AI Morning Brief · Nov 13, 2026
Top Stories
1.Jensen Huang resigns as NVIDIA CEO; CFO Colette Kress named interim CEO, Harvey C. Jones becomes first board chair
NVIDIA announced on Nov 12 that Huang has resigned as president, CEO and director, effective immediately; CFO Colette Kress will serve as interim CEO while retaining her roles as executive vice president and CFO, and former Synopsys CEO Harvey C. Jones was named the company's first-ever board chair; the board accepted Huang's resignation on the evening of Nov 11 and an 8-K has been filed; the announcement gave no reason for the departure and disclosed no severance terms; Huang still owns about 3.6% of the company and is its fourth-largest shareholder, but no longer sits on the board; the board has begun a formal CEO search covering internal and external candidates, with Ian Buck and Arm CEO Rene Haas seen as potential contenders; Q3 earnings remain scheduled for Nov 17. The trigger: the Wall Street Journal reported on Nov 11 that the board was preparing to oust Huang and would decide that night, and the New York Times reported the same day that Huang had a second extramarital affair, with Kimberly Powell, a vice president in the medical business who reported to him (an earlier affair involved Alice Huang, a woman US officials suspect of spying for China); the board launched an independent investigation on Nov 8.
Sources: reuters.com · techcrunch.com · usatoday.com
2.Google launches Gemini 4 Hydrogen, priced at $1.50/$7.50, rolling out everywhere
Launched Nov 12, Gemini 4 Hydrogen targets fast everyday coding and agentic workflows, with a 1M-token context window and API pricing of $1.50 per million input tokens and $7.50 per million output tokens; it scores 51 on the Artificial Analysis Intelligence Index at the default high setting (above Argon at 50) and 54 at xhigh; it is rolling out across the Gemini API, AI Studio, the Gemini app, Search and Gemini for business, with Antigravity access for AI Pro/Ultra subscribers and the Orchestrator (which switches dynamically between Carbon and Hydrogen) exclusive to Ultra; it replaces the Gemini 3.8 Flash tier. Note: cloud docs label it "short-term GA, with at least 45 days' notice before retirement", and it does not support RAG Engine, fine-tuning or the Live API.
Sources: blog.google · unite.ai
3.Microsoft 365 suffers a global outage of over five hours, knocking out Outlook, Teams and Copilot
Microsoft 365 went down globally from 13:26 ET on Nov 12, affecting Outlook, Teams, SharePoint and Microsoft 365 Copilot; Downdetector recorded a peak of more than 97,000 Outlook outage reports; Microsoft said some servers became unexpectedly unavailable, overloading the remaining infrastructure; OpenAI and Anthropic reported errors in their Microsoft 365 connectors (new connections could fail; Claude for Outlook was down), though ChatGPT and Claude themselves stayed up; service was still not fully restored by the evening of Nov 12; a separate, independent Teams outage ran 09:45–10:53 ET.
Sources: securityaffairs.com
4.Nonprofits sue OpenAI, Anthropic, Google, Meta and xAI, seeking a halt to frontier AI development
Nonprofits including the Irreplaceable Education Fund, Public Citizen and the Heat Initiative sued the five companies in San Francisco on Nov 12, invoking California's unfair competition law and asking the court to order a halt to development of the most advanced models, including ChatGPT, Gemini and Claude, on the grounds that the companies' own executives have warned the technology could threaten human civilization; the complaint quotes Anthropic CEO Dario Amodei and Elon Musk. Background: Florida sued OpenAI in June and in September asked that new model development be barred pending external oversight; OpenAI says it has paused training of its strongest models until safety measures are in place.
Sources: mymixfm.com
5.Anthropic IPO follow-up: roadshow book covered on day one, bankers weigh pricing above a $1.5T valuation, but filed terms unchanged
After the roadshow opened on Nov 10, the order book was fully covered on the first day at the $70–$78 range, with sources describing it as "multiple times oversubscribed"; the FT and Bloomberg report bankers are considering lifting the valuation above $1.5T. But the filed S-1/A terms are unchanged: 825M shares, $70–$78, pricing expected Nov 18 and Nasdaq trading on Nov 19 under ticker ANT; the $26B concurrent private placement (Google and NVIDIA at about $10B each, AMD about $2.5B, and others) is also unchanged; the filing also disclosed a Q3 operating loss of $567M (after a $974M profit the prior quarter) and up to $4.3B in additional Microsoft compute commitments.
Sources: opentools.ai · aiagentupdate.com
6.Baidu World 2026: ERNIE Lightning deep-thinking model, ERNIE-Wizshot-ME micro-expression model, a "Token Exchange" platform, and a 1,024-card Tianchi Supernode
At Baidu World 2026 on Nov 12, Baidu launched the ERNIE Lightning deep-thinking model and the ERNIE-Wizshot-ME micro-expression model (which can produce 4–15 seconds of 1080p video of real people, listed on Qianfan at 0.7 yuan per second), both available via the Qianfan API; it also launched "Token Exchange", a trading platform for large-model tokens (over 100 resources onboarded in its first month, with more than 50M calls and 700B tokens per day); Baidu Cloud president Shen Dou announced a 1,024-card Supernode called "Tianchi" for 2027, claiming it can train 5-trillion-parameter models and pairing it with Kunlun M300 chips (the 5-trillion figure is Baidu's own claim, with no independent verification).
Sources: unwire.pro · ai-club.cn · quantisnow.com
7.ByteDance: Zhang Yiming says catching up with OpenAI and Anthropic will take until 2029–2030; a Muse-style personal AI assistant is coming in weeks
At a recent Seed team all-hands, Zhang Yiming said ByteDance's Seed pre-training dataset is orders of magnitude smaller than those of top rivals, and that the company can hope to compete at the frontier with OpenAI and Anthropic only in 2029–2030 at the earliest; Bloomberg separately reports ByteDance is preparing a personal AI assistant similar to Meta's Muse, expected to debut within weeks, integrating the internal "Spell" project (started in April) with the Doubao conversational team; Doubao had 419M monthly active users in September, up 143.6% year over year, making it China's top AI-native app (Zhang's remarks come from an internal meeting and have not been confirmed by ByteDance).
Sources: bi.run · techritual.com · zgeo.com.cn
8.Zhipu and MiniMax plunge in Hong Kong on Nov 12; HSBC slashes Zhipu's target price by 64%
Hong Kong closed lower on Nov 12, with the Hang Seng down 1.26% at 24,008.55 and Hang Seng Tech down 2.36% at 3,931.93; Zhipu (02513.HK) closed at HK$454, down 9.05% from its prior close of HK$499.2, after falling more than 5% intraday (HK$470.4 at 11:40); MiniMax-W (00100.HK) closed at HK$166, down 7.83% from HK$180.10, after falling 5.66% intraday to HK$169.9; HSBC cut Zhipu's target price from HK$1,375 to HK$500, keeping a Hold rating. The conflicting reports on the Nov 11 closes are now clarified: Zhipu closed Nov 11 at HK$499.20 (up 3.57%) and MiniMax at HK$180.10 (up 4.04%), with Finnhub daily data consistent with HSBC's warrants page.
Sources: finnhub.io · stockanalysis.com · tw.tradingview.com · stock.10jqka.com.cn
9.Fireworks AI follow-up: Notion confirms chat transcripts of about 1% of users were stolen
TechCrunch reported on Nov 12 that Fireworks AI's latest security bulletin confirms attackers entered its development systems through a cloud provider and stole customer billing information and some datasets; Notion confirmed the stolen material includes Notion AI chat records of users of its "Insights" feature, affecting about 1% of Notion users; Notion disconnected from Fireworks on Nov 3 and deleted the related data. This is the second breach at Fireworks in two months (the first was in October).
Sources: techcrunch.com
10.US stocks on Nov 12: NVIDIA closes down 4.34% at $166.19, AMD rises 5.86%; major indexes fall about 0.4% and the 10-year Treasury yield tops 5.1%
NVDA opened at $151.01, hit an intraday low of $148.80 (about 14.4% below the prior close) and a high of $167.14, closing at $166.19, down 4.34% (-$7.54), on volume of 642M shares; it recovered slightly after hours to about $167.48 (+0.78%, Finnhub); since the Nov 6 close of $186.64, about $658B in market value has evaporated in a week (secondary-source figure). The S&P 500 fell 0.35% to 7,592.40, the Dow 0.36% to 52,297.09 and the Nasdaq 0.39% to 25,366.09; the Russell 2000 dropped 1.66% and the VIX rose 6.9% to 18.70; the main drivers were the NVIDIA CEO resignation shock and inflation worries, with the 10-year Treasury yield rising to 5.10%–5.12% (from 4.98% on Tuesday) and Brent crude up to about $94–96; October PPI is due Nov 13. Chip stocks diverged: AMD closed at $585.30, up 5.86% (intraday high $594), seen by the market as a beneficiary of NVIDIA's turmoil; Micron closed at $798.44, down 1.59%, giving back part of the prior day's +11.89%; TSMC's ADR was about $406.27, down about 1.7% (single source, unverified).
Sources: finnhub.io · sfntoday.com · stockmarketwatch.com
Platform & Model Tracking
OpenAI (GPT/ChatGPT)
GPT-6 Luna API inference processing lands in India — announced Nov 12, eligible Indian enterprises and developers can now complete GPT-6 Luna inference processing and storage inside India via the Responses and Chat Completions APIs; India becomes the third market for inference localization after Australia and Japan; at launch only GPT-6 Luna is covered, requiring Zero Data Retention and standard pay-as-you-go billing, with no hosted tools yet. OpenAI held its "Cyber at the Frontier" keynote at 10:00 PT on Nov 12 with no new public model; audience reports say slides showed an unreleased security model scoring 77% on ExploitGym, above GPT-6 Cyber's 52.6% (another early report put GPT-6 Cyber at 48.5%; both figures are unverified). GPT-6.2 remains a rumor, with only leaked model IDs and screenshots and no listing on OpenAI's pricing page, model list or changelog. ChatGPT Finances expanded to the Mac desktop app (launched Nov 11, reported Nov 12; US Free/Go/Plus/Pro users, read-only bank connections via Plaid). The proposed Nov 12 cutoff of OpenAI models in Cursor did not take effect and remains unconfirmed: as of 9:00 AM Central on Nov 12, neither company had confirmed it; Cursor's site still lists GPT-5.6 Sol/Terra/Luna, and OpenAI's help page still calls Nov 12 a "proposed" transition date, with the formal date to be announced once both sides confirm. Two brief incidents on OpenAI's status page on Nov 12 (Workspace Agent failing to start; elevated GPT-Image-2.5/2.6 error rates) were resolved (single source, unverified).
Sources: newzora-pied.vercel.app · enjoy.dev · macobserver.com
Anthropic (Claude)
IPO follow-up — order book covered on day one, bankers weigh pricing above a $1.5T valuation; Q3 operating loss of $567M; up to $4.3B in additional Microsoft compute commitments (see Top Story 5). Single source: Claude Fable 5.2 fast mode is now on the Messages API ($16/$80 per million tokens, up to 2.5x output speed) and has landed on Google Cloud and Copilot Cowork/Studio — reported by a single API-aggregation briefing, unverified. No new developments in the South Korean PIPC investigation; Sonnet 4.5 deprecation unchanged (deprecated Sep 30, hard retirement Nov 30, with Sonnet 5.5 as the recommended replacement).
Sources: opentools.ai · aiagentupdate.com
Google DeepMind (Gemini)
Gemini 4 Hydrogen launched (see Top Story 2). Google AI Plus loses Gemini Pro on Nov 13 (today, as planned, with no hands-on confirmation yet that the change has taken effect): per an email Google sent subscribers in mid-October (disclosed by 9to5Google), from Nov 13 choosing the Pro model in the Gemini app and Search AI Mode is limited to AI Pro and above; AI Plus ($4.99/month) keeps Flash and Flash-Lite, and existing Pro conversations can continue by switching models; Google's help page confirms the change without a public date, and the plans page now lists only Flash. AI Plus will also lose Claude in Antigravity on Nov 18. Gemini Decision 1 and the Decisions API launched Nov 11 and were covered in the previous issue.
Sources: blog.google · enjoy.dev · notebookcheck.net
Meta
The five-company lawsuit seeking a halt to frontier AI development (see Top Story 4). The publishers' class action against Meta has been verified as an old case: Elsevier, Cengage, Hachette, Macmillan, McGraw Hill and author Scott Turow sued Meta and Mark Zuckerberg (Elsevier Inc. et al. v. Meta Platforms, Inc., No. 1:26-cv-03689) on May 5, 2026, in the Southern District of New York, alleging Meta copied millions of works from pirate sites including LibGen to train Llama; the Nov 11 23:31 UTC timestamp was an aggregator page's update/crawl time — there is no new publishers' class action in this window. The nearest real procedural step is an Oct 27 joint case management statement (the Kadrey case is set for trial May 3, 2027), which falls outside this window. No new Llama or Muse model, pricing or availability release in the window.
Sources: mymixfm.com · reuters.com
xAI (Grok)
Grok 4.8 has no official confirmation; only social-media posts claim a quiet limited rollout to Bot users. Musk said on Nov 10 that pre-training was complete, the model was in RL, and a soft launch to Bot users would come later in the week; on Nov 12 social posts said it had been quietly deployed to a limited set of Bot users, but xAI's news page, model directory and API show no Grok 4.8, with no model ID, pricing or benchmarks — treat as planned, not confirmed as shipped. No Grok 5 release and no pricing change in the window.
Sources: explainx.ai · grokshelf.com
Microsoft
Microsoft 365 global outage (see Top Story 3). Reminder: UK Copilot tenants will have Claude switched on by default from Dec 9 — tenants that have not made a choice will get Anthropic Claude by default, and admins must opt out before Dec 9; the data is processed by Anthropic in the US, and EU/EFTA tenants are unaffected. No new Copilot or Azure AI model, pricing or availability announcement in the window.
Sources: securityaffairs.com
NVIDIA
Jensen Huang resigns; Colette Kress named interim CEO (see Top Story 1). Shares closed down 4.34% at $166.19 on Nov 12 (see Top Story 10). President Trump said on Nov 12 that former Apple CEO Tim Cook is the only person who should replace Huang, praising Huang's work; this is the president's personal view, with no response from NVIDIA's board and no comment from Cook. Senator Jack Reed, the top Democrat on the Senate Armed Services Committee, said on Nov 12 that Congress should assess whether NVIDIA is "too big to fail", demanding full transparency about the complex financing deals supporting the AI ecosystem and an assessment of the impact on the US banking system and credit markets (Reuters exclusive; a senator's statement, not a formal investigation). Q3 earnings are due Nov 17; whether the Nov 18 closed-door briefing with Huang will take place, and who would attend, is unconfirmed (see Unverified Items).
Sources: reuters.com · reuters.com · reuters.com
Mistral
No major update in the window. Orange launched a free app that lets users compare Claude, Gemini, Mistral and GPT under one account (a €10/month tier unlocks premium models); this is a distribution partnership, not a model release (single source, unverified).
Perplexity
Decider v1.3 released on Nov 12 — Decider v1.3 and Decider Flash v1.3 answer yes/no and multiple-choice questions with probabilities rather than text, adding a Locate feature that returns exact matching text spans with start and end character positions (for named-entity recognition, PII detection and extractive QA); Perplexity says the large model averages 91.6% accuracy across 63 query-classification tasks and agrees with frontier-model judges 93.98% of the time on a sample of 1M real API requests (v1.2 was 93.68%); built on Qwen3.8-27B (Flash on Qwen3.5-4B), with Apache-2.0 weights downloadable and available on Hugging Face, OpenRouter and the Perplexity Decisions API. Conflicting reports on new funding continue (see Unverified Items).
Sources: unite.ai · dailyaibrief.com
China AI Companies
Alibaba (Qwen)
Qwen4-Max preview details confirmed — checks of Alibaba Cloud's official docs on Nov 12 confirm Qwen4-Max-Preview has a 1M-token context window and a maximum output of 393,216 tokens (reasoning process and final answer share the allowance); it is available only through Token Plan subscriptions, with no pay-as-you-go API; the preview runs Nov 11 10:00 to Nov 25 24:00 Beijing time, with the model unchanged during the window and retired at the end; Alibaba has published no parameter count, model card, benchmark scores, per-token pricing or weights, and there are no Qwen4 weights on Hugging Face; Token Plan credits are consumed at as little as 20% of standard price during the preview. Follow-up (single source): Alibaba said on Nov 11 that its chip subsidiary T-Head has confidentially filed an A1 form with the Hong Kong Stock Exchange, applying to list H shares on the main board — a confirmation of the earlier spin-off report, with the filing itself unverified. Alibaba shares fell 4.97% to HK$95.6 in the Nov 12 morning session (reported only in a Chinese-language roundup, unverified).
Sources: aiagentupdate.com · aifrontierpost.com · xueqiu.com
Baidu (ERNIE)
Baidu World 2026 — two application models, the "Token Exchange" platform and the Tianchi Supernode (see Top Story 6). No new development on ERNIE 5.5 Flash Preview; weights remain on schedule to be open-sourced before Nov 25.
Sources: unwire.pro · ai-club.cn
Tencent (Hunyuan/Yuanbao)
Hy4 weights still not released as of Nov 12 — the Hy4 full release (Nov 4) remains API-only; tencent/Hy4 on Hugging Face is still not publicly accessible, and neither weights nor a system card have been published, despite the Nov 4 "planned open source soon" commitment; Hy4 Flash has also not appeared.
Sources: pchome.tech
Zhipu (GLM) / MiniMax
MiniMax's H3-Image tops both Artificial Analysis open-weight image lists — the Nov 12 rankings show H3-Image, released and open-sourced late on Nov 11, ranking first among open-weight models in both text-to-image and image editing; the model generates a native 4K image (4096×4096, 16MP) in about 10 seconds, roughly 5x faster than mainstream models, and supports up to 16 reference images; this item is a follow-up to the Nov 11 release. Both stocks plunged in Hong Kong on Nov 12 (see Top Story 8). Single source: Webull cites Zhipu ARR of $1.6B at the end of August (up 60% from $1.0B in early July), over $2.0B on a weekly run rate and a $2.4B target by end of 2026 — no second source, and the publication date is unclear (see Unverified Items). Per a Pandaily reading, the H3-Image license is MiniMax's own community agreement (not applicable in the US, EU, UK or South Korea; companies with over $1M in annual revenue need a separate written license) — single source, unverified.
Sources: news.marsbit.co · cn.tradingview.com
ByteDance (Doubao)
Zhang Yiming's 2029–2030 timeline and the Muse-style personal assistant (see Top Story 7). Rumors about the departure of Seed chief Zhou Chang remain unverified.
Sources: bi.run · techritual.com
DeepSeek
Follow-up on the second funding round — Caixin reported on Nov 12, citing multiple deal sources, that DeepSeek is about to close a second round of about 100B yuan, up from the previously planned 50B yuan (nearly doubled in under half a month), while the pre-money valuation remains about 500B yuan, not raised with the larger size; if it closes, the two rounds total about 150B yuan; CATL plans to invest about 10B yuan in this round after about 5B yuan in the first, with Tencent, IDG and other existing investors adding more, and new investors including Shixi Capital, SMIC Juyuan, CPE, Legend Capital and Geely. The 105.5B yuan / 615.5B yuan post-money "completed" figure still appears only in media such as 21st Century Business Herald, and DeepSeek has issued no confirmation; the two figures (100B vs 105.5B yuan) remain in conflict and unverified. A separate claim that companies linked to Alice Huang invested in DeepSeek and promised NVIDIA chip access appears only in a single KuCoin aggregation flash, with no mainstream corroboration — not used as news (see Unverified Items). No new model release and no official funding news in the window.
Sources: okx.com · jiougeo.cn
Moonshot AI (Kimi)
Menlo Ventures report (single source) — the VC firm's latest "State of Enterprise AI 2026" report says global enterprise AI spending is expected to reach $176.8B in 2026; 89% of US enterprises' open-model spending goes to Chinese models, with Moonshot's Kimi at 61%, DeepSeek at 15% and Zhipu GLM at 11%; OpenRouter data shows Chinese models' token calls exceeded US models' for 27 consecutive weeks through the week of Oct 26–Nov 1; Zhipu's GLM-5.3 Flash had about 15 trillion calls in the last 7 days, second on the platform; the report also says open models' share of token consumption rose from 11% to 19%, while their share of spending fell from 6% to 4% (single source — a 21st Century Business Herald write-up of a Science and Technology Daily reading of the report; the underlying report data is unverified). No new Kimi K3.1 release in the window; K3.1 Pro weights were released Nov 5 (outside the window).
Sources: 21jingji.com
Meituan (LongCat)
No new developments in the window; LongCat-2.5 weights were open-sourced on Nov 9–10 (covered earlier); a Nov 11 follow-up data point says LongCat-2.5 has about 109B cumulative tokens and about 4,700 unique users on OpenCode (single source).
Sources: m.aitntnews.com
Manus
No new developments found in the window.
Other China developments
No new news on China's AI export-control draft; the comment period still ends Nov 15, and Reuters cited sources in an Oct 28 report (outside the window) saying the proposal is expected to be formally adopted by year-end.
Sources: uanalyze.com.tw
Policy and regulation
BAAJA follow-up — no final vote in the window; the Senate is in recess for Veterans Day (no votes Nov 11–13), with the final vote expected in the week of Nov 16, at the earliest Nov 19. Confirmed procedure: a 73–25 cloture vote to begin debate on Nov 9 and a 62–21 motion-to-proceed vote on Nov 10; next comes amendment debate (Democrats plan tribal-rights/environmental amendments, Republicans plan amendments protecting farmers from transmission-line land takings), followed by another 60-vote cloture and a simple-majority final vote; even if the Senate passes it, the House must agree to the Senate text, and that procedure is unconfirmed. Three senators pressed Commerce Secretary Lutnick on Nov 12 (single source, unverified): Warren and Coons wrote to Lutnick, Rubio and Miller demanding preservation of all NVIDIA chip export records and confirmation of a freeze by Nov 17; Wyden and Neal opened an inquiry into approvals of advanced NVIDIA chip sales to China, questioning the H200 decision and the constitutionality of a 25% export levy; Cotton demanded that Chinese memory maker CXMT be added to the Entity List; the backdrop is a report that Lutnick has delayed executing Trump's three approvals (May, August and Nov 5) of a crackdown on diversion, with the Commerce Department yet to act; about 10 Chinese companies have been approved to buy H200, while Blackwell remains banned. EU AI Act: the transparency guidance was published in the Official Journal on Nov 12, with the Dec 2 new bans 20 days away — the European Commission's guidance on Article 50 transparency obligations (adopted Nov 5, first published Jul 20) was published as C/2026/6005 in the EU Official Journal on Nov 12 (43 pages, 155 items), covering chatbot disclosure, machine-readable marking of AI-generated content, emotion-recognition notices, deepfakes and public-interest AI text labels; the guidance itself is non-binding, and breaching Article 50 carries fines of up to €15M or 3% of global turnover; from Dec 2, two new obligations take effect: generative systems placed on the market before Aug 2, 2026 must meet the Article 50(2) machine-readable marking rule, and new bans cover generating non-consensual intimate images of real people ("nudify" apps) and child sexual abuse material, punishable by up to €35M or 7% of global turnover.
Sources: beyondnuclear.org · theapexwire.com · ppc.land
Data centers and energy
Epoch AI dataset shows top AI data centers at 15.1 GW of compute, and Exowatt launched a new solar-thermal system (single source, unverified): Epoch AI's dataset of the 96 largest AI data centers (downloaded Nov 12) shows combined IT power rising from 2.0 GW at the end of 2024 to 15.1 GW in November 2026, projected at 38 GW by end-2028; the largest is Amazon's New Carlisle campus (used by Anthropic, about 1.04 GW), followed by xAI's Colossus 2 in Memphis (946 MW); Exowatt launched its P3 Gen 2 modular solar-thermal power system and a 1 MW validation unit, Exobloc, on Nov 12, and plans the 2.4 GW Athena data center project in West Texas (80% renewable). Nebius' Q3 missed expectations and dragged AI cloud stocks down, with CoreWeave giving back gains: Nebius Q3 revenue was $980.8M (+571% YoY) but cut its 2026 revenue guidance to the low end of $3.0–3.4B (some third-party capacity delayed to H1 2027, and the New Jersey Microsoft site delayed by power constraints), raised annual capex to about $27B (from $20–25B) and cut its annualized recurring revenue guidance, falling about 19% intraday on Nov 12 (low $118.81); CoreWeave closed at $56.60, down 6.86%, giving back part of the prior day's +24.7%; IREN also fell; Huang's resignation was another driver of the sell-off.
Sources: ai-magazine.com · marketspeaker.com
Capital markets
US stocks on Nov 12 (see Top Story 10). Anthropic IPO (see Top Story 5). TSMC's board approved up to $1B in investments in Anthropic and OpenAI (follow-up; the resolution was Nov 10, disclosed Nov 12): up to $500M to subscribe to Anthropic common stock at the IPO price and up to $500M for OpenAI preferred stock, both caps rather than completed deals; the board also approved up to $500M to set up a semiconductor investment company and about $40.93B in capacity capex; Anthropic's IPO range is $70–$78 with a valuation of up to about $1.4T, plus a $26B post-listing private placement (Google and NVIDIA about $10B each, AMD $2.5B, Broadcom, Micron and Microsoft about $1B each, TSMC $500M), with pricing expected Nov 18 and trading from Nov 19; analysts say the Huang turmoil has not yet affected TSMC's orders. New funding and M&A disclosed in the window, all single source and unverified: Mirendil's $1B Series A (San Francisco agent-tools company, led by Kleiner Perkins and Clearlake Capital, which claims one researcher pre-trained the Mirendil-Base coding model from scratch in 3.5 months on under $10M of compute); UFORCE's $700M Series A at a $6B valuation (Ukrainian-founded defense tech, with Valor Equity Partners, the UK National Wealth Fund and the US-Ukraine reconstruction fund); Agile Robots' $500M round at an $8.5B valuation (led by SoftBank, with NVIDIA and Hitachi); Luminal's $25M Series A (AI inference compiler, led by Standard Capital); Databento's $100M Series C at a $1.1B valuation (led by Fifth Down Capital). Note: Fractile's $720M Series C (at a $7B valuation) was Nov 11 and was covered in the previous issue; the roughly £540M figure repeated by UK media on Nov 12 is the same round converted, not new funding.
Sources: communicationstoday.co.in · crosswalk.to · techstartups.com
Security incidents
Fireworks AI's second breach affecting Notion and Cursor (see Top Story 9). Palo Alto Networks Unit 42: Chinese attackers used AI to autonomously breach 19 organizations, succeeding at 10 (single source, unverified) — Unit 42 reported on Nov 12 that a Chinese attack group paired the open-source Cairn penetration-testing framework with Zhipu GLM 5.2 and DeepSeek V4 Flash/Pro to run 25 autonomous intrusion operations against at least 19 organizations in August–September 2026, breaching 10, stealing dozens of valid credentials, over 1M records from one organization, military/government identities, persistent admin accounts and unauthorized cloud storage access, and tampering with live ballots in one industry-standards vote. Chainstack disclosed an AI-driven attack leaking customer data (single source, unverified): disclosed Nov 12, the attack began Oct 26 (via default admin credentials on a development Grafana) and was contained Oct 31; customer names, emails and password hashes leaked, and node credentials must be rotated; the incident itself falls outside the window, the disclosure inside it. No new developments on the NEAR AI SGLang breach or the leak of 2,539 GitHub secrets via Claude Code settings files. No new developments in the South Korean PIPC investigation into the Anthropic Claude enterprise account data leak.
Sources: pivotnews.ai · coinpost.ai
Worth Reading
1."Nvidia's board is reported to be preparing to remove Jensen Huang; no decision announced" — TheDesk, 2026-11-12
A disciplined item-by-item separation of what is confirmed versus merely reported, the most restrained account of the NVIDIA turmoil so far.
Sources: thedeskresearch.com
2."OpenAI's November 12 Cursor cutoff is still unconfirmed. How to keep using GPT there" — enjoy.dev, 2026-11-12
Checks both companies' official sites as of 9:00 AM Central on Nov 12, confirms the cutoff did not take effect, and lists workarounds including bring-your-own API keys.
Sources: enjoy.dev
3."Cyberattack on Fireworks AI allowed hackers to access Notion customers' chat transcripts" — TechCrunch, 2026-11-12
The full account of Fireworks' second intrusion, the exposure at Notion and Cursor, and the types of data leaked.
Sources: techcrunch.com
Unverified Items
1.
NVDA's Nov 12 close: ChartExchange records $164.27 (-5.45%), conflicting with Finnhub and multiple closing summaries at $166.19 (-4.34%); the Finnhub figure is used for now.
2.
TSMC's ADR close of about $406.27 (about -1.7%) on Nov 12 rests on a single MarketBeat table, unverified.
3.
The three senators' letters pressing Secretary Lutnick: only two secondary sources, with no confirmation from Reuters, Bloomberg or other major outlets.
4.
Claims that the Senate "passed the GAIN AI Act / AI chip export limits": the articles found carry inconsistent dates and contradict a Nov 12 Tikr piece saying the chip provisions did not make it into the House version; not usable as a new development in this window.
5.
Huang's Nov 18 closed-door briefing: no public record of a reply before the Nov 11 17:00 ET deadline, and with his Nov 12 resignation it is unclear whether the briefing will happen or who would attend.
6.
Grok 4.8 quiet rollout: only social posts and aggregation headlines, with no confirmation from xAI.
7.
The unreleased OpenAI security model scoring 77% on ExploitGym: audience reports only, with no official post; the GPT-6 Cyber comparison figure appears as both 48.5% and 52.6% in different reports.
8.
The Unit 42 autonomous-intrusion report: single source (the report relayed by one outlet), unverified.
9.
CoreWeave's Q3 figures: single-source aggregation, with no original Reuters or Bloomberg link found.
10.
Claude Fable 5.2 fast mode on the Messages API: a single API-aggregation briefing, unverified.
11.
Mirendil's $1B Series A, UFORCE's $700M Series A, Agile Robots' $500M round, Luminal's $25M Series A and Databento's $100M Series C: all single source or aggregator relays, unverified.
12.
DeepSeek's second funding round: the 100B yuan (Caixin) and 105.5B yuan (21st Century Business Herald and others) figures coexist, with pre-money valuation of about 500B yuan versus post-money 615.5B yuan; no official confirmation.
13.
Claims that companies linked to Alice Huang invested in DeepSeek and promised NVIDIA chip access: a single KuCoin aggregation flash, with no mainstream corroboration.
14.
Menlo Ventures report data: single source (a 21st Century Business Herald write-up of a Science and Technology Daily reading of the report); the underlying report data is unverified.
15.
Zhipu's ARR of $1.6B: no new corroboration found in the window; still single source and unverified.
16.
T-Head's confidential A1 filing: single source (China Fund News via Xueqiu); the filing itself is unverified.
17.
The H3-Image license reading (not applicable in the US, EU, UK or South Korea; companies with over $1M in annual revenue need a separate written license): single-source interpretation, unverified.
18.
The Epoch AI dataset (top data centers at 15.1 GW of compute, 38 GW by end-2028): a single dataset, with power figures as estimates, unverified.
19.
Chainstack's AI-driven attack disclosure: single source; the incident itself falls outside the window, the disclosure inside it, unverified.
20.
Nebius' closing price: no reliable quote obtained; described by its intraday decline instead.
21.
Alibaba's Nov 12 morning-session fall of 4.97% to HK$95.6: reported only in a Chinese-language roundup, unverified.
22.
Zhang Yiming's remarks quoted verbatim: disclosed from an internal meeting, with no confirmation from ByteDance.
23.
Alibaba's claim that complex web tasks use half the tokens of DeepSeek 4.1 Flash: an official Alibaba claim appearing only in a Chinese-language roundup, unverified.
24.
Two brief incidents on OpenAI's status page on Nov 12: single-source aggregation, unverified.
25.
Anthropic's $10B-each private placement from Google and NVIDIA: partly confirmed — Fool and AI Agent Update reported the split from the filing on Nov 11, but no Reuters or Bloomberg cross-check; still unverified pending official or mainstream confirmation.
26.
Anthropic IPO pricing on Nov 18 and trading on Nov 19: CNBC, Reuters and WSJ differ (Nov 19 or 20); the FT says pricing is expected Nov 18; the official filing leaves the dates blank; unverified, pending the official announcement.
27.
DeepSeek Harness at 92 trillion tokens per day: company-reported, official-API-only figure with no methodology disclosed and no independent audit.
28.
Tencent Hy4 weights, system card and full-release license: Tencent only promised release "in the coming days" with no date; still not out as of Nov 12; the full-release license terms are unpublished, and the preview's Apache 2.0 license does not imply the full release's terms.
29.
Perplexity's new funding figures conflict: DefiLlama still lists a Nov 2 Series F-1 of $3B at a $30B valuation (updated Nov 6), but several Oct 31 checks say the last confirmed valuation was $20B in September 2025, with anything above $30B only in talks in August; no company announcement or mainstream confirmation — the $3B/$30B figures are not used.