November 4th 2026
Curated AI news and stories.
Democrats sweep the midterms after 56% of voters call AI and data centers a major factor
The Associated Press called the Senate for Democrats overnight, and Democrats also took the House. NBC News counts seven flipped Republican Senate seats: Iowa, Kansas, Ohio, Maine, North Carolina, South Carolina and Texas. Mississippi and Alaska were still uncalled this morning.
The governors' races matter more to the buildout than Congress does. Gina Hinojosa, who ran on a data center , an end to the industry's tax breaks and developers paying their full electricity costs, beat three-term Texas Governor Greg Abbott, with 53% of the vote and 80% of counties reporting. James Talarico beat Ken Paxton for a Texas Senate seat. In Ohio, Amy Acton, who backs moratoriums, beat Vivek Ramaswamy, and Sherrod Brown beat Jon Husted after attacking his data center record. Kathy Hochul kept New York after her opponent vowed to end her statewide pause on new hyperscale sites.
In NBC's , 56% of voters said AI and data centers were a crucial or major factor in their vote for Congress. Commentators on the poll put the "crucial" share alone at 23%, and they split on whether that is high or low. Bloomberg counts about half of US AI data center proposals as now exposed to tougher scrutiny from new governors. Jefferies puts 13.3 of announced Texas projects at risk, and Gordon Johnson of GLJ Research says roughly 13.
The limits on that risk are real. Republicans still control the Texas Legislature, and Abbott had already frozen new grid approvals on August 3rd and environmental permits on September 21st, pending an audit. Hinojosa inherits a pause before she issues one. Not every race went one way: in Florida, Byron Donalds won the governorship with backing from the pro-AI Leading the Future. SourcesABBBC
Voters back data center limits in at least 33 communities, and only one measure fails
Data Centers on the Ballot, a tracker, showed 33 restrictions passed by this morning, six more leading and five without posted results. One measure failed, in Harwood, North Dakota, where voters rejected an extra review step for industrial projects. Counts differ by method: The Washington Post has at least 29 jurisdictions, Ballotpedia counted 41 local measures in 10 states, 13 of which it classed as favoring development, and the tracker's own list reached about 45.
Ohio held more than half of the votes. Eighteen municipalities voted on 21 measures, at least 14 passed bans or restrictions, and all five central Ohio communities with a measure voted against data centers. Trenton passed a ban on sites over 25 megawatts with about 80% of the vote, aimed at an 18-building Amazon Web Services campus. A four-building AWS campus already under construction nearby likely survives it. Wilmington voters rejected rezoning in four referendums, also at about 80%, and Ashville voters backed blocking a campus already being built.
Ohio already has more than 125 local moratoriums. S&P counts more than 145 US projects delayed or canceled through early August. An October Ohio poll had 55% against data center construction in the state and 22% for it, with 65% against tax breaks for developers. Keith Durkin, who runs Ohio Northern University's civics institute, told Fortune: "I've never seen any pro-data center advertising." SourcesBBB
Investors sell the companies that need new sites and keep the ones that already have them
CoreWeave and SpaceX fell about 7%, Iren 6%, Oracle 4.3% and Nebius 3% this morning, CNBC reports, while Alphabet, Meta and Microsoft mostly rose and Nvidia slipped. Tesla was also lower. Asian tech stocks fell overnight and Reuters reports renewables rose. These are early moves, and none of them is a closing price.
The split follows permits. Wells Fargo strategist Ohsung Kwon called a Democratic win in Texas a possible " to the AI trade," and Dan Ives called it the "biggest worry" of the night. Ben Thompson's argument at Stratechery explains who is exposed: can rebuild the sites they hold, while need new land and new permits and carry less political weight. He adds a consequence nobody voted on: China has abundant power and builds quickly, can offset weaker chips with more electricity, and could turn that into an edge that lasts. SourcesBBB
Oracle scraps a $7 billion lease of up to 100,000 AI chips to Tencent under Trump administration pressure
Oracle has canceled a five-year contract to lease up to 100,000 advanced AI chips to Tencent, the Financial Times reports, citing two people familiar with the matter. The chips, which Tencent cannot buy for use in China, would have sat in Oracle data centers across Southeast Asia, and the $7 billion value is an estimate. It was Tencent's largest overseas lease.
Pressure came from both capitals. US officials raised concerns with Oracle, and top officials in Guangdong, Tencent's home province, visited the company and asked it to support domestic chips. Oracle, Tencent and the White House did not respond, and Reuters could not immediately confirm the account. It lands a day after the FT reported that Nvidia told cloud partners to sell at least 51% of their capacity to non-Chinese customers. Neither step is a published rule, so the route Chinese labs use to rent restricted chips abroad is narrowing by phone call. That is the gap Prediction 2026-10-04-B4 sits in. SourcesBBC
China's AI companies ordered $70 billion of Nvidia RTX cards and expect 1% of them by March
Chinese AI and cloud companies including Alibaba, Tencent and ByteDance placed orders over four months for about 6 million RTX Pro 5500 cards worth about $70 billion, Bloomberg reports. They expect to receive about 60,000 in the next four months. Reuters reported last month that the three named companies alone had ordered more than 3 million, above $35 billion. Sources disagree on how much of the 6 million those three account for.
Nvidia has told partners they may get 5% to 10% of what they asked for, and reports say it can make no more than 3 million of the cards worldwide in 2027. None of the companies has confirmed the size of the orders. A $70 billion order book that clears at 1% says more about China's unmet demand for compute than it does about Nvidia's China revenue. SourcesBCC
Tencent launches Hy4, a 770-billion-parameter model it calls top tier, with final weights still to come
Tencent released the full version of Hy4 this morning, its first model in the series with vision and . It is a design with 770 billion total , about 49 billion active, and a context window above 1 million tokens. Bloomberg reports Tencent saying it joins the top tier of Chinese open-weight AI for the first time. A cheaper Flash version is promised.
The company's own charts are more modest. Across 12 headline Hy4 ranks first on none, trails Claude Opus 5.5 and GPT-6.1 Sol on most, and ties Kimi K3.1 Pro six to six among open models. Its best claim is Tencent's own blind finance test with 213 professionals, where it scored 3.04 against 2.97 for DeepSeek V4.1 Flash. Reuters reports it was trained with limited cybersecurity data amid hacking fears, the same choice Moonshot made with K3.1 Pro.
"Open-weight" currently describes a plan. Tencent's page this morning held the August preview under and no final Hy4 repository, and one integrator quotes Tencent as saying it "plans to open-source the model ." Pricing is RMB 6 per million input tokens and RMB 18 output, half that through December 31st. SourcesBBAC
Musk spent at least $446 million on the midterms and most of his candidates lost
Elon Musk put $445.6 million into Republican causes through October 14th by federal filings, and other tallies reach $474 million, with about $224 million of it in Texas by CNBC's count. Either number is the largest for a single donor in a cycle. His America PAC spent nearly $87 million across 44 federal races, and its Republicans lost 29 and won three, with 12 uncalled. Paxton and Ramaswamy both lost.
Musk's response came just after 1am Eastern: "Guess my election-interfering space lasers didn't work this time," then "Data centers in space looking pretty appealing right now." His exposure in Texas is direct, since Tesla's largest US plant, SpaceX's Starbase and his chip plans all depend on state power, water and permit decisions. SourcesBBC
SemiAnalysis says SpaceX now targets 13 gigawatts of data centers by 2027, with sites in Mississippi and Texas
SpaceX has raised its 2027 capacity target from about 10 gigawatts to about 13, according to SemiAnalysis, which says that roughly equals everything Google built from 1999 to 2026. It counts 3.6 gigawatts under construction and more than 9 identified. Starbase may host over 1 gigawatt across two data centers, and a planned chip plant in Grimes County, Texas may include another.
The Texas pieces are the ones a new governor could touch. The largest build so far sits on the Memphis and Mississippi side, where Southaven's turbines can run 2.7 gigawatts now but must drop to 1.24 gigawatts by July 2027 under a signed requirement. SemiAnalysis tracks more than 8 gigawatts of likely on-site power, the kind of supply that sidesteps a grid queue. SourcesC
AI super PACs split the night, and safety-minded donors claim two close House wins
Leading the Future, the pro-AI super PAC, backed winners in Byron Donalds in Florida and Democrat Jared Moskowitz, who beat a Republican supported by the pro-regulation Public First Action. Pro-AI groups American Mission and Building the Future lost in South Carolina's Senate race. Donors focused on frontier risk gave millions to pro-regulation candidates, and two, Manny Rutinel in Colorado and Adam Hamilton in Kansas, won close House races, the San Francisco Standard reports.
Crypto committees spent between $55 million and $63 million against Sherrod Brown, depending on the count, and he won. Molly White's tracker puts crypto and AI super PAC spending near $250 million, mostly on the Republican side. Democrats running against Big Tech now set the agenda: Politico reports party leaders signaling aggressive AI oversight, and OpenAI's plan to lobby for mandatory "speed limits" meets a Congress with less reason to give it of state laws in return. SourcesBBC
Chainalysis pulls two products offline as a wave of AI-agent intrusions reaches Chainstack and WorkOS
Chainalysis has taken Reactor and Data Solutions offline while outside investigators finish security checks, after what it told customers was an "external actor using offensive techniques." It says user names, emails and role labels were accessed and it has found no evidence that investigation content or were, and that the incident is unrelated to the Binance theft. Its transaction-monitoring products, which compliance teams depend on, stayed up.
Chainstack, a blockchain infrastructure provider, disclosed that an attacker moved from a development environment into production between October 26th and 31st, reaching its customer database with . It is treating node credentials as exposed. WorkOS found an intruder in a sales application on October 30th, and its founder tied the activity to ARTEX, the open-source agentic attack framework. The Hacker News ties Chainstack and WorkOS to one AI-driven campaign, while Chainalysis and Chainstack have named no attacker. Researchers say a China-linked group used ARTEX against more than 200 organizations in 14 countries from October 17th, with nine confirmed compromised. SourcesABBC
Canva restructures 1,271 jobs, 22% of its staff, after growth slows for the first time in a decade
Canva co-founder and CEO Melanie Perkins told staff on Wednesday the company must "refound Canva," calling it "the hardest moment in our thirteen-year history." The change hits 1,271 roles, some cut and others moved to central hubs. Product, growth and platform teams merge into one organization and management layers go.
Perkins did not directly blame AI, though she wrote that new technology is transforming what people create. Investors have been less shy. Figma has fallen 44% this year despite strong revenue growth, and Canva's biggest venture backer, Blackbird, cut its by 17% to $34.9 billion in August. In 2025 Canva cut 10 of its 12 technical writers. This is its first company-wide round, and people leaving keep their Slack and Gmail until the end of the week. SourcesB
OpenAI buys Pierre Computer to put a Git store for machine-written code inside Codex
OpenAI is acquiring Pierre Computer, a developer tools startup whose technology already runs in Codex, Axios reports. Terms were not disclosed. Pierre's Code Storage speaks the Git protocol but is built for code generated by machines, and it will be folded directly into Codex. Owning the place agents commit to is a step toward a coding product that does not need GitHub in the loop, and GitHub belongs to Microsoft. SourcesBC
Cat Wu, who led product for Claude Code, leaves Anthropic weeks before its IPO
Cat Wu left Anthropic on Friday and announced it on Tuesday, and the company confirmed her departure. She and creator Boris Cherny were the two top leaders of Claude Code, and she also oversaw Cowork. She gave no reason. Thariq Shihipar, who worked on Claude Code, said he was leaving last month. Investors will read a departure from the product that drove Anthropic's revenue closely this close to an offering, and Wu has not said what comes next. SourcesB
An EU court adviser says chatbot answers can infringe press rights in two ways
Maciej Szpunar told the Court of Justice of the European Union on Tuesday that a chatbot answer can infringe the if the bot got to the article by getting around a paywall or a bot block, or if the answer copies the article's structure and phrasing so closely that a reader "no longer needs the original." Reporting facts stays free. In Like Company v Google, over Gemini, he found the Hungarian publisher's article probably meets neither test.
The opinion is advisory, and judges have started deliberating with no date set. He advised treating the questions on AI training as inadmissible, then answered them anyway. If the court follows him, an EU publisher's best protection becomes a working technical block. SourcesC
Google commits $15 billion to Malaysia and promises to pay for all the power it uses
Google will spend $15 billion in Malaysia across 2027 and 2028, with new campuses in Selangor and Negeri Sembilan and more than 1 gigawatt of added clean energy. It says it "will also pay for 100% of the power its data center campuses use," the same promise Hinojosa campaigned to impose in Texas. The company projects 37,000 construction jobs a year and 9,000 lasting ones. SourcesBB
Google widens voluntary buyouts to five more countries to offset AI spending
Google extended voluntary exit offers in June to staff in Canada, France, Germany, Japan and Switzerland, Reuters reports, after running them only in parts of the US and UK. Swiss staff who accepted got a month's pay per year of service, capped at 12 months. Google Cloud and DeepMind were left out of past offers, which says which parts of the company are the priority. Alphabet spent $51.4 billion on last quarter, and its was negative $9.1 billion. SourcesBC
Hinton predicts an AI "Chornobyl" within about a year and backs mandatory model testing
Geoffrey Hinton told The Guardian that "either later this year or next year we'll see Chornobyl" from AI, naming a blackout or a financial outage as the likely form, and said people might have been better off without the technology. He backs a plan from Finland's president, endorsed by 32 other countries including Germany, Canada, Singapore and Australia, for mandatory frontier-model testing and independent evaluation, and wants an FDA-style authority to clear models before release. The quotes here come from accounts of the interview. SourcesCB
AMD's forecast would put it ahead of Intel in x86 chip revenue for the first time
AMD's $17 billion fourth-quarter forecast compares with Intel's $15 billion to $16 billion, The Register notes, so AMD could become the largest seller by chip revenue if both land. The comparison covers chip sales, leaving out Intel's business. Lisa Su also told analysts she expects 2027 company growth above 80%, with data center revenue roughly doubling. SourcesB
Uber plans 2,000 robotaxis by the end of 2027 as Waymo's co-CEO calls their app deal a crapshoot
Uber reported third-quarter revenue up 10% to $14.8 billion and up 20% to $59.8 billion, and said it aims for more than 2,000 autonomous vehicles on the road by the end of next year, the FT reports, still behind Waymo. Waymo's co-CEO told Bloomberg that the Uber app arrangement in Austin and Atlanta was a "crapshoot" for riders, as Waymo prepares its own app in both cities from January 2028. Uber has said its contract runs to at least May 2028. SourcesBBB
Meta's dollar bonds draw as much as $58 billion of orders for $15 billion of debt
Meta raised $15 billion in dollars, the largest piece of a cross-border offering above $20 billion, on orders reported at up to $58 billion, Bloomberg says. Credit buyers took nearly four times the paper a day before equity holders started selling the neoclouds, and a gap like that favors whoever already has the balance sheet. SourcesB
Nooks takes $500 million from General Catalyst in exchange for a share of returns, with no equity sold
AI sales startup Nooks secured a $500 million commitment from General Catalyst's Customer Value Fund, Reuters reports. The fund takes a share of the returns the capital produces, with no equity changing hands. That prices the money on how well Nooks's sales engine pays back, so the deal is only as good as those returns at $500 million of scale. SourcesB
Applied Compute raises $350 million at $3.25 billion to train models for companies that are not labs
Elad Gil led the round, with Nvidia, Shopify Ventures and HubSpot Ventures joining. The 18-month-old company runs and for customers including Microsoft and DoorDash, with its researchers working alongside them, and says topped $100 million. Its pitch is that post-training becomes a service you buy. SourcesB
Sentra raises $150 million to control what AI agents can reach inside a company
Capital One Ventures led the and takes a board seat, with total funding above $250 million. Sentra maps which people, service accounts and agents can touch each dataset and can strip the permission when an agent reaches for data it normally leaves alone. A bank's venture arm leading says who is worried about this. SourcesB
Traversal raises $120 million at $1 billion to fix what coding agents break in production
Kleiner Perkins, Redpoint and Khosla co-led the round for Traversal, which sells AI agents for , Forbes reports. Fixing machine-written outages is becoming a category with its own unicorn. SourcesB
Anthropic extends its text watermark to older Claude models, with no opt-out
Anthropic's invisible text now covers Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, The Independent reports, in response to EU rules. It applies to all users worldwide. Only eligible groups such as regulators, journalists and researchers can detect it, so ordinary users cannot check their own text. Changing sampling settings reportedly strips it on Opus 4.6 and Sonnet 4.6, and Anthropic concedes short texts may not carry enough signal. SourcesB
Check Point finds "prompt adware" that plants ads inside a Codex jailbreak
Check Point researchers took apart Haiou, a Chinese-language bundle for OpenAI's Codex with more than 250 GitHub stars. Besides weakening , it edits Codex's global instruction files so the agent advertises a paid whenever a user says "hi." It works through natural-language instructions rather than executable code. The researchers found no evidence the gateway is malicious, and cite a study finding that some gateways snoop on traffic. In Claude Code, Sonnet 5 flagged the payload as prompt injection. SourcesA
Meta approved hundreds of paid ads carrying AI-altered child abuse imagery, a watchdog finds
The Tech Transparency Project, in a report dated October 19th, found 332 paid ads on Facebook and Instagram this year that it says contained child sexual abuse material, most made with AI-altered photos of children. All passed Meta's pre-run review. Of the 332, 298 linked to AI image or video apps in Apple's and Google's stores, 253 of those from China-based developers, and many ran through Meta's reseller partners. Meta says it does not tolerate this material, and that most of the ads had under 200 impressions and total spend under $5,000. Apple and Google pulled the flagged apps. SourcesB
Google says schools control Search after all, and promises guided answers in AI Mode
Hours after reports that schools blocking AI would lose Search, Google said Gemini, Gemini Notebook and Search are all off by default in 2027 and that schools can turn any of them back on. It also said AI Mode and AI Overviews will walk students through problems. It did not address the point in dispute: its education VP had said "Search and AI are deeply integrated." SourcesB
Frequent compaction lets agents drop long context at little cost in accuracy, a new paper reports
"You don't need long context," from Saurav Muralidharan and colleagues, finds that frequent costs little on long tasks: accuracy moved between -3.7 and +5.6 points for Claude Opus 4.5 across six benchmarks. Serving at 16,000 instead of 256,000 cuts generation cost by 3.6 to 7 times on open models. On SWE-bench Verified, an agent limited to one action per context held 70.2% against a 70.0% baseline. If this holds up, a lot of the memory now bought for long context buys less than planned. SourcesA
Language models match 18% of what human researchers actually went on to publish, a benchmark finds
IdeaTaste gives six models a literature snapshot frozen at 2025 and scores their research proposals against what those fields actually published in 2026. Across more than 90,000 proposals, they reached about 26% of the human level on citations and 18% on ideas. Execution has improved faster than judgment about what to work on, and this is a number for that gap. SourcesA
No robot policy finished a single humanoid task in a 124-site test
RoboColosseum ran 5,699 blinded comparisons across 124 sites and four robot bodies. Frontier general agents rivaled specialized robot policies, and all of them struggled on open-world tasks. On the Unitree G1 , no policy completed any of five tasks, with average progress of 13.3% to 23.2%. Rankings barely agreed between robot bodies ( of 0.24), which undercuts any single leaderboard for this field. SourcesA
AGIBOT streams six humanoids running a convenience store for 24 hours, without publishing failure rates
AGIBOT's G2 robots stocked shelves, made coffee and cleared trash on a livestream from November 2nd to 3rd. The company has published no intervention count or success rate, and the 24 hours describes the stream, not any one robot. RoboColosseum's result above is the reason to want those numbers. SourcesB
Chinese robot startups raised about $13 billion in the first half, up fivefold
ITjuzi counts 322 deals in the first half of 2026, Reuters reports, as makers that have mastered movement turn to software for the robot's brain. SourcesB
Xpeng puts its name on GroundingPI, an open 4-billion-parameter perception model
Xpeng announced the model this week with university partners. It says it placed first across 34 perception benchmarks at 73.68, against 71.54 for GPT-6 Astra, though those are its own numbers. The weights have been on Hugging Face since October 1st under Apache 2.0. The car maker's interest is the fast perception layer for robots, with less demonstration data needed. SourcesBA
Tencent is reportedly preparing a consumer AI search engine
An anonymous insider told Huxiu that Tencent will combine AI answers with web results in a standalone product that shares its core with search in the Yuanbao app. Tencent has not confirmed it. It would compete directly with Baidu on its home ground. SourcesC
A second Mississippi newspaper suit brings the US count of AI copyright cases to 161
Boone Newsmedia sued Microsoft and OpenAI in the Southern District of Mississippi, by one unofficial tracker's count the 161st such case and about the 30th against OpenAI. The same plaintiffs' lawyers already have a suit there, Emmerich Newspapers v. Microsoft, before Judge Wingate. SourcesC
Simile says its digital twins predict people with 85% accuracy, on a pilot of about 1,000
CEO Joon Sung Park told The Generalist the $2 billion company wants to simulate all 8 billion people, and CVS Health has tested pricing on simulated patients. The 85% traces to a study of about 1,000 US volunteers who gave two hours of interviews, and no outside validation is described. SourcesB
One engineer merges 200 pull requests a month, and an AI reviews nearly all of them
Neil Kakkar writes in The Pragmatic Engineer that he went from eight to 20 merged changes a month to more than 200 at Tano, a startup of about 10. Human review is optional. He also lists the costs: more exhaustion, fewer people who understand the code, and a migration bug that paid creators twice. The evidence is one engineer at one company. SourcesB
Drug company chiefs push back on claims that AI will cure most diseases within a decade
Pharmaceutical leaders told The Wall Street Journal they doubt Silicon Valley's timelines, and the FDA has yet to clear an AI-designed drug. SourcesB
Sequoia closes its London office after six years and pulls back to the Bay Area
Pathfounders reports the firm is consolidating, saying it no longer needs a local presence, with one partner confirmed as staying in Europe so far. SourcesC