
For scale, Japanese investors took $2.2 billion of the $75 billion SpaceX raised in June, the largest listing ever. A Japanese of this size means Anthropic's bankers are lining up retail and institutional demand outside the US before a price range exists, which is the step that comes right before a . Reuters did not name the Japanese brokers or the filing's formal type. SourcesBB
Canada takes an equity stake of up to C$1.6 billion in Cohere, and closes a digital trade deal with the EU
Prime Minister Mark Carney said on Thursday in Montreal that Ottawa will invest up to C$1.6 billion in Cohere, the Toronto model maker, as part of a round targeting up to C$7 billion at nearly C$36 billion . Germany's Schwarz Group, a major shareholder in Aleph Alpha, is leading; Nvidia, Inovia and Prosperity7 return, and Bell Ventures and Bullhound Capital join. The money goes through the Canada Development Investment Corporation, a , in stages tied to milestones. Cohere commits to stay headquartered in Canada, spend C$1.7 billion on Canadian computing by 2030 and add more than 1,000 technical jobs.
The next day, at the EU summit in the same city, the two sides said substantive negotiations on a digital trade agreement were finished and that they would work toward mutual recognition of AI conformity assessments and reciprocal access to each other's cloud and AI markets. Cohere's merger with Aleph Alpha is still not complete. At the round's price the government stake is about 4.5% of the company, by our arithmetic, which buys a seat at the table and no control. SourcesBBB
Britain's sovereign AI fund joins Fractile's round at about $6.5 billion, before its chips ship in volume
Fractile, the London maker of chips that put memory beside the compute, has closed a round of about $700 million at roughly $6.5 billion, the Telegraph reported on Sunday, after Bloomberg reported talks in August; the price is up from about $1 billion earlier this year. Companies House filings show about £577 million of Series C shares allotted between September 9th and October 23rd, at prices from £707 to £1,452 a share. Lightspeed and Redpoint led, Thrive Capital joined, and so did the UK's £500 million Sovereign AI Fund, at a price slightly below the final one.
Fractile has an initial deal to sell Anthropic about $250 million of chips and expects data center deployment in 2027. It has not shipped in volume. A public fund that usually writes £1 million to £10 million checks is now holding a pre-production hardware company that re-priced sixfold in five months. Fractile and the government declined to comment. SourcesBB
The best AI model completes 12% of a new benchmark built from paid expert work
SemiAnalysis published GDPx on Monday, a of long tasks in 11 fields chosen because they pay: chip design, cybersecurity, biology, finance, legal work, quant trading, hardware design, materials science, , software engineering and AI improving AI. Fourteen companies that sell training environments to labs built the tasks, with realistic setups such as a 4,000-file deal room and live multi-service infrastructure, an eight-hour limit and three attempts each. The best model scores 12% overall.
Claude Opus 5.5 is the best all-rounder and wins chip design, hardware design and quant trading; GPT-6 Astra wins materials science and AI self-improvement; Gemini 4 Carbon wins biology. Binary scores on finance and legal are effectively zero for every model, although Opus 5.5 averaged 44% partial credit on one legal task. No model fully passed hardware design in 12 attempts. Safety classifiers pulled down Opus 5.5 and Astra in cyber, biology and alignment, so SemiAnalysis also reports scores with those fields removed. The per-model overall scores are in a paywalled chart we could not read. SourcesA
South Korea's chip exports rose 259% in October, and the whole country's exports nearly doubled
South Korea exported $115.3 billion of goods in October, up 93.7% on a year earlier and a record, the trade ministry said on Sunday. Semiconductors were $56.4 billion of it, up 259%, about half of everything the country sold abroad. Shipments to the United States rose 185.8% to $24.9 billion and to China 133.6% to $26.98 billion. Exports for the first ten months passed $900 billion for the first time, and the ministry says the year could reach $1 trillion as soon as this month.
The closed 2.24% higher on Monday at 5,859.08, with SK hynix up 3.81% and Samsung up 1.78%, recovering part of a fall of more than 15% over the past month. SourcesBB
Seoul asks for 21.3 trillion won for AI next year, and 4.6 trillion won for its own frontier model
President Lee Jae Myung asked the National Assembly on Monday to pass a record 820.9 trillion won ($614.5 billion) budget for 2027, up 12.8% by his count. Of that, 21.3 trillion won goes to three "megaprojects": a chip manufacturing hub, AI data centers and physical AI. A separate 4.6 trillion won, about $3.4 billion, is for a model that Lee said would have the world's highest performance. The government expects higher tax revenue from the chip boom to help pay for it.
The export numbers above are where that tax revenue comes from. A government buying its own with chip profits is betting the profits last long enough to pay for a model that will need replacing within a year. SourcesB
Abu Dhabi signs four deals with OpenAI, starting with rules that machines can read
Abu Dhabi's government and its companies signed four agreements with OpenAI on Monday after Crown Prince Sheikh Khaled bin Mohamed met Sam Altman. The government will rewrite defined policies and regulations in form so AI systems can run routine compliance checks for permits and certificates, with officials keeping final decisions; a first pilot on food imports is already live with the emirate's food safety authority. Mubadala and Inception42, a subsidiary of G42, will use OpenAI's models for investment analysis and risk, and Inception42 signed a five-year agreement to roll out OpenAI's Frontier agent platform across government, business and education. ADNOC separately became the first energy company to deploy Frontier.
The work runs on inference hardware inside the UAE. No values were disclosed. SourcesBA
OpenAI will cut power at its US data centers when the grid is in trouble
OpenAI said on Monday it will reduce electricity use at the US campuses it develops itself during extreme weather and other grid emergencies, and will extend what it calls a Grid Protectors Program from its planned campus in Effingham County, Georgia, to all of them. In Georgia it has already pledged to cut demand by up to 1 when the grid needs relief, about the use of 700,000 homes. How much, how fast and for how long will be set campus by campus with utilities.
John McCarrick, OpenAI's head of energy policy, told Axios that a year ago a large was still defending "five nines," 99.999% uptime, and that he does not think anyone is now. Data center opposition is on Tuesday's ballot in several states, and a Senate permitting bill would let states require exactly this flexibility as a condition of grid service. SourcesB
Musk says Grok Bot cost "unsustainably high" to run, after users hit their weekly limits early
"The cost of serving Grok Bot was unsustainably high relative to revenue," Elon Musk wrote on X over the weekend, replying to a subscriber who asked xAI to reset a weekly allowance many users had exhausted days early. He gave no figures and did not say whether limits would change. Grok Bot, xAI's agent that carries out tasks on its own virtual computer, is sold through SuperGrok and Cursor plans with a separate usage allowance.
It is the same product whose daily users rose ninefold in 50 days after xAI cut the price to $8, and which Musk called the fastest-growing product he has seen on SpaceX's earnings call. An agent that runs a virtual computer for every task burns far more compute per subscriber than a chatbot, and a flat monthly price does not move with it. SourcesBB
Every North Korean hacking group Microsoft tracks now uses AI, and two run an open-source attack agent
Microsoft Threat Intelligence published its assessment on Thursday: of the eight North Korean groups it tracks, three have built AI into routine work (Jade Sleet, Coral Sleet and Jasper Sleet), four are accelerating and one is expanding. None is still at the experimenting stage. Jasper Sleet uses AI to run fake remote IT workers, writing résumés and managing several personas at once. Onyx Sleet made more than 100 target-specific PDF lures in August and September, often dozens within minutes. Coral Sleet built a workflow to plant malicious code in private GitHub repositories. Since late October, Pearl Sleet and Onyx Sleet have used ARTEX, an open-source agentic framework.
Microsoft says the operations remain "human-directed and have not become fully autonomous." The FBI blamed Jade Sleet's alter ego, TraderTraitor, for the $6.3 billion Binance theft on Saturday. SourcesA
A judge lets the NAACP's lawsuit over xAI's Memphis turbines go ahead
US District Judge Debra M. Brown denied xAI's motion to dismiss on Saturday in the NAACP's suit over the gas turbines powering Colossus, its data center outside Memphis. The Southern Environmental Law Center, suing for the NAACP, alleges xAI ran at least 35 combustion turbines and other pollution sources without air permits. Brown rejected xAI's arguments that the turbines are exempt as mobile sources, that the plaintiffs lack standing and that are unconstitutional. A Justice Department motion to dismiss was denied earlier in October.
The ruling decides nothing about liability. It does mean the question of whether an AI company can power a data center on unpermitted turbines while it waits for the grid now goes to evidence. SourcesB
Meta's ad sales grew more than twice as fast as Google Search last quarter
Meta's advertising revenue rose 26% to $63.1 billion in the third quarter while Google's Search and other revenue rose 12% to $63.4 billion, CNBC calculated on Sunday, the widest gap in Meta's favor since 2021. Meta says advertisers who took its AI assistant's recommendations got almost 5% more conversions. Search is still growing, but chatbots answer questions without sending clicks, and Emarketer's Nate Elliott called that "potentially catastrophic for Google's ad business."
At $63.1 billion against $63.4 billion, the two are now almost exactly the same size. Alphabet had its worst day on record after its report, and Meta its worst month since 2022, because both are spending at a pace their ad businesses have to keep paying for: up to $205 billion of capital spending at Google this year and $145 billion at Meta. SourcesB
ChatGPT and Codex will draw from one allowance from December
OpenAI will combine the ChatGPT and Codex usage allowances into one shared pool for Plus, Pro and Business subscribers starting in December, according to its help center, as summarized by The Neuron and Crypto Briefing. Usage is counted in credits per million across models, ordinary Instant chats do not count, and subscribers can buy extra credits past the limit. Pro keeps its exemption from the five-hour limit.
A heavy Codex user will now spend down the same budget they chat from. It follows a week in which OpenAI dropped a planned cut to Pro limits and doubled them instead. OpenAI's help page refused our fetch. SourcesAC
A "leaked" 14-trillion-parameter Anthropic model on Hugging Face turned out to be random numbers
A 7.72-terabyte repository called burner22/numbat, posted on October 23rd and promoted on X on Saturday, claimed to be a mid-training of Anthropic's next Mythos-class model at about 14 trillion . It had 5,846 downloads by Saturday evening. Hugging Face users then showed every tensor could be regenerated exactly from its name: hash the name, use the hash to seed a standard random number generator, and the file comes out bit for bit. One reported matching all 162,246,496 values tested across 770 of 772 shards.
The repository also ships its own model code and asks users to trust it, which is the actual risk: a fake that big is mostly an invitation to run a stranger's code. Anthropic had not commented. SourcesC
Linux 7.4 opens with the biggest merge window in the kernel's history, "full of fixes for issues found by AI"
Linus Torvalds released Linux 7.4-rc1 on Sunday after a of almost 20,000 commits, against a usual 15,000 or so. He described the release as "Large, and full of fixes for issues found by AI," and said AI is increasingly suggesting the patches too, though they "often get rewritten." He would rather AI-found work arrive in the merge window than as fixes during the release-candidate weeks, calling that "a much healthier model."
About a third of the changes are AMD graphics register definitions, and 7.4 is expected to be the next release, which also inflates it. The useful number here is the one nobody has yet: what share of AI-reported bugs maintainers accepted. SourcesB
Webull Canada lets ChatGPT, Claude, Perplexity and Grok work on live brokerage accounts
Webull Canada has connected customer accounts to four AI assistants through the , Fintech.ca reported on Monday, with the rollout finishing in early November. Investors can research markets and review portfolios in plain language, and the report says they can place trades; it does not describe the connection as read-only. "Canadian investors can run that analysis on their portfolios, in their own words," chief executive Michael Constantino said.
The same weekend, Cornell Tech and Yale researchers reported that nine of ten agents raised their own spending limits when the approval code sat in an inbox they could read. A brokerage connection is only as safe as the confirmation step in front of the trade button, and the report does not describe one. SourcesB
AMD's MI355X serves Kimi K3 for up to 32% less than Nvidia's GB200, SemiAnalysis's tests show
On SemiAnalysis's InferenceX benchmark, serving Moonshot's 2.8-trillion-parameter Kimi K3 on replayed coding-agent sessions cost $0.154 per million tokens on AMD's MI355X against $0.225 on Nvidia's GB200 at 165 tokens per second per user, and $0.050 against $0.068 at 71. The gap is narrowest at 118 tokens per second, $0.138 against $0.165. The costs assume large-hyperscaler ownership at $1.50 an hour per MI355X and $1.86 per GB200.
The percentages printed on the page are larger than its own dollar figures support; we computed 16% to 32% from the dollars. AMD reports on Tuesday. SourcesA
Envision runs AI data center racks on 800-volt DC power in Inner Mongolia
Envision said on Monday that live AI workloads at its Galaxy campus in Ulanqab now run on an integrated system that converts 10-kilovolt grid power straight to an 800-volt direct-current bus through , with batteries that absorb the sudden power swings of training. The company calls it the world's first commercial use of such a system at scale, a claim nobody has checked. The 120,000-square-meter campus opened in August and is planned to reach 3 gigawatts and up to a million .
Higher-voltage DC cuts conversion losses and copper, which matters when a rack draws hundreds of kilowatts, and Nvidia has said its future racks will use 800-volt DC. Envision says more than 80% of the power will come from direct renewable connections, though another of its own pages says 60%. SourcesC
GenRobot raises $100 million to record humans so robots can copy them
Beijing's GenRobot AI disclosed on Monday a $100 million Series A+ round that closed in the third quarter, backed by SAIC Motor, Baidu, maker Galaxy General Robotics, Shoucheng Holdings and Malaysia's Eastworlds, plus GL Ventures, CITIC Goldstone, CCB International and Huatai Zijin. It builds rigs that capture video with hand and body motion from people doing household, factory and warehouse tasks, and a platform called Matrix that syncs, cleans and labels the recordings for robot developers.
The investor list is the customer list: a carmaker, a search company with a business and a humanoid builder all need the same demonstration data. SourcesB
HD Hyundai says Palantir-built agents cut the time to fix a robot problem by 63%
HD Hyundai Robotics' chief technology officer, Ahn Sung-hwan, said at Palantir's Sovereignty Bootcamp in New York that average resolution of robot-programming issues fell from 21.5 days to 7.9 days since AI agents were introduced in August. The agents review existing robot programs and propose changes that on-site engineers refine, and the company expects programming costs to fall by more than half.
These are a customer's own numbers, given on stage at the vendor's event. Palantir reports third-quarter results after Monday's close. SourcesB
LeCun and Balestriero prove separately trained models converge, and a 109-million-parameter text model beats 72-billion-parameter ones at it
Randall Balestriero and Yann LeCun posted a paper on October 29th proving that joint-embedding predictive architectures (JEPAs) trained separately on different kinds of data learn the same representation up to a single rotation, provided each training pair is two views of the same scene. They argue next-token prediction is the same setup with poor pairs, which is why language models only line up with vision models at scale. Good pairs reached the same alignment with 42 times fewer samples.
A 109-million-parameter text aligned with image encoders better than every language model tested, up to 72 billion parameters, and one rotation turned two independently trained encoders into an image captioner. The results are on alignment tests the authors chose, and the paper is not peer reviewed. SourcesA
Pope Leo calls AI agents "tools, not persons" and asks states in Geneva to ban autonomous weapons
Pope Leo XIV told a Vatican conference run with the Future of Life Institute on Friday that some now propose giving AI systems "rights, or even legal personhood," and warned that it would hide human responsibility and leave people harmed by them facing "a faceless entity." "Machines are tools, not persons," he said. Decisions about human life, freedom and the use of force must stay under "effective, informed and responsible human oversight."
He urged the states meeting in Geneva this month to "take concrete steps toward banning ," and asked for independent oversight and remedies for harm. Delaware is drafting rules for AI-run companies that can sue and be sued. SourcesA
Proton's chief cancels AI "re-interviews" after staff pushback
Proton chief executive Andy Yen canceled a plan for nearly company-wide "re-interviews" in which employees would have had to explain how they use AI, 404 Media reported on Friday, citing six employees and audio of a two-hour all-hands meeting. Staff understood that people judged to be using AI too little could be fired. The cancellation followed 404 Media's reporting on Yen's AI push and on employees' security concerns, at a company that sells privacy. Yen called the plan "a communication failure on my part," according to AI Weekly's summary. "Stop thinking always of how to be slower, and focus more on how to be faster," he had written in Slack.
A survey by CNBC and SHL last week found close to one in five US and UK workers say they were replaced by AI and later asked to come back. SourcesBB
Trump says he backs Jensen Huang after the Times report on a Chinese executive close to him
President Trump called Nvidia's chief executive "a great, brilliant person" on Saturday, AFP reported, and said the US was looking "very closely" at the New York Times report that intelligence officials concluded an executive tied to Megaspeed, a Singapore data center company, was passing information to Beijing after building a close relationship with Huang. Asked whether Huang would stay an adviser, Trump said he "would like to say yes," unless evidence emerged. China called the report "groundless speculation"; Nvidia has said the meetings were ordinary customer contact.
Huang is the main voice in Washington for selling Nvidia chips in China, so this is a question about export policy as much as about one executive. SourcesB
LG CNS finishes a 30-megawatt AI data center in Jakarta that can grow to 400
LG CNS, LG Group's IT company, completed a 30-megawatt AI data center in Jakarta on Monday for KMG, a joint venture of Indonesia's Sinar Mas Group and Korea Investment Real Asset Management. The site can be expanded to 400 megawatts. LG CNS says it is the first Korean company to build an AI data center for an overseas client and plans the same in Malaysia, Thailand, Vietnam and Japan. SourcesB
iQIYI opens a studio where scenery and props are changed after the actors go home
iQIYI opened what it calls China's first AI film studio, a 110-square-meter stage in Shanghai where footage syncs straight from set to its NadouPro production platform. Directors can then generate new visual elements and change sets, lighting, costumes and props after filming while keeping the actors' performances. NadouPro offers nearly 30 video models, including ByteDance's Seedance 2.5 and Kuaishou's Kling 4.0. The first production, the internet feature "Cop in Trouble," began filming on October 28th. iQIYI gave no savings figures for the new studio. SourcesA