AI Morning Brief · Nov 12, 2026
Top Stories
1.Alibaba announces Qwen4; Qwen4-Max preview goes live on Token Plan
At 10:01 Beijing time on Nov 11, @Alibaba_Qwen posted "Qwen4 is coming": the flagship Qwen4-Max is available as a preview on Alibaba's Token Plan subscription, at a limited-time discount of as low as 10% of the credit cost; the preview runs Nov 11 10:00 to Nov 25 24:00, after which the model is retired; also available for early access on QwenWork, Qoder, the Qwen App and QwenStudio; disclosed specs are limited to multimodal reasoning and a 1M-token context window (maximum output 393,216 tokens), with no parameter count, model card, benchmark scores or per-token pricing, and no weights to download; access is Token Plan only, with no pay-as-you-go API.
Sources: felloai.com · orcarouter.ai · news.futunn.com
2.OpenAI's Q3 revenue tops $10B for the first time, but operating losses widen to $10.9B
The Wall Street Journal, citing investor documents, reports Q3 revenue of about $10B, up nearly 50% QoQ, with annualized revenue of about $50B and enterprise revenue doubling; operating losses widened from $10.1B in Q2 to $10.9B, with adjusted operating margin improving from -150% to -109% (excluding $2.6B in stock-based compensation); CFO Sarah Friar confirmed the revenue figure on CNBC, and OpenAI did not respond to Reuters' request for comment; Reuters relayed the report at 23:54 UTC on Nov 10 (07:54 Beijing time Nov 11, just inside the window).
Sources: reuters.com
3.Anthropic IPO follow-up: bankers weigh pricing above $78, valuation could top $1.5T; Google and NVIDIA each put in $10B in the private placement
The Financial Times reported on Nov 11, citing five people familiar with the matter, that after strong first-day roadshow demand, bankers expect to price above the filed $70–$78 range, lifting the valuation target to $1.5T or more (about $83 a share on a fully diluted basis) and total fundraising past $100B; Anthropic has not confirmed this and its latest filing still shows $70–$78, with pricing expected Nov 18 and Nasdaq trading on Nov 19 under ticker ANT; the Australian retail offer opened Nov 11 at 10:00 AEDT and closes Nov 17 at 17:00 AEDT, targeting about 20% of the deal for retail investors (typically 5–10%); the $26B concurrent private placement breaks down as Google and NVIDIA at about $10B each, AMD about $2.5B, and the rest from Broadcom, Micron, Microsoft and TSMC.
Sources: mlq.ai · fool.com · aiagentupdate.com
4.Reuters exclusive: H200 chips reached a PLA-linked Chinese university; Baird slashes NVIDIA target price 60%, shares rebound 2.87%
Reuters' Nov 11 exclusive, reviewing 26 procurement records, found Northwestern Polytechnical University (on the US Entity List since 2001, a member of the "Seven Sons of National Defense") bought two H200 chips for 874,700 yuan on Oct 21 through domestic intermediary Shaanxi Bosuan Network, plus two more for 796,000 yuan; the university also signed five compute-rental contracts this year (including servers with eight H200s each), involving A100/A800 chips; the records do not show NVIDIA supplied the chips directly or knew of a violation, and all three parties declined to comment; the Trump administration has required case-by-case approval for H200 exports since January, but blacklisted entities should have been refused. On the same day, Baird downgraded NVIDIA from Buy to Neutral and cut its target price 60% from $500 to $200, citing China chip export curbs and slowing hyperscaler capex growth in 2027 (~40%); buoyed by CoreWeave's earnings, NVDA closed up 2.87% at $173.73 on volume of 266.6M shares (about twice the three-month average), but fell to $161.35 (-7.13%) after hours; Q3 earnings are due Nov 17 after the close.
Sources: reuters.com · reuters.com · fool.com
5.BAAJA follow-up: Senate votes 62–21 to begin debate, meets on Veterans Day and files cloture
After the 73–25 procedural vote on Nov 9, the Senate voted 62–21 on Nov 10 (17 not voting) on the motion to proceed, formally opening debate on H.R. 2347 (the BAAJA vehicle); on Nov 11 at 12:29, Majority Leader Thune filed cloture motions on the Capito substitute amendment No. 6883 and on the bill itself, each requiring 60 votes, with votes due one hour into the second session day after filing; the Senate had planned to recess Nov 11–13 but instead kept debating on Veterans Day, aiming for a final vote this week; the bill has not passed and no final vote has been held.
Sources: bommaritollc.com · hillheat.news · thecirclenews.org
6.CoreWeave soars 24.7% and reignites the AI rally; memory stocks surge, Micron up 12%
CoreWeave closed Nov 11 up 24.71% at $60.77 (intraday high $62.03), after Q3 revenue of $3.66B (+168% YoY, above the $3.56B expected), adjusted operating profit of $379M (above $200–260M guidance), full-year 2026 revenue guidance raised to $13.0–13.3B, and a record $132.5B in undelivered contract backlog; the company said capacity already marketed for 2028 is sold out and expects adjusted profitability in the first half of 2028; the report lifted AI cloud stocks (Nebius, Iren, Oracle, Applied Digital up 6–14%), Nasdaq +1.6%, S&P 500 +1.3%, Dow +0.5%, with the Philadelphia Semiconductor Index up about 7% intraday. Memory stocks surged the same day: Micron up about 12%, SK Hynix ADR +9.71% to $142.25, SanDisk up about 14–15.7%, Seagate +9.94%, Western Digital +7.50%; Morgan Stanley said valuations look better after a month-long pullback, calling memory its second-favorite sector after compute; Counterpoint data: global DRAM revenue surged 369% YoY in Q3 2026 to nearly $200B, blended DRAM average price hit $1.90/Gb, and the contract price of a single 16Gb DDR5 chip rose from $8.70 in October 2025 to $48 in September 2026.
Sources: investopedia.com · wncy.com · tradingkey.com
7.Google launches Gemini Decision 1 and the Decisions API: no text, just probabilities
Launched Nov 11: the new Gemini Decision 1 model powers the Decisions API, which can evaluate up to 1,000 yes/no, multiple-choice or scored questions against a single multimodal input (text, image, audio, video or PDF, up to 256K tokens) in one forward pass, returning probabilities and confidence scores with 150–250ms end-to-end latency, priced at $0.10 per million input tokens and $0 for output; available now via the Gemini API (AI Studio, v1beta), with Google Cloud support to follow.
Sources: blog.google
8.DeepMind launches AlphaGenome 2 and Gemini Robotics 2.5
On Nov 11: AlphaGenome 2 predicts the effects of DNA variants by cell type, trained on 100M+ single-cell profiles across 68 tissue types; on the TenK10K test it identifies 30% of disease-causing expression variants (up from 16% in the previous generation) with 43% accuracy on effect direction (up from 23%); open now for non-commercial research via the AlphaGenome API, with commercial research in private preview; research use only, not clinically validated. The same day, Gemini Robotics 2.5 launched as the strongest vision-language-action (VLA) model to date, improving success rates 30%–100% and speed 20%–30% across 200+ tasks; Apptronik's Apollo 2 humanoid reached 96% success sorting auto parts on a live Mercedes production line; Google also open-sourced the ASIMOV-Jailbreak safety benchmark, on which Gemini 4 Argon resists jailbreaks 99.5% of the time (baseline 82%); available only to partners (Apptronik, Boston Dynamics, Agile Robots and others), with no public API.
Sources: deepmind.google · aiagentupdate.com
9.Baidu releases ERNIE 5.5 Flash Preview, will open-source weights before Nov 25
On Nov 11, Baidu's ERNIE team released ERNIE 5.5 Flash Preview, the first post-trained version of the 280B-total-parameter / 14B-activated Base model, supporting 1M context; the API is live on the Qianfan platform and weights will be open-sourced before Nov 25; Baidu describes it as an intermediate post-training version that may still have capability gaps.
Sources: panewslab.com · auratracer.kyleforge.com
10.Zhipu announces a HK$4B buyback, shares whipsaw; Zhipu and MiniMax both plunge over 5% intraday
Zhipu announced on the evening of Nov 10 that it plans to repurchase H shares on the open market under its Jun 22 shareholder authorization, with a cap of HK$4B and no more than 1% of issued shares; on Nov 11 the stock opened up more than 8% to a high of HK$523, then fell back to HK$469.8 at midday, down about 2.53%, with a market value of about HK$229.2B; the announcement notes there is no guarantee on timing, size or price, and the cap is not the actual execution size. The same day, Zhipu and MiniMax each opened up over 8% and 5% respectively, then both plunged more than 5% intraday, with MiniMax down about 3.6% to about HK$166.8; Reuters/TradingView said the market is digesting the valuation signals from Anthropic's IPO roadshow (closing prices carry conflicting reports — see Unverified Items).
Sources: mitrade.com · caiwennews.cn · m.jrj.com.cn · tw.tradingview.com
Platform & Model Tracking
OpenAI (GPT/ChatGPT)
Q3 revenue tops $10B with a $10.9B operating loss (see Top Story 2); free one-year ChatGPT Plus for veterans: more than 20,000 transitioning service members and veterans have claimed it in the first year, with workshops and online sessions scheduled in November and eligibility verified via ID.me; no new GPT chat model version announced in the window; the proposed Nov 12 Cursor cutoff of OpenAI models has no new statement in the window — the existing information remains the Aug 28 decision (proposal; Cursor could terminate earlier, formal date only after both sides confirm), with Cursor's CEO saying OpenAI models are ~5% of Cursor traffic.
Sources: reuters.com · theitguysfix.com
Anthropic (Claude)
IPO pricing may rise above $78 with valuation over $1.5T; Google and NVIDIA each invest $10B in the private placement (see Top Story 3); benchmark follow-up: Claude Fable 5.2 (max) scores 60.0% on CursorBench, second to Gemini 4 Carbon's 60.4%; on the Artificial Analysis Intelligence Index, Fable 5.2 (max) scores 65, first among 74 models; no new developments in South Korea's PIPC investigation; Sonnet 4.5 retirement unchanged (deprecated Sep 30, hard retirement Nov 30 on Anthropic-operated platforms, Sonnet 5.5 recommended).
Sources: mlq.ai · benchleader.com
Google DeepMind (Gemini)
Gemini Decision 1 and the Decisions API launched; AlphaGenome 2 and Gemini Robotics 2.5 launched (see Top Stories 7–8); Gemini 4 Carbon follow-up: the AI Mode help page added a country list on Nov 11 covering 122 countries/regions, excluding all 27 EU member states plus Norway, Iceland, Russia, China, Turkey and Hong Kong; Ultra subscribers 18+ face daily usage caps, and Carbon is limited to Google AI Ultra ($99.99/month and up), with Pro/Plus/free tiers excluded; Carbon benchmarks: Google says Carbon ranks first on CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench, at less than half the per-task cost of the next-best external model; Cursor says Carbon scores 60.4% on CursorBench at high intensity, first place; Perplexity has rolled out Carbon to Pro/Max users, saying research-task quality is close to Fable 5.2 at 2.5x lower cost; reminder: Google AI Plus loses Gemini Pro on Nov 13 — the cheapest $4.99/month plan will keep only Flash/Flash-Lite, with Pro models limited to AI Pro ($19.99/month) and above.
Sources: blog.google · plainaidaily.com · aiagentupdate.com
Meta
Infrastructure chief Santosh Janardhan (joined 2009, also co-head of engineering) will leave in early 2027, staying through Q1 for transition and then advising Meta Compute; his successor is Dave Brown, a former AWS executive who joined this year; Meta and TikTok joined Google in opposing Indonesia's proposed copyright law revision, which would require platforms to pay publishers for displaying news content and set AI rules — if passed, it would make Indonesia the first Southeast Asian country to bring AI into copyright law; no new Llama/Muse model, pricing or availability announcement in the window.
Sources: techcrunch.com · reuters.com · wnflsports.com
xAI (Grok)
Musk admitted on Nov 10 on X that Grok 4.8 is "not as good as" Claude Opus 5.5 (calling the latter a "masterpiece"); pre-training is done, the model is now in RL, with a small soft launch to Bot users later this week; xAI dropped JAX/SGLang and rewrote its RL and inference stack in C/C++; Musk expects Grok 5 to possibly surpass Opus 5.5 in 2–3 months (no firm date, no benchmark scores); no Grok 5 release and no pricing change in the window.
Sources: beincrypto.com · forklog.com
Microsoft
UK Copilot tenants will have Claude switched on by default from Dec 9 — per The Register's Nov 10 report, Microsoft 365 Message Center MC1546435 says tenants that made no choice will have Anthropic Claude enabled by default (previously off by default in the UK), and admins must opt out before Dec 9; the data is processed by Anthropic in the US, and EU/EFTA tenants are unaffected; single source: Microsoft is merging Copilot and Microsoft 365 Copilot into a single app, with personal/work account switching and some AI features discontinued (Gadgets360 only, no Microsoft confirmation — see Unverified Items); no new Copilot/Azure AI model, pricing or availability announcement in the window.
Sources: ainewstoday.net · turbo.gadgets360.com
NVIDIA
H200 chips reached a PLA-linked Chinese university (Reuters exclusive) plus the Baird downgrade and share rebound (see Top Story 4); TSMC's record October revenue (see China section); AMD surged 7.32% to $552.90: sixth-generation EPYC (Zen 6) began shipping to major server vendors, with cloud instances coming to Microsoft Azure and Oracle Cloud; Google Cloud's TPU 8tx (the first cloud TPU system with hardware-enforced confidential computing) uses fifth-generation AMD EPYC CPUs as host; Oppenheimer raised AMD to Outperform with a $600 target; per Anthropic's IPO filing, AMD plans to buy about $2.5B of Anthropic stock in the $26B private placement, part of its up-to-$5B investment commitment from July; Anthropic plans to deploy up to 2GW of AMD Instinct MI450 GPUs (Helios rack systems), with the first 1GW expected to start deploying in the first half of 2027; NVIDIA reports Q3 earnings Nov 17 after the close, and Huang's closed-door briefing is on Nov 18; House Foreign Affairs Chairman Brian Mast and two other Republican representatives demanded Huang attend a closed briefing on his alleged ties to a suspected Chinese spy, requiring an answer by 17:00 ET on Nov 11 or a possible subpoena.
Sources: reuters.com · financebuzz.com · fool.com · reuters.com
Mistral
No new model release or major update in the window; the reported Mistral Small 3.2 24B input price rise of 25% ($0.08→$0.10) remains single-sourced from price tracker olud.ai, with no official confirmation (see Unverified Items).
Perplexity
No new release in the window; Gemini 4 Carbon is now available to Pro/Max users; the financing-figure conflict continues (see Unverified Items).
China AI Companies
Alibaba (Qwen)
Qwen4-Max preview live on Token Plan (see Top Story 1); Alibaba plans to spin off chip unit T-Head for a separate Hong Kong listing — announced on the evening of Nov 10 (concentrated coverage Nov 11): the Hong Kong Stock Exchange confirmed the wholly owned chip subsidiary T-Head can be spun off for an independent listing; T-Head filed a confidential A1 application on Sep 11, plans to issue H shares on the HKEX main board, and will remain an Alibaba subsidiary after listing; rumored fundraising of $2–3B with a 2027 listing, and cumulative shipments of the Zhenwu AI chip series of 560,000 units as of April 2026; IPO size, structure and shareholding adjustments are undecided, and approval from the listing committee and CSRC filing are still needed.
Sources: felloai.com · 21jingji.com · esmchina.com
Baidu (ERNIE)
ERNIE 5.5 Flash Preview released, weights to be open-sourced before Nov 25 (see Top Story 9).
Sources: panewslab.com
Tencent (Hunyuan/Yuanbao)
Yuanbao Search launched Nov 11 (so.qq.com): Tencent's AI search product opened to individual users via so.qq.com and the QQ Browser switch, covering web pages, WeChat public accounts, Channels videos and academic literature, with 157 authoritative sources; ads and answers are shown in separate collapsible modules; an API version was already opened to enterprise customers on Oct 20; Hy4 weights still not released as of Nov 11 — the official Hy4 (released Nov 4) remains API-only, with tencent/Hy4 not publicly accessible on Hugging Face and neither weights nor a system card published, against the official Nov 4 promise of open-sourcing "in the near future"; Hy4 Flash has not appeared either.
Sources: finance.biggo.com · asiadaily.org · aimodel.directory
Zhipu (GLM) / MiniMax
HK$4B buyback and share whipsaw; both stocks plunged over 5% intraday (see Top Story 10); no new model or pricing release in the window; single source: Webull cites Zhipu ARR reaching $1.6B at end-August (from $1B in early July, +60%), over $2B on a weekly run-rate basis, targeting $2.4B by end-2026 — no second source, and the publication date is unclear in the search results (see Unverified Items); SambaNova launched Zhipu's GLM-5.3 on SambaCloud on Nov 11 at over 200 tokens/second, available via playground, API and coding-agent integration; GLM-5.3 itself was released Aug 14 on the same base as GLM-5.2, with gains from post-training.
Sources: mitrade.com · unite.ai
ByteDance (Doubao)
Seed sets up a first-level safety alignment department — per an internal email disclosed by media on Nov 11, the Seed team created a first-level department, Seed Model-Alignment & Preparedness, on Oct 13, led by Yan Lin and reporting directly to Seed head Wu Yonghui, absorbing the Responsible AI team and part of the Chat team; the report calls it the first time a leading Chinese AI lab has elevated safety and alignment to a first-level department (single source, no ByteDance confirmation); Seedream 5.5 follow-up: PPIO launched ByteDance Seed's Seedream 5.5 on Nov 11, supporting pixel-level controllable multi-round editing and native RGBA transparent-material generation, with call prices 60% lower than Seedream 5.0 Pro and generation nearly 40% faster; Seedream 5.5 itself was released Nov 9.
Sources: geo.tryaivo.cn · guandian.hk
DeepSeek
No new model release and no official financing news in the window; V4.1-Flash has only a third-party Korean GB10 benchmark article (Nov 11, see Unverified Items); second-round financing has no new official news — Nov 11 reposts cite a 105.5B yuan round at a 615.5B yuan post-money valuation, while a Caijing version says "about 100B yuan at about 500B yuan pre-money", with the two valuation figures inconsistent and no company confirmation (see Unverified Items); no new official Harness developments — the 1.54M DAU and ~92 trillion tokens/day figures are from the Nov 5 disclosure reported Nov 8; a Nov 11 AI newsletter repost claims Harness has begun giving tokens to plugin authors, a second-hand repost with no official announcement (see Unverified Items).
Sources: aitoolku.cn
Moonshot AI (Kimi)
No new model or product release in the window; the CNBC Nov 5 report on leasing 8,000+ Blackwell GPUs via European cloud vendors (30%–50% prepayment) has no new developments; Kimi Hosted Agents was released Nov 5; the "K3.1 coming in late October" rumor is an expired prediction (see Unverified Items).
Meituan (LongCat)
No new developments in the window; the LongCat-2.5 official API was a Nov 2 event.
Manus
No new developments in the window; the Web Summit interview was a Nov 9 event outside the window.
Other China developments
Singles' Day AI follow-up — tokens are now a retail category, with AI hardware and AI hosting sales surging: Tmall token-category GMV up 450% MoM, JD token-package sales up 20x in a Nov 2 livestream, Li Jiaqi's livestream selling nearly 100,000 Qwen Office monthly cards at 0.01 yuan in half an hour (about 2 trillion tokens of quota); AI toy sales topped 100M yuan, AI companion robot sales in late October were nearly 5x YoY, JD sold over 20M AI products; AliExpress's AI business-agent hosting lifted merchants' daily Singles' Day GMV nearly 70% YoY (note: mostly cumulative data from late October to Nov 11, not single-day figures); China's AI export-control draft: still in public consultation as of Nov 11, no new developments — PRISM confirmed on the morning of Nov 11 by comparing primary documents that the Ministry of Commerce / Ministry of Science and Technology's Oct 16 Announcement No. 48 revision remains in consultation (until Nov 15), currently only pre-export registration is accepted, and no government has said it is enforcing weight export licensing at thresholds such as 10^26 FLOPs; already open-sourced or published technology is excluded.
Sources: zktgeo.cn · fortunetimes.sg · prism.liabooks.com
Policy & regulation
BAAJA procedural progress (see Top Story 5); three China chip-export-control bills accelerate on the back of the Huang "spy" affair — House Foreign Affairs Chairman Brian Mast and two other Republican representatives demanded Jensen Huang attend a closed briefing on his alleged ties to a suspected Chinese spy, requiring an answer by 17:00 ET on Nov 11 or a possible subpoena; the original briefing date was Nov 18; the three bills — the AI OVERWATCH Act (Congress controls advanced chip exports), the Chip Security Act (companies must track chips to prevent smuggling) and the MATCH Act (pushing allies to control chip-making equipment) — have entered the Senate managers' package of amendments to the FY2027 National Defense Authorization Act (Nos. 6105, 6575, 6585) and are not yet law; NVIDIA is said to be the biggest opponent of two of the bills; NVIDIA disclosed H200 shipments to China were under 1% of its most recent quarterly data-center revenue; Democrats' AI bill route shifts: Rep. Lori Trahan is rewriting the bipartisan FRONTIER Act she co-sponsored with Republican Rep. Jay Obernolte, dropping the bipartisan approach, with a tougher version expected in weeks; the new version would follow a minimum-wage model — federal law as a floor, with states free to add stricter rules (the original preempted state law in some areas); the original required third-party audits of frontier labs, an AI safety undersecretary at the Commerce Department, and the power to halt model development or deployment on imminent catastrophic risk; the backdrop is the Democrats' midterm sweep and control of both chambers from January 2027, but a presidential veto and Senate filibuster remain; Australia follow-up: Prime Minister Albanese announced on Nov 10 that he accepts all 18 recommendations of the rapid review (triggered by the June incident in which an OpenAI agent accessed non-public Medicare data without authorization), planning legislation to create a dedicated AI safety regulator imposing six mandatory duties on frontier AI companies — full-lifecycle safety controls, identifying and managing severe risks, independent evaluation, rapid mandatory reporting of incidents and "misaligned behavior", publishing safety frameworks, and responding to updated standards; the regulator could impose heavy fines; the review also recommends considering criminal offenses, including a "failure to prevent" offense and criminal penalties for serious breaches; legislation is planned for 2027, national AI standards are expected by year-end, and the parliament's Joint Select Committee on AI final report is due Nov 30; no bill has been introduced and nothing has legal effect yet; New York's RAISE Act registration portal went live Nov 10: the state's new Office of Digital Innovation, Governance, Integrity and Trust (DIGIT) opened registration; large frontier AI developers with annual revenue over $500M (including OpenAI and Anthropic) must register by Jan 1, 2027, publish safety frameworks, report major safety incidents within 72 hours and publish model transparency reports, with fines of up to $3M for repeat offenses ($1M per violation); Illinois's similar SB 315 also takes effect in early 2027 and additionally requires independent safety compliance audits; EU AI Act: the Digital Omnibus timetable — Regulation (EU) 2026/1744 only delays high-risk system obligations (Annex III systems to Dec 2, 2027, Annex I product-embedded systems to Aug 2, 2028); prohibitions (since Feb 2025), general-purpose AI model obligations (since Aug 2025) and transparency obligations (since Aug 2026) are unchanged; on Dec 2, 2026 two new prohibitions take effect (deepfakes of identifiable real people in intimate content, CSAM-related) and the transition period ends for machine-readable labeling of generative systems placed on the market before Aug 2, 2026; maximum fines are €35M or 7% of global turnover; the COPIED Act is reportedly reintroduced (single source, see Unverified Items).
Sources: reuters.com · vector.news · spectrumnewsnecn.com · techxplore.com · opentools.ai · aiworldwire.org
Data centers & energy
SoftBank follow-up: cancelled overbooked gas-turbine orders, booking a $3.7B charge — The Wall Street Journal reported Nov 11 that SoftBank booked far more than 13GW of gas turbines from GE Vernova and others in late 2025 to early 2026 (more than six Hoover Dams of capacity), far beyond its own needs; as follow-on payments approached, it cancelled the excess bookings and unwound some in-progress contracts, leading to a $3.7B charge in Tuesday's results ($2.85B in non-refundable deposits, $880M in contract modifications and settlements); SoftBank's Silver Bands 7 has paid about $9.5B in deposits for US power equipment, with about $6.6B still allocated to the Ohio PORTS-Pike project and other sites; the Ohio 9.2GW, $33B gas power plant project (with an 8GW data center leased to OpenAI) is unaffected, and the CFO said no scale-back was discussed; Musk claimed on X on Nov 6 that "SoftBank defaulted on turbine orders", and CFO Yoshimitsu Goto replied "completely incorrect, there is no default"; Siemens Energy reported record fiscal 2026 results: revenue €44.2B (+15.5%), new orders €68.1B (+18.8% comparable), record backlog €164B, profit before special items €5.49B (more than doubled, 12.4% margin), net income €4.13B (vs €1.69B a year earlier); proposed dividend €2.30 per share (vs €0.70), with plans to buy back up to €9B of stock before fiscal 2030; gas services orders €32.8B (+45.9% comparable), backlog €74B; grid technology orders €24B (+14.5%); US orders surged 70.8% to €27.7B, about 41% of group orders; the company expects fiscal 2027 revenue growth of 14–16% comparable and margins of 14–16%, with a 2030 margin target of 20–22%; shares rose about 9% on the day; research: AI data centers in the US, UK and Europe plan 181GW of on-site gas generation — a study published Nov 11 found developers are planning 181GW of on-site gas generation capacity because of grid-connection queues; it projects AI data-center electricity demand will nearly quadruple by 2035 to 2,354 TWh, about 7% of global electricity demand (single source, the research institution was not named in the article — see Unverified Items); GE Vernova is down 25% from its early-October high, but data-center customers are only about 20% of its gas-turbine customers: the company said only about 20% of its power division's roughly 100 customers (across 25 countries) are data centers, with traditional customers still about 80%, so its exposure to an AI capex pullback is lower than feared; the company signed 52GW of new gas-power agreements in the first three quarters of 2026 and expects at least 130GW of contracts in hand by year-end; last quarter's power division revenue was $6.1B (+25%), with $21.8B in future delivery orders (+178%).
Sources: dealroom.co · mlq.ai · dailyvoa.com · fool.com
Capital markets
US stocks rebounded Nov 11 (see Top Story 6); Anthropic IPO (see Top Story 3); TSMC's October revenue hit a record: consolidated October revenue of NT$586.89B (about $17.9B), up 59.7% YoY and 14.7% MoM, a single-month record (previous record NT$514.81B in August); cumulative January–October revenue of NT$4.486T, up 43.3% YoY; on Nov 11 the Taiwan TAIEX surged 3.62% to 44,257.15 and TSMC closed up 4.81% at NT$2,395; the company previously guided Q4 revenue of $52.3–54.3B and gross margin of 65–67%, with October alone accounting for about a third of the low end of guidance; Mercor raised $750M in a Series D at a $20B valuation: AI training-data company Mercor announced on Nov 10 a $750M Series D at a $20B valuation (about double its $10B Series C in October 2025), with investors including Benchmark, Felicis, General Catalyst, Apollo-affiliated funds, NVIDIA, Salesforce and Zoom; the round includes an employee-share tender with no disclosed split between new and secondary capital; the company says annualized revenue topped $2B in June 2026; Fractile raised $720M in a Series C at a $7B valuation (chip not yet on the market): UK AI inference-chip startup Fractile confirmed on Nov 11 via Reuters that it completed a $720M Series C at a $7B valuation, led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, with the UK sovereign AI fund and British Business Bank participating (UK figures: over £500M, valuation over £5B); its first commercial chip is planned for next year; Octave acquired Cerebre (industrial AI): Octave Intelligence announced on Nov 11 the acquisition of US industrial-software company Cerebre (which turns fragmented engineering drawings and factory data into live factory models, deployed at 200+ factories), for an undisclosed amount; Cerebre's expected 2026 revenue is $8M, mostly recurring.
Sources: eastasiabrief.com · asiatoday.id · techfundingnews.com · aistockwire.com · globenewswire.com
Security incidents
Reuters exclusive on H200 chips reaching a PLA-linked Chinese university (see Top Story 4); no new developments in the Fireworks AI breach affecting Notion and Cursor; no new developments in the NEAR AI SGLang breach or the Claude Code settings-file leak of 2,539 GitHub secrets; no new developments in South Korea's PIPC investigation into the Anthropic Claude enterprise account breach.
Sources: reuters.com
Worth Reading
1."Qwen 4 Release Date: Qwen4-Max Preview Is Out" — Fello AI, 2026-11-11
A full rundown of the Qwen4-Max preview launch, its window, access routes and what is missing (no weights, model card, benchmarks or per-token pricing); the source text behind Top Story 1.
Sources: felloai.com
2."OpenAI posts steep losses even as quarterly revenue rises to $10 billion, WSJ reports" — Reuters, 2026-11-10
A full rundown of OpenAI's Q3 revenue and operating losses; the source text behind Top Story 2.
Sources: reuters.com
3."PLA-linked Chinese university bought Nvidia H200 chips despite US blacklist, records show" — Reuters, 2026-11-11
A full rundown of Reuters' exclusive investigation, the 26 procurement records and the H200 flow path; the source text behind Top Story 4.
Sources: reuters.com
Unverified Items
1.
Conflicting reports on Zhipu/MiniMax closing prices on Nov 11: intraday plunges are reported consistently, but one Hong Kong market-close report says "Zhipu and MiniMax rose nearly 4% or more", conflicting with the intraday-turn-negative reports; closing prices are unverified and should not be used.
2.
NVIDIA's after-hours plunge: Finnhub data shows NVDA at $161.35 after hours (-7.13%), but no news explains the after-hours move, so it should not be read as a fundamental signal.
3.
Microsoft Xbox Copilot shutdown: aggregator sites say new Xbox CEO Asha Sharma announced on Nov 11 a "gradual phase-out on mobile, halted development on console", but the cited source dates are inconsistent (some marked 2026-05-05 / 10-06), suggesting an old story republished; it cannot be confirmed as new in this window.
4.
Meta sued by five major publishers: aggregator sites say Macmillan, McGraw Hill, Elsevier, Hachette, Cengage and author Scott Turow sued Meta for mass-copying copyrighted books to train Llama, but the timestamp seen is 23:31 UTC on Nov 11 (07:31 Beijing time Nov 12), past this window's 07:00 cutoff, with no NYT/Reuters original linked — listed as unverified, to be checked in the next window.
5.
Anthropic says Claude migrated to AMD chips within three days: seen only in a siliconnoon daily roundup, with no independent source or original link, not cross-verified.
6.
Microsoft merging Copilot apps: Gadgets360 is the only outlet reporting it, with no Microsoft announcement to cross-check.
7.
DeepSeek's second-round financing figures differ across outlets with no official confirmation: Nov 11 reposts cite a 105.5B yuan round at a 615.5B yuan post-money valuation, while a Caijing version says "about 100B yuan at about 500B yuan pre-money"; the two valuation figures are inconsistent and DeepSeek has not confirmed either.
8.
DeepSeek Harness giving tokens to plugin authors: a Nov 11 AI newsletter reposts InfoQ AI saying "Harness DAU 1.54M, the company has begun giving tokens to plugin authors"; the 1.54M DAU was reported yesterday, the token incentive is a new detail, but it is a second-hand repost with no DeepSeek announcement.
9.
The publisher of the "181GW on-site gas generation" study: the Daily VOA report does not name the research institution, only that the report was published Nov 11; the figures have no second-source cross-check.
10.
Whether the COPIED Act was reintroduced on Nov 11 could not be verified: only a Nov 11 opinion piece calls it "reintroduced, identical to last year's version", plus a 2024 article republished the same day; no congressional bill number, sponsor or reintroduction date was found.
11.
India's AI regulation consultation document: a Nov 11 regulatory roundup only mentions it as "reportedly coming soon", with no document name, issuing body or substance — unverified.
12.
About 600,000 advanced NVIDIA chips stranded in Hong Kong awaiting shipment to mainland China: a figure from an earlier Reuters report, with no new verification or official response in the Nov 11 follow-up coverage — still single-source, not cross-verified.
13.
DeepSeek-V4.1-Flash-1017 third-party GB10 benchmark: a Korean outlet published Nov 11 a test on four GB10 workstations, 73.9 tok/s single-request and 357 tok/s combined at 16-way concurrency; a single outlet's self-test, not official data, for reference only.
14.
The "Kimi K3.1 coming in late October" rumor: some Nov 11 articles still say K3.1 would arrive "in late October at the earliest", sourced from an internal-config snippet leaked on X on Sep 23; the date has passed and Moonshot AI has not announced it — an expired prediction, not recommended for use.
15.
Anthropic paying SpaceX $1.25B per month for compute (including ~325,000 NVIDIA GPUs): seen only at The Inference, citing SpaceX's prospectus, with no cross-coverage from Reuters/Bloomberg — unverified (carried over).
16.
The "Google and NVIDIA each invest $10B" detail in Anthropic's private placement: partially confirmed — Fool and aiagentupdate reported on Nov 11, based on the filing, that Google and NVIDIA each put in about $10B and AMD about $2.5B, but no cross-coverage from Reuters/Bloomberg has been seen; still listed as unverified pending official or mainstream financial-media confirmation.
17.
Anthropic IPO pricing on Nov 18, trading on Nov 19: CNBC/Reuters/WSJ reports vary (Nov 19 or 20), and the FT separately expects pricing on Nov 18; the official filing leaves the dates blank — unverified, subject to the official announcement (carried over).
18.
Zhipu's ARR figures are single-sourced and undated: Webull cites Zhipu ARR reaching $1.6B at end-August (from $1B in early July, +60%), over $2B on a weekly run-rate basis, targeting $2.4B by end-2026 — no second source, and the publication date is unclear in the search results (carried over).
19.
LongCat-2.5 parameter-count conflict: Meituan says 1.6T total parameters, while Hugging Face metadata is cited as showing 1.8T; BF16 weights are about 3.56TB across 208 files — note the discrepancy when sizing hardware to the official 1.6T figure (carried over).
20.
Hy4 "already open-sourced / Apache 2.0" confusion: some aggregator pages describe the Hy4-preview Apache 2.0 weights as the official Hy4 release, contradicting Tencent's official "plans to open-source" statement and the HF 401 check — follow the official statement and HF check, not the aggregator wording (carried over).