AI Morning Brief · Nov 19, 2026
Top Stories
1.Anthropic IPO priced at $83 a share, above the $70–$78 range; trading starts November 19
Reuters first reported the pricing at about 05:26 Beijing Time on November 19 (evening of November 18 ET), with Bloomberg, CNBC and TechCrunch citing sources; Anthropic declined to comment and no Form 424B4 pricing filing had appeared by the end of the window, so $83 remains "reported, not yet confirmed by an official filing." Trading begins at 09:30 ET on November 19 on Nasdaq under the ticker ANT. The offering was set at 825 million shares (801,351,354 from the company plus 23,648,646 from existing holders), worth about $68.5B at the original share count, plus a $26B Class N private placement (AMD, Broadcom, Google, Micron, Microsoft, NVIDIA, TSMC Partners); multiple reports say the deal was upsized, but the final share count was not disclosed. Valuation is about $1.5T fully diluted (estimate) / about $1.08T on roughly 13.05 billion shares outstanding, up from $965B in the May $65B private round. Lead underwriters Goldman Sachs and Morgan Stanley head a 22-bank syndicate; the greenshoe covers up to 123.75 million additional shares. Q3 2026 revenue was $17.0B (up more than 12x year over year); revenue for the first nine months was $33.2B. (November 18–19; Reuters, TechCrunch, Bloomberg, CNBC)
Sources: techcrunch.com · wixx.com · fool.com
2.Google launches Gemini 4 Neon, priced at $3/$15, replacing Gemini 3.1 Pro and Argon
Google announced Gemini 4 Neon on its official blog on November 18, positioned between Carbon (hardest long-horizon tasks) and Hydrogen (fast everyday work). Pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 (a 95% discount) and a 1-million-token context window. Artificial Analysis gives it an Intelligence Index of 61 at the xhigh tier (3rd overall) and the top Agentic Coding Index score of 58.6. It launched the same day on the Gemini API (AI Studio), Gemini for Business, Antigravity (Pro/Ultra) and the Gemini App (Pro/Ultra), replacing Gemini 3.1 Pro and Argon; a cybersecurity edition follows within days for trusted defenders under the Fairwind Program. From December 1, Carbon/Argon introductory pricing ends at $4/$20, making Neon 25% cheaper than Carbon. (November 18; Google official blog, 9to5Google)
Sources: blog.google · 9to5google.com
3.BAAJA clears final cloture 65–31; only a simple majority now needed for passage
The Senate voted 65–31 on November 18 (vote ending 14:30 ET, 03:30 Beijing Time on November 19) to invoke final cloture on H.R. 2347 carrying the BAAJA text (Roll Call 268, 60 votes required) — the last 60-vote hurdle, after which final passage needs only a simple majority. Five Democrats flipped from yes to no (Ossoff, Blumenthal, Kim, Peters, Shaheen); only two Republicans opposed (Rounds, Hawley), with independent Sanders also voting no and the remaining 28 no votes from Democrats; more than a dozen Democrats voted yes. A manager's amendment passed 69–27 the same day. Under the 30-hour rule, the final passage vote is expected at 20:30 ET on November 19 (08:30 Beijing Time on November 20), outside this window; Bloomberg Law says a final vote is expected early Friday, after which the bill goes to the House. (November 18; Bloomberg Law, Washington Examiner, Senate vote record)
Sources: news.bloomberglaw.com · washingtonexaminer.com · pranava0x0.github.io
4.Nvidia falls 2.89% to $150.96 after earnings beat, fifth straight daily loss; congressional probe escalates
The Q3 FY2027 results themselves (revenue $115.8B, +103% year over year, above the expected $109.2B; adjusted EPS $2.68 vs $2.47 expected; Q4 guidance $130B ±3%, above the roughly $125B expected; FY2028 revenue growth outlook raised from 70% to 75%) were reported previously. The new item is the stock reaction: shares rose about 2.7% premarket (high about $159.45–$159.89), opened at $159.19, then gave back all gains to close at $150.96 (-$4.50 / -2.90%), with an intraday low of $150.40; volume of 339.5 million shares was about 2.17x the three-month average of 156.7 million — a fifth consecutive daily decline. Chairman Harvey Jones said on the November 17 earnings call that the board unanimously accepted Jensen Huang's resignation and has "no intention" of restoring him as CEO; hours earlier Huang posted on X that he would be "deeply honored" to return if the board judged it in the interests of shareholders, employees, the company and the US. The board has hired Heidrick & Struggles to find a permanent CEO, with CFO Colette Kress as interim CEO. The congressional probe escalated: House Energy and Commerce Chair Brett Guthrie wrote to Kress demanding a briefing by December 2 on chip export controls and the Huang–Alice Huang relationship, and also wrote to Commerce Secretary Lutnick and Deputy Secretary Kessler; three House committees had already announced Huang agreed to a closed-door transcribed interview on December 1 and general counsel Tim Teter agreed to December 2; Senate Banking Chair Tim Scott said on CNBC on November 18 that Huang has agreed to testify publicly (date undetermined, hoped before Christmas) and Lutnick has also agreed to appear. (November 18; Motley Fool, Stocktwits, AI Affairs)
Sources: fool.com · stocktwits.com · aiaffairs.com
5.US stocks split on November 18: Dow -1.11% for a third straight 1%+ drop, S&P 500 down a fifth day, Nasdaq ends four-day slide; software selloff continues
The Dow fell 561.25 points (-1.11%) to 50,198.17, a third consecutive session down more than 1%; the S&P 500 fell 16.90 points (-0.23%) to 7,323.63, a fifth straight decline (longest since June); the Nasdaq rose 101.50 points (+0.42%) to 24,429.88, ending a four-day slide; the Russell 2000 gained 0.52%. Minutes from the Fed's October 27–28 meeting showed almost all participants saw further rate hikes as possibly appropriate. Chips split: AMD +5.31%, Broadcom +3.65%, Philadelphia Semiconductor Index +0.1% to 9,652.8. The software selloff continued: Adobe fell 3.87% to $226.43 on November 18; Salesforce had already fallen 9.45% on November 17 (S&P cut its rating to A from A+), ServiceNow -6.51%, software ETF IGV -3.29%, with the market attributing the rotation to Anthropic IPO positioning and fears that AI agents will disrupt traditional software business models. (November 18; TickerGrove, TradingKey, Wisevoter)
Sources: tickergrove.com · tradingkey.com · wisevoter.com
6.Tencent Q3: revenue +8%, capex +527%; Hy4 Flash weights open-sourced
Tencent reported Q3 results after the Hong Kong close on November 18: revenue 207.336 billion yuan (+8% year over year); net profit attributable 63.612 billion yuan (+0.8%), non-IFRS net profit 66.519 billion yuan (-6%); capex 81.433 billion yuan (+527%), free cash flow -69.1 billion yuan; cloud revenue +35%; R&D spending 32.4 billion yuan (+42%). The same day it open-sourced Hy4 Flash weights in three versions — Hy4 flash (instruction-tuned), Hy4 flash-FP8 (8-bit quantized) and Hy4 flash Base (pretrained) — 224B total parameters, 12B activated per token, 1-million-token context and multimodal input, via GitHub, Hugging Face, ModelScope and GitCode under the Tencent Hy License (MaaS providers with over $50M annual revenue or products with over 100M monthly active users need separate written authorization); the flagship Hy4 (770B class) remains in accelerated open-sourcing with no date. WorkBuddy data disclosed with the results: daily active users above 10 million, daily token calls above 100 quadrillion, which Tencent calls the highest LLM token volume of any AI product in China; Hy4 flash exceeded 50 quadrillion token calls on its first day, topping OpenRouter's daily LLM chart; WorkBuddy's annualized revenue over the past 7 days exceeded $500M, with Q3 quarterly revenue up more than 5x from the prior quarter; it has processed 9 billion user task requests since its March launch. Management said on the earnings call that usable compute will fall in the near term, with both leased and self-built compute affected and model training under pressure; Hy4 trains with relatively high precision efficiency, but the next generation of models may take longer; pressure from AI investment is expected to ease gradually by 2028; new AI products dragged adjusted operating profit by about 15.2 billion yuan in the quarter. (November 18; 21st Century Business Herald, 17173, AF, IT168 — Chinese-language sources)
Sources: 21jingji.com · news.17173.com · af.net · cloud.it168.com
7.Hong Kong AI "twin stars" rebound: Zhipu +9.61%, MiniMax +4.40%, ending the slide
Hong Kong stocks fell on November 18: the Hang Seng dropped 213.42 points (-0.91%) to 23,294.75 (intraday low 23,208, a monthly low); the Hang Seng Tech Index fell 0.58% to 3,736.33; the HSCEI fell 0.86% to 7,836.27; turnover was HK$177.15B. The AI twin stars rebounded, ending their losing streak: Zhipu (02513) rose 9.61% to HK$367.20; MiniMax-W (00100) rose 4.40% to HK$121.1 — confirmed by three sources. The driver: Nvidia's earnings beat repaired compute-demand sentiment, with AI software names rallying in the afternoon. Alibaba, Tencent and Baidu all edged lower (see Unverified Items for a discrepancy in Alibaba/Tencent closing prices). (November 18; CnYES, Finet, Longbridge — Chinese-language sources)
Sources: gfe-desktop.cnyes.com · finet.hk · longbridge.com
8.DeepSeek's free-tier limits enter a second day with still no official response
A report at 19:03 on November 18 said the limits have been in place for two days, with users reporting "my quota was gone after two sentences" and "three or four questions and it says the server is busy"; as of then DeepSeek had still not responded officially or published unified paid rules or explanations, and market rumors of a membership fee or "Coding Plan" package remain unconfirmed. The limits (from November 16): single-input cap of about 40,000 characters on web/App, long chats inactive for over 7 days can no longer be resumed, and a daily conversation quota that triggers a "server busy, usage limit reached" message. The free tier is being limited for the first time; the pay-as-you-go API is unaffected. (November 18; 21st Century Business Herald — Chinese-language source)
Sources: 21jingji.com
9.Trump's Asia trip: Xi–Putin meet in Shenzhen for over two hours; first-ever China–US–Russia trilateral set for November 19
Xi Jinping and Vladimir Putin held a bilateral meeting in Shenzhen at noon on November 18 lasting more than two hours; Xi called it their third face-to-face meeting this year and their 50th since 2013 (including video calls), while Putin said at the APEC business summit that tools of economic coercion are failing and looked forward to the November 19 trilateral as "very useful"; Xi's opening address at the APEC CEO Summit called for stronger AI cooperation and free trade. Kremlin aide Ushakov said the agenda covered energy, finance, high technology and AI. The first-ever China–US–Russia trilateral summit is scheduled for November 19 in Shenzhen, with AI, nuclear disarmament, Ukraine and Iran, and trade expected on the agenda; Reuters said AI and advanced-chip export controls may be discussed; a bilateral Xi–Trump meeting is also set for November 19 (both outside this window; no outcomes had been reported by the end of the window). After the Quad leaders' meeting (reported in the previous edition), Trump met Philippine President Marcos Jr. in Manila on November 18, calling the US–Philippines alliance America's oldest and most important in the Pacific; no US–Philippines tech agreement will be signed on this trip. (November 18; The Chronicle, The Hindu BusinessLine, BusinessWorld)
Sources: thechronicle.com.ph · thehindubusinessline.com · bworldonline.com
Platform & Model Tracking
OpenAI (GPT/ChatGPT)
New frontier-model lifecycle safety review framework (November 18, US afternoon): every stage from research to internal deployment to external release must pass a safety review, with training paused or a model withheld if evidence is insufficient; four layers of protection — model alignment, safety and sandboxing, blocking safeguards, and monitoring and response; a Control Room already deployed on research infrastructure can halt model output within 35 seconds of a shutdown instruction. Follow-up: early GPT-6.2 reviews are mixed (November 18) — Artificial Analysis puts Astra's Intelligence Index at 62 (second only to Claude Fable 5.2's 65, above Gemini 4 Carbon's 60) and Sol at 57, but Astra's agentic coding score at maximum effort matches GPT-6 Astra; VentureBeat cites critics including Dan Shipper and Theo Browne, who called it "OpenAI's worst model ever" (third-party opinion, not official data).
Sources: news-block.com · therundown.ai
Anthropic (Claude)
IPO priced at $83 (see Top Stories 1). Valuation about $1.5T fully diluted (estimate) / about $1.08T on roughly 13.05 billion shares outstanding, up from $965B in May; Q3 2026 revenue $17.0B (up more than 12x year over year), first-nine-months revenue $33.2B — the nine-month figure corroborates the previously unverified "$33.2B" (operating loss of $1.3B and $70.7B of forward commitments remain unverified; see Unverified Items).
Sources: techcrunch.com · unite.ai
Google DeepMind (Gemini)
Gemini 4 Neon launched (see Top Stories 2). Gemini OneGov federal discount extended three more months (November 18, 12:18 ET): federal agencies keep using Gemini at $0.47 per agency through February 28, 2027 (was due to expire November 15); this is the second extension (previously September 30 → November 15); still includes 20% off first-party cloud services and 50% off networking/GCE egress. Follow-up: Gemini can now take tasks by text message or email, starting in the US (November 18) — via Google Messages, iMessage or email, US-only and English-only; SMS is not end-to-end encrypted; no date for outside the US. Single source: Gemini App comes to low-memory Android (Go edition) phones (November 18, Google blog) — rebuilt core experience runs on low-memory devices; help page sets the bar at 2GB RAM and above; Gemini Live in 45+ languages; the blog gives no rollout date or device list.
Sources: blog.google · nextgov.com · tbreak.com
Meta
Loblaw becomes the first Canadian retail partner for Muse (November 18, 08:00 ET) — Loblaw's product catalog connects to Muse via a partner connector, letting Canadian users pick a store, plan meals, build a basket and check out at participating banners; no list of participating banners, no word on pickup/delivery or PC Optimum points; Muse voice launched in Canada the same day (mobile, web, Mac). Single source: Meta agreed to pause issuing Class C shares during the AI-funding lawsuit (November 18) — a Delaware court-approved settlement bars issuing non-voting Class C shares (intended to fund AI expansion without diluting Zuckerberg's voting control) during pre-trial and trial; the restated charter and 2025 equity incentive plan revisions effective January 19, 2027 still stand; trial set for April 19–23, 2027 (Unite.AI only; see Unverified Items).
Sources: about.fb.com · unite.ai
xAI (Grok)
Single source: after the Irish DPC intervened, X has suspended using personal data collected by Grok Bot agents to train Grok (November 18, MLex 13:22 GMT) — Irish Data Protection Commissioner Dale Sunderland said the processing stopped on Monday (November 16); it covers third-party personal data collected by agents (including other people's information shared by users in messages); the DPC is also discussing similar training-data issues with other companies (MLex citing the commissioner only; no X/xAI confirmation; see Unverified Items). X and xAI settled with South Africa's Competition Commission with no compensation (November 18) — both sides withdrew appeals against the media and digital platforms market inquiry remedies; X committed to two years of publisher training (no funding commitment), xAI made a voluntary commitment to train publishers twice a year on using Grok settings and robots.txt to control content use for training; Google's November 2025 settlement of the same inquiry included a R688M support package. Follow-up: Grok Bot security incident details (eWeek, November 18 16:00 UTC, citing NewCore's November 17 report) — the 31 unauthorized read requests in five days from two Grok Bots were all intercepted by NewCore's security gateway and no data was taken; the source was Grok Bot's built-in background check-in (runs even with no one chatting) and it was not logged in Grok chat history; xAI support called it expected functionality, not a vulnerability; xAI added an admin switch on November 16 to turn off proactive messages (still on by default); eWeek also found that deleting a Grok Bot leaves saved files and logged-in browser sessions on its cloud computer (NewCore report and eWeek verification are a single-source chain).
Sources: mlex.com · techcentral.co.za · eweek.com
Microsoft (Copilot/Azure AI)
No major standalone release in the window — the November 18 ET Ignite schedule was breakout sessions: Entra session BRK343 (identity/high-assurance credentials, Conditional Access policy plan previews), Azure Cosmos DB session BRK367, Agent 365 governance session and others. Items appearing in November 18 media — Copilot Workspaces, Microsoft IQ, seven MAI models (Decision 1.1 GA, Code 2, Cyber 2 Flash), Agent 365 Business Analytics, Perception GA, HorizonDB GA, Cobalt 200 GA, Maia 300, GPT-6.2 and Meta's latest models joining Foundry, and Copilotforce with Salesforce (38+ prebuilt sales skills) — were all from the November 17 keynote, already reported; not repeated as new.
Sources: epcgroup.net · technologymagazine.com
NVIDIA
Post-earnings share reaction (see Top Stories 4). Board rules out Jensen Huang's return (November 17 ET earnings call = morning of November 18 Beijing Time) — Chairman Harvey Jones said the board unanimously accepted Huang's resignation and has "no intention" of restoring him as CEO; hours earlier Huang posted on X that he would be "deeply honored" to return if the board judged it in the interests of shareholders, employees, the company and the US; the board has hired Heidrick & Struggles to find a permanent CEO (internal and external candidates considered), with CFO Colette Kress as interim CEO. Congressional probe escalates (see Top Stories 4). Nvidia said on the earnings call it has stopped shipping licensed Hopper data-center products to China, with $2.4B of such sales to China in Q3 — a second-hand account of the call, not directly confirmed in the company's filing (see Unverified Items).
Sources: fool.com · aiaffairs.com
Mistral
Vibe desktop app launched for macOS and Windows (announced by Mistral on X on November 17, widely reported November 18) — previously web, CLI and mobile only; the desktop version can dispatch tasks to multiple agents in parallel and view them side by side, combining Vibe CLI local capability with Vibe Code Web cloud sessions, with local and remote agents in parallel, diff review and PR tracking; Linux version "coming soon" (no date); European developers could get three free months of Vibe Pro with code LECHONK (normally $14.99/month; the first batch of codes ran out the same day and was restocked). CEO Arthur Mensch said Mistral is preparing a European IPO, but not this year (November 18, Semafor event in Silicon Valley) — Europe is the "natural" listing venue, with no exchange or timetable named; he said the French state holds an indirect stake and the company is valued at $24B; most investors and half of customers are in Europe; he added that OpenAI and Anthropic compete with his own customers. Single source: Mistral is developing industrial engineering AI models for regulated industries (Pulse 2.0, November 17, republished November 18) — covering research, engineering and product development for semiconductors, manufacturing, automotive, aerospace and energy, based on Emmi AI acquired in early 2026, offered alongside Mistral Large 4 and deployable on customer-controlled infrastructure; no release date given (see Unverified Items).
Sources: testingcatalog.com · dailyaibrief.com · aiaffairs.com
Perplexity
Perplexity Brain can now automatically create and update personal agent skills (November 18) — from August 1 to November 14, 2026, Brain wrote 80,000+ personal skills for 20,000+ users; from September 4 to November 14 it made 100,000+ automatic updates to 30,000+ existing skills; in an eight-week randomized production test, 72.9% of users with auto-improvement on gained at least one new skill, versus 6.5% manually adding in the control group; tool-call failure rate fell 11% overall (about 7% excluding background runs); offline SkillLearnBench task success rose from 43.8% to 62.5%; now open to all US Max subscribers, on by default and switchable off. Computer's new Multiagent tier (up to 16 agents collaborating) was announced by Perplexity at 22:52 UTC on November 17 (06:52 Beijing Time on November 18, eight minutes before this window opened) and was already reported in the previous edition; not repeated.
Sources: unite.ai
China AI Companies
DeepSeek
Free-tier limits enter a second day with still no official response (see Top Stories 8). OpenRouter data showing V4.1 Flash-1017 at 59.2 trillion tokens in the week of November 9–15, ranked first globally, was cited in November 18 reports but is not new data for the day; the week-over-week growth figures conflict (mxcos says +6%, zennew says +801%; see Unverified Items).
Sources: 21jingji.com
Alibaba (Qwen)
Earnings preview (not a result): Alibaba reports quarterly results for the period ended September 30 before the US market open on November 19; consensus expects revenue of about 269.849 billion yuan (+8.9% year over year), with focus on whether cloud growth breaks 50%, flash-sale losses narrowing and AI capex. Single source: Citi said Qwen4-Max cuts per-task cost 58% and kept a Buy rating with a $193 target (reported November 18) — average cost per task fell from 26.58 yuan to 11.16 yuan (-58%); 2.4T total parameters, 71B activated; domestic API pricing 10 yuan per million input tokens / 30 yuan output, overseas $1.5/$4.5 — this is a Marsbit AI-generated summary of a November 16 Citi note, with no Citi original seen for cross-checking (see Unverified Items). Qwen4-Max remains closed-weights-first (restated November 18 from the November 16 launch): the flagship launched as a closed API with no weight-download timetable; only smaller Qwen4 derivatives were promised to stay open-weights. No new development on November 18.
Sources: news.marsbit.co · ad-hoc-news.de
Tencent (Hunyuan/Yuanbao)
Hy4 Flash weights open-sourced (see Top Stories 6). Q3 results (see Top Stories 6). WorkBuddy data (see Top Stories 6). Management remarks on the November 18 earnings call (see Top Stories 6).
Sources: 21jingji.com · news.17173.com
Zhipu (GLM) / MiniMax
Hong Kong rebound (see Top Stories 7). A CnYES report saying Zhipu "kept a JPMorgan Overweight rating" actually restates JPMorgan's November 3 downgrade (target cut from HK$2,000 to HK$1,200, Overweight maintained) — not a new rating on November 18. Zhipu's HK$4B buyback plan dates from a November 10 announcement, with no execution update in the window. MiniMax's H3-Image "50% off requests above 16,600,000 pixels" offer expired at 00:00 Beijing Time on November 18, a scheduled expiry with no new action.
Sources: gfe-desktop.cnyes.com · finet.hk
ByteDance (Doubao)
No important update in the window. No new November 18 release was found for Doubao, Seedance or Spell. Seedance 2.5 Flash (released November 12, API opened November 15) and the Spell Muse-like assistant plan (Bloomberg, November 11, said to launch in the coming weeks; Zhang Yiming said catching the frontier would take until 2029–2030 at the earliest) are both outside the window.
Moonshot (Kimi)
Single source: Kimi K3.1 Pro listed on OpenRouter — reported at 11:31 on November 18 at $2 per million input tokens / $10 output; 1M context, multimodal, open weights (only one blog report; no OpenRouter announcement seen for cross-checking; see Unverified Items). No new Kimi model or product launch in the window. The November 15–17 Web Summit remarks on shifting "from code to biology/chemistry/pharma research" and the Deep Potential BioMaster case (100 candidate genes screened in 90 minutes vs about 4 hours by hand) were published at 16:12 on November 17, outside this window.
Sources: anygen.xin
Meituan (LongCat)
No important update in the window. LongCat-2.5 weights were released under the MIT license on November 9–10 (1.6T total parameters, about 48B activated per token); related explainers were published November 15–16, outside the window.
Baidu (Ernie)
No important update in the window. Ernie Lightning and Ernie Mirror were released at Baidu World on November 12; Ernie 5.5 Flash Preview weights are planned for Apache 2.0 open-sourcing on November 25, with no new development in the window.
Manus
Minor: Manus became a Platinum Sponsor of Godot — announced by the Godot Foundation on November 17 and reported November 18; no project-level change announced and the scope of sponsorship was not detailed. Single source: a paid-promotion controversy report — a November 18 repost said Meta-owned Manus pays third-party agencies for "easy side hustle, $5,000 a month" style paid UGC ads on TikTok, Instagram and YouTube; a Manus spokesperson confirmed working with third-party agencies on paid creator programs, rejected exaggerated income claims and said it is reviewing the matter; Meta did not respond. The source is a thevergetoday.pages.dev mirror page, not theverge.com itself, and is uncorroborated (see Unverified Items).
Sources: gamedev.net · thevergetoday.pages.dev
Other China developments
China's AI export-control final catalog was not published in the window. The status quo: the October 16 draft Announcement No. 48 (40 proposed new technologies, 23 in AI; 14 banned, 26 restricted; frontier-model weight trigger at training compute ≥10^26 FLOPs, or parameters above 2 trillion, or a global top-ten model), with consultation ending November 15; the interim pre-export registration requirement has been in force since October 16. Follow-up on the Bloomberg report about Chinese firms curbing AI hacking skills: no new Bloomberg original appeared in the window, only November 17–18 reposts. Their key points: Moonshot removed cyberattack/vulnerability-research tasks from Kimi K3.1 Pro's mid-training data and RL environments, with its internal xdev-bench-hard score falling from 90.5% (unfiltered) to 81.0% (filtered, Moonshot internal benchmark, unaudited); Tencent deleted exploit-development, reverse-engineering and penetration-testing material before Hy4's November 4 open-sourcing, keeping defensive content; Alibaba's Qwen4-Max keeps weights closed; Zhipu took a different path, open-sourcing GLM-5.3 after a security review, with its most sensitive capabilities to sit behind "Cybersecurity Trusted Access" identity verification (no opening date). The reposts also say CrowdStrike reported hackers using DeepSeek models in attacks on organizations in 14 countries.
Sources: news.bloombergtax.com · longbridge.com
Policy & regulation
BAAJA (see Top Stories 3). Reactions to the BIS remote-compute interim final rule (follow-up): China's Foreign Ministry spokesperson Guo Jiakun said on November 17 the rule is typical economic coercion and long-arm jurisdiction and China will take necessary measures to defend its lawful rights, without naming specific measures; the Commerce Ministry said the same day it firmly opposes it and will take countermeasures if the US persists; neither ministry published a specific countermeasure list. Malaysia's trade minister said Kuala Lumpur has no intention of negotiating with Washington and expects data-center operators to comply; Singapore said it is studying the rule and has asked Washington about credible access arrangements; Brazil called it "not a ban." Industry: TSMC said on November 18 it is assessing the requirements and will comply; Quanta and Wistron expect no impact, Compal and Inventec say impact is limited. Timing confirmed: the rule will be published in the Federal Register on November 19 and take effect then, with a 60-day comment period and a 90-day transition for data centers already serving affected customers. Megaspeed probe (follow-up): Senate Banking Chair Tim Scott confirmed in a CNBC interview on November 18 that Huang and Lutnick have both agreed to testify, hopefully before Christmas; Scott said Commerce was warned as early as March to put Megaspeed on the entity list but did not act; the probe will also cover whether Chinese firms remotely accessed advanced chips via third-country data centers. On the House side: Huang agreed on November 17 to a closed-door transcribed interview on December 1 with the Foreign Affairs, Intelligence and China committees; Nvidia general counsel Tim Teter was asked to appear December 2; interim CEO Colette Kress was asked to brief by December 2; House Energy and Commerce Chair Brett Guthrie wrote to Kress demanding a briefing by December 2 on chip export controls and the Huang–Alice Huang relationship. Trump's Asia trip (see Top Stories 9). NDAA chip bills: no new development in the window. The latest status remains the November 12 Semafor report: Senate Majority Leader Thune and Armed Services Chair Wicker said the three chip-export bills (Chip Security Act, AI Overwatch Act, MATCH Act) have no deal yet to stay in the final NDAA text; Thune put the odds of passage as "high." House Democratic Leader Jeffries said on November 16 he will work to keep the three measures. The fourth bill, the Remote Access Security Act, has been partly covered administratively by the BIS interim final rule. EU AI Act: the European Commission opened a consultation on November 16 on the draft Article 27 FRIA template and invited deployers to test it voluntarily; expressions of interest to test are due December 4, public feedback by January 15, 2027, and the obligations apply from December 2, 2027. Regulatory sandbox rules: Commission Implementing Regulation (EU) 2026/2487 (November 10) sets out how AI regulatory sandboxes are established, run and supervised; member states must have national sandboxes by August 2, 2027; SMEs (including startups) get priority and free regulatory guidance. Copyright pressure: on November 17, 11 member states led by Germany, Spain and Poland jointly urged the Commission to adopt tougher AI copyright rules and reform creator opt-out enforcement; the earlier copyright consultation ended November 10, with a broader copyright package expected in Q2 2027 (single source: Medium column). Employer survey: Littler's 2026 European employer survey published November 18 (nearly 500 respondents) found the share of European employers cutting jobs because of AI doubled from 13% to 28%; 93% have taken at least one AI Act preparation step, but only a quarter have named an internal compliance owner.
Sources: techtimes.com · aiworldwire.org · stocktwits.com · aiaffairs.com · neoteo.com · businesswire.com
Data centers & energy
Google–Quaise superhot geothermal: a power purchase agreement signed November 17 (ET), 50 MW, worth over $500M according to the WSJ, for Project Obsidian beside the Newberry volcano in Oregon, with power expected from 2030 at the earliest, supplying three Google data centers in The Dalles; the first commercial offtake agreement for the technology. STN–ChronoScale Odessa AI campus: a 50/50 joint venture announced November 18, each side investing $390M, with the first 65 MW phase ordered and deploying in the first half of 2027 and up to 350 MW across three phases; $750M of financing expected to close November 18; the end customer is an unnamed frontier AI lab. Giga Energy: a 10-year lease worth about $1B signed November 17 with an unnamed US AI cloud vendor, 60 MW Project Clay (Hunt County, Texas), operating from spring 2027. The TensorWave–OpenAI six-year deal for 457 MW of AMD compute was reported on November 16.
Sources: morningstar.com · dealroom.co · superpowerdaily.com · aiagentupdate.com
Capital markets
US stocks on November 18 (see Top Stories 5). Hong Kong AI twin stars rebound (see Top Stories 7). Anthropic IPO (see Top Stories 1). AMD completed its acquisition of inference-chip company Taalas (completed November 17; the technology joins the accelerator roadmap).
Sources: ad-hoc-news.de
Security incidents
NSA NSANet outage (follow-up): full recovery had still not been confirmed as of November 18. Director of National Intelligence Jay Clayton said on Fox Business on the afternoon of November 17 that the outage came from patching the PHP2Shell vulnerability (CVE-2026-109618, CVE-2026-109619, CVSS 9.2) and some systems did not come back as scheduled; the NSA called it a temporary degradation during routine maintenance with mission-critical capability unaffected; Fort Meade was limited to essential personnel. Officials and sources agree there is no sign of a cyberattack. The vulnerability was disclosed on October 21 by OpenAI, Wiz and California researchers using AI tools. Grok Bot security incident (follow-up; see xAI tracking).
Sources: thebuyerspulse.com · thedefensenews.com
Worth Reading
1."Anthropic reportedly prices IPO at $83 a share, above projected range" — TechCrunch, 2026-11-18
How Anthropic priced its IPO at $83, and what is known and still unknown — the verification source for Top Story 1.
Sources: techcrunch.com
2."Introducing Gemini 4 Neon" — Google Blog, 2026-11-18
Gemini 4 Neon's positioning, pricing and launch scope — the verification source for Top Story 2.
Sources: blog.google
3."Senate Permitting Legislation Advances, Setting up Final Push" — Bloomberg Law, 2026-11-18
BAAJA's final cloture vote of 65–31 and the final-passage timetable — the verification source for Top Story 3.
Sources: news.bloomberglaw.com
Unverified Items
1.
Anthropic IPO final share count and total proceeds at the $83 price: Reuters and Bloomberg both say the deal was upsized, but the post-upsizing share count, total proceeds and fully diluted valuation (about $1.5T is an estimate) had not been confirmed in an official filing by the end of the window; the Form 424B4 is pending.
2.
Meta agreed to pause issuing Class C shares during the AI-funding lawsuit: seen only at Unite.AI, with no court filing or Meta confirmation.
3.
After the Irish DPC intervened, X suspended using personal data collected by Grok Bot agents to train Grok: seen only in MLex citing the DPC commissioner, with no X/xAI confirmation.
4.
Mistral is developing industrial engineering AI models for regulated industries: seen only at Pulse 2.0, with no Mistral confirmation.
5.
Kimi K3.1 Pro listed on OpenRouter at $2/$10 pricing: seen only in a single blog post at anygen.xin, with no OpenRouter announcement.
6.
Qwen4-Max "per-task cost down 58% (26.58 → 11.16 yuan)" and Citi's $193 target: seen only in a Marsbit AI-generated summary of a Citi note, with no Citi original seen.
7.
Alibaba/Tencent Hong Kong closing prices on November 18: Longbridge lists both closes as HK$372.8 with the same -0.85% move, which looks like a copy error in the source; prices and moves are uncorroborated and should not be quoted directly.
8.
Hy4 Flash total parameter count: Tencent/ITHome and other Chinese sources say 224B, while an MLQ English report says 220B; 224B is used as the main figure.
9.
Gemini 4 Hydrogen/Argon returning 429 on Vertex AI's EU endpoint: only a Google developer forum user post from November 18, with no Google confirmation.
10.
Mistral model API pricing changes (November 18): olud.ai says Medium 3/3.1 input rose 10% to $0.44 per million tokens and Small 3.2 24B fell 20% to $0.08; no Mistral pricing page corroboration.
11.
Whether Senator Merkley's point of order against BAAJA was raised before final cloture and how it was resolved: only a pre-vote preview was seen, with no result reported.
12.
The claim that the Xi–Trump bilateral will focus on "AI semiconductors and rare-earth supply chains": seen only at el7.ai, with no Reuters or official confirmation — single source.
13.
Perplexity's annualized revenue passing $1B: CEO Aravind Srinivas said at a Semafor event on November 16 that it reached a $1B ARR run rate that day (not realized annual revenue); the event itself is outside the window and the claim is the CEO's own, unverified independently.
14.
Nvidia "has stopped shipping licensed Hopper data-center products to China, with $2.4B of such sales to China in Q3": a second-hand account of the earnings call, with no direct confirmation in the company's filing.
15.
Eleven European member states urging tougher AI copyright rules: seen only in a Medium column, with no European Commission or member-state statement.