AI News: November 4, 2026

1. Fal Raises $1.3 Billion Led by Bessemer, Doubling Its Valuation to $16 Billion in Seven Months

Fal. The San Francisco company, which hosts and runs image, video, and audio generation models for developers, raised a $1.3 billion Series F led by Bessemer Venture Partners, with Nvidia, Sound Ventures, and T. Rowe Price participating, at a $16 billion valuation, double the $8 billion it was valued at in April, Bloomberg reports via TechCrunch. Fal announced the round together with a Sequoia-led Series E, for $1.65 billion across the two, and Bloomberg says it passed $800 million in annualized revenue earlier this year; SiliconANGLE reports the money will go to compute and to hiring beyond its current 185 employees. TechCrunch notes that inference rivals Baseten, Fireworks AI, and Modal are also raising at rising valuations, while people familiar with their finances describe gross margins across the category as thin because of the cost of buying or leasing compute. Source SiliconANGLE

2. AMD’s Data Center Revenue Nearly Doubles, and Lisa Su Expects Server Revenue to Double Again in 2027

AMD. Third-quarter revenue rose 45% to $13.38 billion against a $13.09 billion consensus, with adjusted EPS of $2.09 versus $1.95 expected, and the data center segment nearly doubled to 64% of revenue (about $8.5 billion), up from 47% a year earlier. CEO Lisa Su said she expects server revenue to double year over year in 2027 and data center AI growth to accelerate faster than AMD projected 90 days ago, and that AMD has secured wafers, substrates, advanced packaging, and memory for “much higher growth next year”; CNBC ties much of the demand to server CPUs that run the virtual machines behind long-running agents, and Su said “we expect demand to grow sharply as agents are deployed more broadly.” Fourth-quarter guidance is about $17 billion versus $16.2 billion expected, at a 55% gross margin, while the combined PC and gaming segment fell 8% to $3.7 billion amid a memory-driven PC slump; shares rose 8% after hours. CNBC adds that Meta, Microsoft, Anthropic, OpenAI, and Oracle are starting to deploy AMD’s Helios rack-scale systems. Source

3. Super Micro Reports $19.9 Billion in Quarterly Sales, Four Times a Year Earlier, and Guides to $72 to $82 Billion for the Fiscal Year

Supermicro. Net sales for the September quarter (fiscal Q1 2027) were $19.9 billion, versus $11.1 billion in the June quarter and $5.0 billion a year earlier, with net income of $1.59 billion and non-GAAP diluted EPS of $1.81. GAAP gross margin fell to 12.7% from 17.5% in the prior quarter (9.3% a year earlier), operations used $434 million of cash, and the company ended September with $6.9 billion in cash against $8.6 billion of bank debt and convertible notes. Supermicro guided December-quarter sales to $19.5 billion to $21.5 billion and fiscal 2027 to $72 billion to $82 billion, with CEO Charles Liang crediting its Data Center Building Block Solutions, from “gigawatt-scale datacenters down to single-container PODs.” Source

4. CrowdStrike Agrees to Buy Agent Observability Startup Fiddler and AI Security Tester Dreadnode, Reportedly for About $495 Million

CrowdStrike. CrowdStrike signed definitive agreements on November 3 to acquire Fiddler AI and Dreadnode, for about $400 million and $95 million respectively according to The Wall Street Journal (via Techmeme), paid mostly in cash with a stock portion subject to vesting. Fiddler’s pre-deployment evaluation, decision tracing, and drift detection for agents, instrumented through OpenTelemetry with support for LangGraph, Amazon Bedrock, AWS Strands, and Google ADK, will be sold as Falcon Observe; Dreadnode runs more than 100 attack strategies, including jailbreaks, prompt injection, and multi-turn and agent-specific attacks, against models, agent reasoning loops, applications, and MCP servers, and maps findings to the OWASP Top 10 for LLMs, OWASP Agentic Security (ASI01 to ASI10), MITRE ATLAS, and NIST AI RMF. Both products are slated for general availability in CrowdStrike’s fiscal fourth quarter (November 2026 to January 2027), and the company announced an AI kill switch for its Falcon Guardian agent security product the same day. Separately, cybersecurity VC Ballistic Ventures, which backed Pangea before CrowdStrike bought it, filed to raise $490 million across two new funds, and TechCrunch cites a JPMorgan report that AI-enabled companies made up about 72% of US cybersecurity venture deals through May. CrowdStrike: Fiddler CrowdStrike: Dreadnode Techmeme TechCrunch

5. Traversal Raises a $120 Million Series B at a $1 Billion Valuation for Its AI Site Reliability Agents

Traversal. The round was co-led by Kleiner Perkins, Redpoint Ventures, and Khosla Ventures, with Sequoia and Battery Ventures participating, and will fund hiring across research, engineering, and go-to-market. Traversal’s agents investigate production incidents using a “Production World Model” built from telemetry, code, deployment history, and team knowledge, plus a “Causal Search Engine” that tests hypotheses in parallel, both introduced in March; the company says monthly investigations grew from 513 to nearly 73,000 this year and that about two-thirds of usage now goes beyond root cause analysis to alert triage, release risk, and production support. It pitches the product as cleaning up after AI-written code, citing a survey in which 52% of developers spend more time debugging in production, and offers Workers (long-running agents triggered by alerts) and an MCP server that feeds production context to coding agents; Forbes reports American Express and Pepsi as customers. Source Techmeme

6. Revolut Says Self-Hosting GLM for 2,000 Engineers Cut Its AI Coding Inference Costs by About 90%

Revolut. Speaking at the AI Engineer conference in New York last month, ops and infrastructure lead David Abitbol said the fintech made self-hosted Z.ai models its default for coding after budget-capped rollouts of Gemini (early 2025) and Claude Code (early 2026), when some engineers were spending $1,000 to $10,000 a month; self-hosted GLM-5.1 cost about $0.30 per million tokens, roughly a tenth of Codex or Claude Code, The Register reports. Revolut now serves GLM-5.3 on 56 Nvidia B200 GPUs, and Abitbol said the bigger gains came from tuning, NVFP4 quantization, disaggregated prefill and decode, and holding context to about 256,000 tokens (“Long context is a scam”), which gave under 3 seconds to first token and about 200 tokens per second. With tokens effectively unlimited, engineers turned the models on debugging and Jira and Confluence tickets, Revolut reports 60% faster coding time and 70% shorter lead time since its AI rollout, and a router still sends vision work to Gemini or GPT Luna because GLM is text only. Source

7. Gartner Predicts Open-Weight Models Will Run 90% of Enterprise Agent Workflow Steps by 2030

Gartner. In a research note titled “Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps,” senior director Mark Petty and group vice president Raymond Paquet argue that nine in ten agent workflow steps (classification, tool calling, validating, and reconciling data) are routine, and that open-weight models handle them at parity for a quarter of the price of the cheapest closed model with the same score and as little as a twentieth of the cost of running a frontier model on every step, The Register reports. They concede that current open models are more verbose and retry more often, so the realistic saving per task against a frontier model is closer to five or six times, and they call keeping workflow history inside the enterprise perimeter the main reason to switch, with the cost saving “a bonus.” The note expects mostly domain-specific open models on stable steps that run millions of times a month, frontier models to remain the engine of R&D, and the harnesses that route work between models to be where product value sits. Source

8. Stanford Retests 50 Widely Cited LLM Findings on 20 Models and Finds 44% Have “Expired”

Stanford. Owen Queen and James Zou reran 50 published claims, from sycophancy and jailbreaks to long-context behavior and cultural biases, on 20 models released between July 2024 and September 2026 by eight developers, and fit trends against release date. 44% of the effects have largely disappeared (a fifth of the original effect or less), 34% have weakened substantially, 16% persist, and 6% have strengthened; among expired claims with an estimated date, the average lifespan was 15.7 months. Claims about what models cannot do expire fastest (five of nine reasoning findings, including letter counting and lost-in-the-middle), while evaluator self-preference, sycophancy, and a left-leaning political bias persist; rerunning a 2023 study in which GPT-3.5 fabricated 55% of its citations, GPT-6.1 Sol fabricated none and Claude Sonnet 5.5 1%, but Qwen3.7-Flash fabricated 69%. The authors ask papers to report model, version, and date and to publish prompts and analysis code so findings can be retested. Source The Decoder

9. Critics Say the Weight-Space Programming Post Blurs Tool Use and Raw-Weight Reading, as Simon Willison Reproduces One Test on a Laptop

Weight-space programming. spaCy author Matthew Honnibal called the MIT CSAIL “weight-space programming” post (published November 2) misleading: the no-tools setting (“just by looking at the raw weights”) appears only in experiments on networks of 17,000 to 172,000 weights, while every experiment on a real language model, including the headline Qwen3.5-0.8B MATH-500 jump from 7.9% to 48.1%, has GPT-6 Astra write Python and run the model on up to 50 batches of practice problems, which he says is work a human ML engineer could do. He adds that the from-scratch MinAtar networks beat the PPO-500M baseline only on Breakout (517 vs. 25) and lose on Seaquest (53 vs. 89) and Asterix (53 vs. 55), and Sean Goedecke separately argued the Qwen edit follows standard researcher steps. Simon Willison then ran the paper’s no-tools “which game was this network trained on” prompts against Qwen 3.8 27B in LM Studio on his laptop: after 46 to 65 minutes of reasoning per prompt it identified Pac-Man, Pong, and Breakout (stated confidence 72% to 88%) and missed Freeway, while Kimi K3.1 Pro identified Breakout in under 10 minutes via OpenRouter. Honnibal Simon Willison

10. ByteDance’s Axion 1.0 Releases a Three-Stage Agent Training Recipe, With Its 122B Model Scoring 59.0 on HLE

ByteDance. Axion 1.0, from a ByteDance-led team with Duke, Maryland, and Buffalo co-authors, fine-tunes Qwen3.6-35B-A3B and Qwen3.5-122B-A10B with two SFT stages followed by asynchronous GRPO, skipping dedicated mid-training, DPO, and on-policy distillation. Stage one uses 10,819 reasoning trajectories made by rewriting problems the base model already solves into harder ones, and stage two uses 7,544 tool-use trajectories (about 312 million tokens) whose reference answers are computed programmatically from sources such as the NASA Exoplanet Archive, OpenAlex, UniProt, and Ensembl. The 122B model reports 59.0 on full HLE with restricted web access (the table lists Claude Fable 5 at 63.8), 87.0 on FrontierScience Olympiad, and 66.7 on FrontierScience Research, and the 35B model reports 62.2 on SEAL-0; several baselines were evaluated by the Axion team. The Apache 2.0 repository includes data generation, filtering, prompts, SFT, RL, and grading code, with training data on Hugging Face, but the 35B checkpoint it links is a community reproduction rather than an official release. Source GitHub

11. Berkeley’s SpecPilot Has Agents Tune Speculative Decoding Per Deployment, Lifting Single-Request Speedups From 4.46x to 10.87x

UC Berkeley. SpecPilot, from a team including Kurt Keutzer, Jiantao Jiao, and Kannan Ramchandran, targets parallel drafters, whose released systems it finds fall well short of the speedup parallel drafting allows. Each round, a cost model locates the largest gaps between the current system and its ceiling, parallel agents explore fixes and integrate the ones that work, “token forcing” checks every output token against the stock system to catch reward hacking, and a structured memory carries validated mechanisms to new deployments. The loop discovered a lossless 12-bit weight format, tree verification for hybrid recurrent targets, two-level verification, and three batching mechanisms; across four target models, three drafters, two GPU platforms, and five benchmarks, it raised speedup over autoregressive decoding from 4.46x to 10.87x with one request and from 1.94x to 3.93x with batching, without training or quality loss. Source GitHub

12. CMU’s IdeaTaste Finds LLM Research Proposals Reach Only 18% of Human Level at Matching the Ideas Fields Went On to Pursue

Carnegie Mellon. IdeaTaste, from Chenyan Xiong, Aviral Kumar, and colleagues, gives models a literature snapshot frozen at 2025 and scores the research proposals they write by how closely they match what the same communities published in 2026, across six models, four research areas, and more than 90,000 proposals. Model populations reach about 26% of the human level on citation matching and 18% on idea matching: they cite as diversely as a community and pick up its tasks and methods, but rarely arrive at the concrete approaches researchers pursued, and neither richer field context nor larger populations closes the gap much. In blinded expert review, AI proposals trailed the 2026 papers least on presentation and most on expected success, and proposals whose ideas no human paper shared scored worst, which the authors read as weaker directions rather than overlooked ones; the benchmark runs as a continuously updated leaderboard. Source

13. CUA-Master-27B Reports 80.2 on OSWorld-Verified After Mining Seed Tasks From Instructional Videos

CUA-Master. The paper describes a two-loop recursive improvement pipeline for computer-use agents: an outer loop folds in open-world data and aims each iteration at the model’s current weaknesses, while an inner loop mines real workflows from instructional videos as seed tasks, reconstructs their trajectories from visual and semantic cues, and turns them into reusable primitives for synthesizing diverse, verifiable tasks, with a snapshot-and-rollback retry mechanism for collecting trajectories in real environments. The 27B model scores 80.2 on OSWorld-Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents’ Last Exam, and the authors report continued gains across iterations; code is on GitHub. For scale, the NorthCUA report released two days earlier listed 78.9% for its 9B model and 88.6% for its 397B model on OSWorld-Verified under its own settings. Source GitHub

14. Huawei Reportedly Tells Customers to Source HBM for Its AI Chips Themselves, Feeding a Smuggling Market

Huawei. Huawei, which plans to launch an Ascend 960DT chip early next year that it says doubles current performance, is telling some customers to procure the high-bandwidth memory themselves because Chinese suppliers cannot make enough and US rules block direct purchases from Samsung and SK Hynix, which together hold more than 80% of the market, The Information reports via The Decoder. Smugglers bundle HBM with low-end server processors that fall under export thresholds, ship them through Hong Kong, and separate the memory in workshops on arrival; a Shenzhen broker says smuggled HBM3E rose from a few hundred dollars to about $4,000 per unit in a year, and at six to eight units per chip the memory alone can cost $32,000. South Korea’s customs service had found $560 million in trade violations by May, already a full-year record, a third of it involving strategic goods such as chips. Source

15. AI Data Centers Are Installing Converted Jet Engines to Get Power in About Two Years

Data center power. Stuck in multi-year queues for grid connections and new turbines, AI data center builders are buying retired aircraft engines rebuilt to burn natural gas, the New York Times reports via TechCrunch; Rystad Energy’s Audun Martinsen says such units can have a data center running in 24 months, compared with about seven years to connect 100 MW to the grid in Northern Virginia, per the IEA. xAI has installed eight converted turbines from ProEnergy, which rebuilds a GE engine common on Boeing jets, in Mississippi to help power Colossus 2, and Crusoe, which is building OpenAI’s first Stargate campus in Abilene, ordered 13 ProEnergy units in April. The engines are less efficient and pollute more than turbines designed for gas plants, and more of them are needed for the same output, so some experts expect them to be a stopgap; last month Microsoft pulled out as anchor customer for Nscale’s site in Loughton, England, after the utility said it could not deliver enough power until the early to mid-2030s. Source

16. Ukrainian Drone Strikes Push Yandex to Move Its Supercomputers and Russia Toward Underground Data Centers

Russia. Ukrainian drones have degraded or knocked out data centers run by Yandex, Sberbank, VKontakte, and the Russian Finance Ministry, disrupting payment, banking, e-commerce, and ride-hailing apps, and Meduza reports that Yandex has relocated all three of its main supercomputers after strikes targeted the AI data centers that housed them, Ars Technica reports. State telecom Rostelecom has discussed moving data centers underground with miner Norilsk, a regional governor has proposed building them in open-pit coal mines, and a November 2 decree from Vladimir Putin requires operators to reserve 10% of space and capacity for state emergencies and lets a government commission, consulting the FSB, authorize use of foreign data centers, a break from the “Sovereign Internet” rules that kept Russian companies on domestic infrastructure. Ars contrasts this with Ukraine’s more dispersed setup, helped by Amazon and Microsoft moving government services to cloud regions elsewhere in Europe in 2022. Source

17. Netflix Adds an “AI Discover” Tab That Recommends Titles by Mood and Available Time

Netflix. Netflix began rolling out AI Discover on supported TVs to US users who watch in English, a tab it describes as a “test kitchen” for experimental discovery features. The first feature lets viewers pick themes such as “something light,” “scare me,” “family movie night,” or “quick watches” with clips, narrow them further or restart with “something different,” and use a “Tune Your Recommendations” button to save or exclude titles; Netflix says the tab runs on a new “intelligence layer” that responds in real time to what users tell it, and that feedback will decide which features move into the main app. More users get the tab in the coming weeks. Source

18. Nexus Mods Makes AI-Use Disclosure Mandatory for New Mods but Declines to Ban AI-Generated Work

Nexus Mods. The largest PC modding site now requires authors of newly uploaded mods to declare whether AI was used, and mods tagged “AI-assisted” must publicly explain how and where; existing mods have a 60-day grace period ending January 2. Owner Victor “Foledinho” Folmann said a ban would only stop people from disclosing AI use, since AI now runs through scripting, textures, voices, and full assets, and that Nexus will not use mods to train AI or sell or license them to anyone who does. The site is also adding controls to hide AI-generated content on game pages and the homepage (search is next), capping how many mods from one author appear in Recently Added, and tightening uploads from new accounts as mod production speeds up. Source