AI News: November 5, 2026

1. Voters Side Against Data Centers in 40 of 41 Local Ballot Measures in the Midterms

Data center ballot measures. Voters sided against data centers in 40 of the 41 local data center measures on the November 3 ballot across 10 states, according to Ballotpedia and CNBC’s results; Ohio had 21 of them, and the one exception was Harwood, North Dakota, where voters rejected an added review of industrial projects. In a CNBC voter poll, 56% of voters called data centers a crucial or major factor in their decision, although AI and data centers still ranked last among the issues in NBC’s exit poll, behind corruption (90%) and the cost of living (89%), The Decoder reports. CNBC notes that the OpenAI, SoftBank and SB Energy campus planned for Ohio, which could cost more than $500 billion, may face resistance from governor-elect Amy Acton, who supports a “conditional moratorium” on data centers. Source The Decoder

2. Texas Elects a Governor Who Wants a Data Center Moratorium, and Data Center Stocks Fall

Gina Hinojosa. The Democrat, who campaigned for an immediate moratorium on data center development, unseated Texas Governor Greg Abbott; Wells Fargo estimates Texas holds 27% of planned U.S. data center gigawatts, with Georgia at 7% and Ohio and Pennsylvania at 5% each, all states where Democrats won or kept the governorship. In Wednesday morning trading CoreWeave and Iren fell 4% while Microsoft rose 5.5% and Amazon 7%, per CNBC, and Reuters reported closing declines of 23% for Hut 8, 11% for IREN, 10% for Core Scientific and 7% for Applied Digital. Wells Fargo’s Ohsung Kwon had described a Democratic win as “a left-tail event to the AI trade.” Source Reuters via Techmeme

3. AI Industry PACs Lose Most of Their Targeted Races, While Safety-Aligned Groups Report Wins

Leading the Future. American Mission, the Republican arm of the industry-backed super PAC, and Building the Future together spent $4.8 million backing South Carolina Senate candidate Darline Graham, who lost, and American Mission’s $500,000-plus efforts for Aaron Flint in Montana and Laurie Buckhout in North Carolina also failed, while its Democratic arm Think Big backed Rep. Jared Moskowitz, who won, TechCrunch reports. Anthropic-backed Public First fared better with Democrats than Republicans and gave $3 million to a PAC supporting California state Sen. Scott Wiener, who was leading at publication, and Guardrails Alliance spent more than $1 million each against Rep. Jimmy Gomez and Flint, both of whom lost. Leading the Future, which spent $27.38 million in the primaries, says it will keep spending in 2028 and push for a national AI framework that overrides state laws. Source

4. Oracle Scraps a $7 Billion Deal to Lease Up to 100,000 AI Chips to Tencent

Oracle. Oracle canceled a five-year, roughly $7 billion contract signed earlier this year that would have given Tencent access to as many as 100,000 advanced AI chips in data centers across Southeast Asia, the Financial Times reports. The contract was dropped after U.S. officials raised concerns with Oracle amid pressure from the Trump administration, while Beijing scrutinized Tencent over the deal and asked it to seek domestic chips. Source

5. Isomorphic Labs Is Reportedly in Talks to Raise About $5 Billion at a $50 Billion Valuation

Isomorphic Labs. Alphabet’s AI drug discovery company, spun out of Google DeepMind, is in talks to raise roughly $5 billion at a valuation of about $50 billion, Bloomberg reports. The company last raised $2.1 billion in May. Source

6. Tencent Releases Hy4 With Vision and Computer Use, and Promises Open Weights Within Days

Tencent. Tencent launched Hy4, adding visual understanding and command-line and GUI computer use, and said it will publish the model weights and system cards “in the coming days.” API pricing matches Hy4 Preview at about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens, with 50% off on Tencent Cloud TokenHub through December 31, and the API is also on OpenRouter. In a Tencent-run blind evaluation with 213 finance professionals across 579 tasks, Hy4 scored 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3, and Tencent says its hallucination rate fell from 12.95% to 7.2% versus the preview. Source

7. AMD Discloses a $2.9 Billion Price for Model-Specific AI Chip Startup Taalas

AMD. A securities filing on Wednesday put the price of AMD’s acquisition of Toronto-based Taalas, announced in August, at about $2.9 billion in cash and stock, with $1.4 billion contingent on unspecified conditions, The Globe and Mail reports. Taalas, founded in 2023 by former AMD executives Ljubisa Bajic, Lejla Bajic and Drago Ignjatovic, designs chips built around specific AI models rather than general-purpose processors, starting with one tailored to Meta’s Llama family, and AMD plans to fold the technology into future inference chips. The startup had raised more than $225 million, including a $169 million round in February, and the deal still needs regulatory approval. Source

8. Ricursive’s Ripple-65B Preview Tops U.S. Open-Weight Models on TerminalBench 2.1

Ricursive Intelligence. The lab released Ripple-65B Preview, a dense 80-layer transformer trained from scratch on 12 trillion tokens with up to 1M tokens of context, as a general-purpose variant of the model behind its chip design platform. It reports 80.52 on TerminalBench 2.1, ahead of Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4) and NVIDIA Nemotron 3 Ultra 550B (56.4), and says full R&D cost about $32 million at roughly 40% MFU without custom kernels or quantization. Weights are on Hugging Face and the model is served through Together AI, GMI Cloud and OpenRouter; the announcement does not state a license. Source

9. TeraWulf Gets a Google Credit Backstop to Finance Its $19 Billion Anthropic Data Center Lease

TeraWulf. Google signed credit support agreements for the debt financing of TeraWulf’s Justified Data Campus in Hawesville, Kentucky, which carries about $19 billion in contracted lease commitments with Anthropic. TeraWulf also signed a definitive EPC agreement with Fluor for three data center buildings, with phased ready-for-service dates from late 2027 through the second quarter of 2028. The release does not disclose the size of Google’s backstop. Source

10. Nooks Secures $500 Million From General Catalyst’s Customer Value Fund for Its Sales Agents

Nooks. The San Francisco company secured $500 million from General Catalyst’s Customer Value Fund (CVF). Its platform combines dialing and sending infrastructure with specialized AI agents and Node, a personal assistant that monitors a rep’s deals, and it says more than 2,000 companies use it, including Glean, Notion, Deel, Replit and Factory. The money is earmarked for new products, enterprise and international expansion, and hiring. Source

11. Palo Alto Networks Is Reportedly Buying Agent Credential Startup Clutch Security for About $500 Million

Palo Alto Networks. The security company is in talks to acquire three-year-old Israeli startup Clutch Security, which develops software for tracking the passwords used by AI agents, for around $500 million, The Information reports. The talks follow CrowdStrike’s agreed purchases of Fiddler and Dreadnode earlier this week. Source

12. Sentra Raises a $150 Million Series C to Limit What AI Agents Can Reach in Corporate Data

Sentra. Capital One Ventures led the round, with Bessemer Venture Partners, Zeev Ventures, Standard Investments, Munich Re Ventures and Key1 Capital participating, bringing total funding past $250 million; no valuation was disclosed. Sentra maps which people, service accounts and AI agents can reach each dataset, including inherited permissions, trims access to the minimum, flags agents that touch sensitive data they normally do not, and can revoke the permissions automatically, with analysis running inside the customer’s environment. The company says its business grew more than 300% over the past year. Source

13. Applied Compute Raises $350 Million at a $3.25 Billion Valuation After Passing $100 Million in ARR

Applied Compute. Elad Gil led the round, with Nvidia, Shopify Ventures and HubSpot Ventures among more than a dozen participants. The company, launched less than 18 months ago, sells a platform for training and serving LLMs that uses asynchronous reinforcement learning and self-distillation without teacher models, an “online RL” feature that periodically fine-tunes deployed models on incoming prompts, and a model router called Shift-1. Customers include Microsoft and DoorDash, and the funds go to market expansion and more infrastructure. Source

14. Microsoft Reportedly Declines to More Than Double Its Texas Data Center Contract With Nscale

Nscale. Microsoft turned down an option to more than double its Texas data center contract with the British AI infrastructure company, a fresh blow to Nscale’s ambitions for a $35 billion IPO, the Financial Times reports. The report came a day after data center stocks fell on expectations of a slower AI buildout following the midterm results. Source

15. China Tells SoftBank It Will Block Its $5.4 Billion Purchase of ABB’s Robotics Unit

SoftBank. China has told SoftBank it will not allow its $5.4 billion acquisition of ABB’s robotics division to go ahead, the Financial Times reports. China is the division’s biggest market, and the unit runs a large factory in Shanghai’s Pudong district. Source

16. SemiAnalysis Says SpaceX Now Targets About 13 GW of Data Center Capacity in 2027

SpaceX. SemiAnalysis reports that SpaceX raised its 2027 data center target to about 13 GW from roughly 10 GW, with 3.6 GW under construction and more than 9 GW of identified capacity, which the firm compares to roughly all the capacity Google built from 1999 to 2026. It says Colossus 2 in the Memphis area added 210 MW in 91 days, and it tracks more than 8 GW of likely behind-the-meter power plus a 5.9 GWh battery installation there. The authors note the figures are estimates from construction observation and filings, and SpaceX has not confirmed a data center at its planned TeraFab site in Texas. Source

17. Qualcomm Targets $5 Billion in Data Center Revenue for Fiscal 2027 as Handset Chip Sales Fall 26%

Qualcomm. Qualcomm reported $650 million in data center revenue for its fiscal fourth quarter, mostly from customization and connectivity work, and set a $5 billion target for fiscal 2027, while handset chip sales fell 26% to $5.14 billion as rising memory costs push up phone prices. CEO Cristiano Amon said the company has engaged an unnamed global hyperscaler and that “We don’t have a demand issue; it’s a supply issue.” Shares fell 7% in extended trading on next-quarter guidance of $10.4 billion to $11.2 billion in revenue. Source

18. Waymo Will End Its Exclusive Robotaxi Arrangement With Uber in Austin and Atlanta in January 2028

Waymo. Waymo confirmed it will end its exclusivity with Uber in Austin and Atlanta in January 2028, ahead of the deal’s May 2028 expiry, and will offer rides through its own app as it does in other cities. Co-CEO Tekedra Mawakana called the rider experience “a little bit of a crapshoot” because riders cannot always tell whether a robotaxi or a human-driven car will arrive. Uber says it still plans to host Waymo vehicles for as long as the companies agree. Source

19. Uber Says AI Agents Booked Four Rides During Its Busiest Week

Uber. In the week of Halloween, its busiest week in October, AI agents booked four Uber rides and placed 500 Uber Eats orders, against roughly 320 million trips and orders across both services, CFO Balaji Krishnamurthy said on the earnings call. CEO Dara Khosrowshahi called it “very, very early and very limited traction,” and the company noted that booking a ride in its app already takes about 50 seconds on average. DoorDash reported only a few hundred agent orders the week before. Source

20. Pangram Extends Its AI Detection From Text to Images and Video

Pangram. The AI text detection company released Pangram Image 1.0, which flags AI-generated images with a heatmap of affected regions, and Pangram Image Flash 1.0, which checks uploaded video frame by frame. Pangram reports 99.96% accuracy on its own set of 1,200 synthetic images from 12 frontier generators and 98.7% on the DF26 video benchmark, 29 points ahead of the next detector, but does not publish false positive rates. Both are available on its site, its Chrome and Firefox extensions and its API, and the minimum input size drops to 256x256 pixels. Source

21. CloudOpsBench Finds the Best Agent Completes 60% of Real AWS DevOps Tasks

CloudOpsBench. A University of Michigan team led by Ang Chen released an Apache 2.0 benchmark of 45 AWS DevOps tasks written by experienced engineers, such as fixing services, restricting access rights and decommissioning projects, which run identically on real AWS and in an emulator. Claude Fable 5.1 solved 60% and GPT-6 Astra 53%, and no model completed every task. About 80% of failures came from misjudging task scope, not checking all relevant services or stopping too early, while 20% were wrong actions, some with irreversible effects on infrastructure. Source

22. ToolSpace Builds Agent Training Environments From Real MCP Traces and Lifts GLM-5.2 to Claude Opus 5 Level

Renmin University of China. ToolSpace mines data, state and retry dependencies between adjacent calls in execution traces from real MCP services, clusters them into environments, and generates tasks with assertion checks that must replay successfully from the initial state. Applied to AutomationBench and tau3-Bench Banking, it produced harder “Pro” versions on which seven frontier models dropped by up to 38.2 points. Training on its data raised Qwen3.6-35B-A3B from 23.0 to 41.0 on average across seven long-horizon tool-use benchmarks (43.2 with RL) and GLM-5.2 from 44.6 to 56.6, which the authors say is on par with Claude Opus 5. Source

23. Route Anything Finds Learned Routers Raise Efficiency 8.82x Across 452 Model and Policy Configurations

Route Anything. The study builds a matrix of 374,401 scored runs from 452 candidate models and policies across 25 benchmarks spanning language, image, video and speech generation, coding and workplace agents, and robot manipulation. Per-input oracle routing beat the best single candidate by 5.4 points, while learned routers beat it by 0.3 points across 16 benchmarks, 2.1 points on video generation and 1.1 points on robot manipulation while using 24% less policy time. Routing raised efficiency 8.82x on average compared with the reference model. Source

24. Context Rot in Retrieval Comes From Competing Statements, Not Length, a New Paper Finds

“Length Is Cheap, Competition Is Not.” Testing five open models from 0.5B to 3B parameters, the single-author paper finds that neutral filler up to 32K tokens leaves retrieval accuracy at 0.97 to 1.00, while statements in the evidence’s format at the same positions drop it to 0.16 to 0.33, and each rival value for the queried key adds to the loss. Across 126,473 trials, the attention mass left on the evidence explains up to 58% of the variance in loss, against at most 8% for log length. The paper also shows that the masked-filler protocol behind length-only claims costs 45 points at 32K tokens in retrieval while the same filler left visible costs none, and it releases all trial-level data. Source