Daily Highlights
- Anthropic kicks off IPO roadshow, targeting listing week of Nov 16
- Anthropic buys 1.7GW of gas generators for $3.4B
- 46 US data-center ballot measures pass post-midterms
- Chinese LLMs lead US in call volume for 28 straight weeks
- SGLang open-sources parallel speculative decoding, up to 4x throughput
TODAY'S HEADLINES
[Anthropic to Begin IPO Roadshow Next Week, Targets Nasdaq Listing Week of Nov 16] According to The Information, Anthropic plans to start its IPO roadshow next week, aiming to go public the week of November 16 ahead of Thanksgiving, with the price range not yet set. The company publicly filed its S-1 on October 26, reported $16.2B in first-half revenue, and is seeking a valuation of up to roughly $2 trillion, which would make it the largest IPO in history, surpassing SpaceX. (Source: The Information)
[Anthropic Spends $3.4B to Take Over 1.7GW of Gas Generators as AI Labs Race for Their Own Power] Anthropic has agreed to pay $3.4 billion to take over an order for 1.7GW of Caterpillar heavy-duty gas generators from OnePWR Solutions, the largest single power-equipment purchase on record — effectively buying its way past grid interconnection queues. Its updated IPO filing shows nearly $900 billion in compute and infrastructure commitments through 2045, while grid interconnection waits in major US markets commonly run 5-7 years; power is replacing chips as AI's new moat. (Source: TMTPost/Edge AI Daily)
[US AI Stocks Slump as Voters Thwart Data-Center Expansion Post-Midterms] The Nasdaq fell 5.52% on November 5, the Philadelphia Semiconductor Index plunged more than 12%, and Nvidia lost over $400B in market value in a single day; per Ballotpedia, 46 of 47 decided data-center ballot measures nationwide supported limiting development. UBS estimates about 30% of planned US capacity for 2028 faces potential construction pauses as AI infrastructure enters a "permitting era." (Source: TMTPost/Edge AI Daily)
NEW PRODUCTS & TECH
[OpenAI Launches GPT-6 Cyber: Finds 948 Linux Kernel Bugs in One Run, First Critical Rating] OpenAI on Nov 6 released GPT-6 Cyber, a security-focused model scoring 48.5% on the ExploitGym benchmark — a ~51% jump over its predecessor — and writing working exploits for 948 local privilege-escalation bugs in the Linux kernel in a single automated run, more than 5x the previous model, plus 81 previously unknown high-severity Firefox vulnerabilities. OpenAI rates it Critical under its Preparedness Framework, and access is limited to vetted Daybreak Red customers. (Source: Ifeng Tech/QbitAI)
[Google Releases Gemini Omni 1.2 Flash: Native 15-Second Video, 30% Cheaper] On Nov 5 Google launched the Gemini Omni 1.2 Flash video generation model, raising the clip-length cap from 10 to 15 seconds (customizable in 1-second steps from 3-15s) and adding multi-turn conversational editing that revises generated video without starting over. Pricing for 720p generation dropped 30% to $0.07 per second; the model is now broadly available via the Gemini API, Flow, and Google Vids. (Source: ITHome/Ifeng Tech)
[TokenRhythm Debuts Agent-Native Model NeoHorse-1.5-Flash Preview, Exploring Recursive Self-Improvement] On Nov 9, TokenRhythm released NeoHorse-1.5-Flash Preview, an agent-native model trained on Qwen3.8-Flash with 180B total parameters (6B active), focused on agent task execution and software engineering. The company feeds execution feedback from real tasks back into training — its "second lap" of recursive self-improvement — with limited beta access now open. (Source: Beijing News)
IPO & FUNDINGS
[Bernstein Estimates OpenAI's Compute Spend at $614B Through 2030, Below the Widely Cited $750B] A Bernstein research note led by analyst Mark Moerdler (Nov 8) itemizes OpenAI's disclosed compute deals and maps them to data-center delivery schedules, estimating actual infrastructure spending of about $614 billion through 2030 (roughly RMB 4.1 trillion), with the spending peak pushed to 2031 at about $184 billion. The report notes chips are largely vendor-financed in installments and some contracts embed stock-price-linked warrants, so actual outlays could shrink further on delays. (Source: TMTPost)
OPEN SOURCE & RESEARCH
[SGLang Parallel Speculative Decoding Hits SOTA: Causality Baked into Draft Network, Up to 4x Throughput] On Nov 9, Machine Heart reported that the SGLang community open-sourced a parallel speculative decoding scheme that embeds causal dependencies directly into the draft network's structure, boosting speculative decoding throughput by up to 4x for SOTA results. Separately, the Beijing Humanoid Robot Innovation Center released Hi-WBC, a general whole-body control model combining human motion references with real physical constraints for autonomous work on complex terrain. (Source: Machine Heart)
COMPANY & PEOPLE
[Apple Posts $114.5B Quarterly Revenue, Net Profit Up 12%] Apple reported fiscal Q4 revenue of $114.5 billion with net profit up 12% year over year. CEO Tim Cook urged the AI team to accelerate the pace of breakthroughs; AI features are becoming an increasingly important growth lever across its services and hardware ecosystem. (Source: C114/Labor Daily)
POLICIES & DATA
[Chinese LLMs Lead US in Weekly Call Volume for 28 Straight Weeks, Taking 4 of Top 5 Spots] Per OpenRouter data cited by Mydrivers on Nov 9, total global LLM call volume reached 210 trillion tokens last week (Nov 2-8), up 8.25% week over week; Chinese models handled 116.18 trillion tokens (+16.71%) versus 13.98 trillion for US models (-32.4%), extending China's lead to 28 consecutive weeks. DeepSeek V4.1 Flash 1017 topped the chart for a third straight week. (Source: Mydrivers)
[46 Post-Midterm Data-Center Ballot Measures Pass as US AI Infrastructure Enters the "Permitting Era"] Data centers became a central issue in the 2026 US midterms: of 50 related ballot measures across 13 states tracked by Ballotpedia, 46 of 47 decided supported limiting development, and Texas's governor-elect pledged to pause statewide data-center construction until regulations pass. Roughly 40% of planned US data-center capacity sits in restriction-minded states, shifting the industry's pricing power from "who has projects" to "who has cheap capital and reliable permits." (Source: TMTPost/Edge AI Daily)
Powered by AOYII.com
