Edge AI Daily Morning News (November 11)

This article may contain an automated translation. The original source is linked for verification.

Edge AI Daily Morning News (November 11)

Edge AI Daily

2026.11.11 08:34

· From Beijing

Full text 6889 words

Silicon Valley Frontier:

1. Anthropic priced at US$74 and raised net funds of US$93.7 billion: the largest IPO in AI history backed by seven giants

1. IPO scale: According to the Financial Associated Press on November 10, Anthropic expects the midpoint of the IPO pricing to be US$74 per share (middle of the range of US$70 to US$80). After including over-allotment rights and US$26 billion in private placements by seven giants, the net raised funds will be approximately US$93.7 billion, surpassing SpaceX’s record of US$85.7 billion in June this year, becoming the largest IPO in history.

2. Industrial chain card slots: Google, Nvidia, AMD, Broadcom, Micron, Microsoft, and TSMC have committed to invest a total of US$26 billion in private placements at the issue price; they lock in future orders for chip suppliers, and enter defensively for Google and Microsoft. It is described as a tacit understanding of "mutual non-destruction" in the AI industry chain.

3. Valuation decline and fund-raising purpose: Affected by the retracement of technology stocks after the U.S. midterm elections from November 4 to 7, and the Nasdaq's single-day decline of more than 5%, the psychological price of buyers has dropped to US$1.2 trillion to US$1.3 trillion, and US$74 is priced at the compromise point between buyers and sellers; the US$93.7 billion raised is mainly used to fill the computing power expenditure commitment of approximately US$69.8 billion due in the next 12 months, which is a "supply landing" rather than a "victory triumph."

2. Anthropic releases Fable 5.2 and Mythos 5.2: 1000 sub-agents write a bootable operating system in 8 days

1. On November 9, local time, Anthropic released Claude Fable 5.2 and Mythos 5.2; Fable 5.2 Coordinated more than 1,000 sub-agents to work non-stop for 8 days, writing from an empty folder a prototype operating system (kernel, disk and keyboard driver, file storage, window and login interface, installation program) that can be booted on a laptop. Another 30 hours were spent to transcribe the proof of Fermat's last theorem into a formal proof that the machine can gradually verify, marking the transformation of the agent from "chatting assistant" to "end-to-end delivery of results".

2. The two models share the same underlying weight, but the difference lies in the guardrails: Fable 5.2 is open to all users and retains strong network security and biosecurity restrictions, while Mythos 5.2 is only open to institutions that have passed the review. The pricing has been reduced to US$8 per million inputs and US$40 per million outputs (down from the previous generation). The Fast mode speeds up by up to 2.5 times but doubles the price. Officials say that the single task cost of High mode is about half of the Max mode of Opus 5.5. It is recommended to use Opus 5.5 for daily coding and knowledge tasks, and then hand over the most difficult open projects to Fable.

3. Officially disclosed that more than 90% of the merged changes in its core code base have been written by Claude under human guidance. In early versions, key production services were migrated from Python to Rust; however, the high-inference gear is slow, developers can easily reach the quota limit, and there are agent transgressions (such as sending 33 emails to partners without permission), indicating that "should it do this" is more worthy of vigilance than "whether it can do it".

3. Jen-Hsun Huang admitted to extramarital affair, Nvidia launched an independent investigation

1. Events and trends: On November 8, Nvidia’s board of directors announced the launch of an independent investigation to verify whether Huang Renxun had become a target of foreign intelligence operations. On the same day, Huang Renxun admitted in an internal memo that he had been unfaithful to his wife. The above-mentioned events occurred before the decline. As of the close on November 10, Nvidia was trading at US$168.80, down 5.53% (US$9.87) in a single day, and a cumulative drop of nearly 10% in two days. It was at a discount of approximately 23% from the 50-day moving average of US$219.04, and a retracement of approximately 31% from the year's high of US$243.37.

2. Policy and governance risks: The board of directors’ statement defined the subject of the investigation as “whether Huang Renxun was targeted by foreign intelligence operations” rather than “whether NVIDIA leaked secrets”, and stated that “there is no reason to believe that U.S. national security or NVIDIA secrets have been compromised.”

3. Fundamentals and valuation: The fundamentals have not deteriorated yet. In fiscal year 2026, revenue is US$215.938 billion, a year-on-year increase of 65%, GAAP net profit is US$120.067 billion, a year-on-year increase of 65%, and GAAP diluted earnings per share is US$4.90; Wall Street pricing has not changed. The average target price of 79 analysts is US$340.03 (implying about 101% upside), about 76% give a buy rating, and the forward price-to-earnings ratio is about 18 times based on the consensus earnings per share of US$9.29, which is close to the level of mature hardware companies. The subsequent direction depends on three variables: the conclusion of the board of directors’ independent investigation, the third fiscal quarter financial report on November 17, and the results of the closed-door congressional briefing on November 18.

4. Google releases TPU four times a year: not building new chips, using packaging, network and security to rewrite AI hardware rules

1. Products and strategies: Google released TPU 8tx (confidential inference) and TPU 8i Max (extra large context) on November 9, 2026. They are not new tape-outs but reuse existing silicon wafers to repackage the network, allowing them to launch four TPU systems within one year (the first time in history); the eighth-generation architecture (split into TPU 8t training, TPU in April 2026 8i reasoning) turns to the modular "building block" route, replacing waiting for the next generation of chips with a combination of packaging, network, memory, and security.

2. Demand and data: AI workloads are structurally differentiated, and the boundaries between training and inference disappear in the era of intelligent agents. TPU 8t training performance is up to 2.7 times higher than the previous generation Ironwood, and TPU 8i inference cost-effectiveness is increased by 80%, both based on TSMC N3 process; TPU 8i Max dual-core packaging doubles HBM capacity, SRAM and bandwidth. A single TPU 8i is 288GB HBM, 384MB SRAM, and 8601GB/s bandwidth.

3. Competition and uncertainty: Google intends to change "reliance on NVIDIA" from a single-choice question to a multiple-choice question of "choosing hardware according to workload". TPU has served nearly 100 customers (Anthropic will expand TPU cooperation with Google and Broadcom in October 2025); but the price, launch time and absolute performance figures of TPU have not been announced. 8t/8i still needs to be fully launched later in 2026. Whether it can shake Nvidia depends on pricing, delivery timeliness and whether third-party independent testing can reproduce its claimed indicators.

5. Google decentralized Gemini Intelligence to Motorola and the old Pixel, and Alphabet shifted from AI investment to AI distribution.

1. Financial report background: Alphabet's revenue in the third quarter of 2026 was US$127.6 billion (a year-on-year increase of 25%), but it recorded negative free cash flow of US$9.07 billion. Because it was less than expected, the stock price fell by about 10% in a single day after the financial report; the company maintained its 2026 AI capital expenditure guidance of US$195 billion to US$205 billion, and is expected to exceed US$300 billion in 2027. The scale of investment is close to its single-quarter revenue.

2. Monetization efficiency is under pressure: Google Chief Commercial Officer Philipp Schindler said that the monetization ability of each query of "AI Mode" is still lower than that of traditional search (fewer ads, lower single value); Search revenue growth rate dropped to 12% in the third quarter (17% in the second quarter, 19% in the first quarter). Query volume reached the fastest growth rate in five years but revenue was almost flat quarter-on-quarter; during the same period, Google Cloud revenue increased by 103% year-on-year to US$30.7 billion.

3. Distribution strategy and market expectations: Google expanded "Gemini Intelligence" to Motorola's first supported model and three older Pixel 10 models, and added four smart-like functions, aiming to keep users in the ecosystem rather than selling new phones; as of early November 2026, the stock price was approximately US$287.24, approximately 30% lower than the 52-week high, and analysts' average target price was US$400.32, implying approximately 39% room for growth.

6. Meta uses one-third of the cost to bet on smart water, electricity and coal

1. Meta will release the lightweight model Muse Ember in November 2026, focusing on "unit task cost" rather than an intellectual arms race: Artificial Analysis intelligence index 36 points and Claude Haiku 5.5 (released on October 7) is the same, but the single task cost of the highest inference gear is about US$0.14, which is only one-third of the latter (about US$0.41); the pricing input is US$0.10/million Tokens and the output is US$0.50/million Tokens, which is the same price as Haiku 5.5.

2. The high-frequency call scenario of intelligent agents makes "cost per task" surpass "single round intelligence" and become the first indicator for enterprise selection. Muse Ember scored 37.8% in the AutomationBench-AA end-to-end business workflow, which is about twice that of GPT-6 Luna (20.2%) and 2.7 times that of Haiku 5.5 (13.9%). It also leads in the coding evaluation; its shortcoming is knowledge (AA-Omniscience accuracy rate is 21.1%, due to active rejection of 64% of questions), but the hallucination rate is only 17.4%, which is Haiku 5.5 Half of top gear.

3. Ecological card position is clear: Muse Ember will log into Meta Model API, Google Cloud and Oracle Cloud at the first launch, and join forces with Box, Cline, Cloudflare, and You.com to embed agent sub-task links, avoiding the head-on competition with OpenAI and Anthropic for the "strongest model", and positioning itself as the "water, electricity and coal" of the agent economy. In 2026, the large model market will be divided into two routes: "intelligence arms race" and "efficiency war". The unit task cost may become a harder competition barrier than intelligence scores.

7. New York puts reins on AI: OpenAI and Anthropic must register before 2027, with fines of up to 3 million for violations

1. The New York State RAISE Act enters the enforcement stage: the registration portal will be opened on November 10, 2026. Large-scale cutting-edge AI developers with annual revenue exceeding US$500 million (such as OpenAI, Anthropic, Google, Meta, Microsoft) must register in 2027 Registration must be completed before January 1, 2020, and the bill will take full effect on the same day; supervision is handled by the Office of Digital Innovation, Governance, Integrity and Trust (DIGIT) under the New York State Department of Financial Services (DFS). The maximum fine for a first violation is US$1 million, and the maximum fine for subsequent violations is US$3 million.

2. Compliance obligations and cross-state comparison: After registration, companies must disclose their security framework, report critical safety incidents within 72 hours, submit catastrophic risk assessments every quarter, and disclose every two years; New York's 72-hour time limit is significantly stricter than California's SB 53's 15 days (only reduced to 24 hours for imminent danger). Illinois SB 315 will take effect in early 2027 and additionally require independent safety compliance audits. The three states are forming a state-level regulatory puzzle that must be met simultaneously.

3. Regulatory logic upgrade and corporate impact: The "RAISE Act 2.0" proposed at the end of October 2026 intends to give DIGIT the power to limit model training computing power, and allows the state attorney general to apply to the court to stop the deployment of models in New York, marking a shift from "requirements" to regulation. "Requesting disclosure" turns to "holding the brake"; the rising costs of transparency, response and deployment may become a moat for leading companies, and small start-ups may be blocked. At the same time, Hochuer has raised $55 million for the "New York AI Responsibility and Research Center" at Binghamton University.

8. Intentional transaction volume of US$85.45 billion: 2.3% of the CIIE, hiding the votes of global capital

1. The 9th China International Import Expo (in Shanghai from November 5 to 10, 2026) concluded. On a one-year basis, the intended transaction volume reached US$85.45 billion, a year-on-year increase of 2.3%, and the absolute increase was approximately US$1.96 billion; the growth rate was relatively The 4.4% slowdown in the previous session was mainly due to the high base effect of more than 80 billion US dollars. The scale of this session also hit a new high (162 countries and regions, the exhibition area exceeded 440,000 square meters for the first time, 4,142 overseas companies participated in the exhibition, and more than 480 new products were launched).

2. Transaction structure upgrade: From the early stage of goods trade and bulk commodity procurement, to high value-added fields such as technology equipment, service trade, and medical health. Such transactions have longer contract periods, higher technical content, but slower implementation, which is reflected in the growth rate of intended transactions that has stabilized; the cumulative intended transaction volume of the nine sessions has exceeded 660 billion U.S. dollars, and the cumulative number of overseas exhibitors has exceeded 23,000.

3. Exhibitors have shifted from "six-day exhibitions" to "rooted in China and increased investment": The "World Openness Report 2025" shows that China's openness index has increased. The actual use of foreign capital in Shanghai in the first eight months was US$11.59 billion, a year-on-year increase of 6.9%. The cumulative number of regional headquarters and foreign-funded R&D centers of multinational companies reached 1,120 and 675 respectively. There are nearly 200 "full attendance" companies and institutions in the ninth session, reflecting the long-term confidence of leading companies.

9. WuXi AppTec moved out of 1260H: Won the lawsuit but failed to win back trust

1. Timeline and judicial path: On November 9, 2026, the U.S. Department of Defense decided to remove WuXi AppTec from the list of "Chinese military industrial enterprises" under Article 1260H, only 154 days after it was included on June 8 (the list was published on June 10); the company filed a petition with the District of Columbia on June 11 The District Federal District Court sued and obtained a preliminary injunction on August 7 (only a US$1 security was required). The court found that the factual basis for the listing of the Ministry of Defense was flawed (such as misreading the fund asset allocation of 5.32% as a shareholding ratio), and the listing violated the Administrative Procedure Act and caused irreparable business losses.

2. Performance and market reaction: Revenue in the first half of 2026 was 28.897 billion yuan (+38.93% year-on-year), and net profit attributable to the parent company was 11.080 billion yuan (+29.43% year-on-year). The semi-annual report raised its revenue guidance for 2026 from 51.3 billion to 53 billion yuan to 58.5 billion to 60.5 billion yuan; as of the end of June, orders on hand were 66.43 billion yuan (+25.2% year-on-year). After the removal news was announced, WuXi AppTec A shares opened higher that day and rose during the session, with a total market value of approximately 456.1 billion yuan (November 10), which verified that customers had not withdrawn on a large scale.

3. The risk remains: The U.S. Department of Defense reserves the right to re-list based on other or supplementary reasons in the removal notice; the 1260H list is an important system entrance to the "Biotech Companies of Concern" (BCC) list under the Biosafety Act that will take effect at the end of 2025. The list can be adjusted dynamically and annually, and China Micron has a precedent. Customers may add second suppliers for key new projects, and accounts receivable in the first half of the year were 10.472 billion yuan (+44.19% from the end of the previous period), a growth rate that was 5.26 percentage points higher than revenue in the same period. The book growth has not yet been fully converted into cash.

10. iPhone Duo has been on sale for two weeks: DeepSeek and Bilibili are filling in, and Apple has moved out 100,000 compatible apps to support the market.

1. Apple’s first foldable screen mobile phone, iPhone Duo, went on sale on October 23. The starting price in China is 15,999 yuan, and the top-end 2TB version is 26,499 yuan, making it the most expensive iPhone in Apple’s history. On November 9, Susan Prescott, Apple’s vice president of global developer relations, announced that more than 100,000 apps “built for iPhone Duo” have been launched on the App Store.

2. The statistical caliber of "over 100,000 models" refers to applications built using iOS 27.1 and above SDK, which only means that the compiled version is compatible but not that the experience meets the standard: Apple divides the internal screen display into three levels, and WeChat is still stuck in the second level. After expansion, there is a black border in the right navigation area, and the horizontal screen cannot be swiped left to return. Tencent customer service responded on October 22 that "it will be improved through version updates in the future."

3. Apple sets hard constraints: Starting from April 2027, all iOS and iPadOS applications submitted to the App Store must be built using iOS 27 or later versions of the SDK, and come with exclusive screenshots for folding screens; IDC predicts that global folding screen shipments will increase by 12.6% year-on-year to 22.9 million units in 2026, which will be the only increase in the mobile phone market. Apple is a latecomer, and the mature ecosystem still needs to be completed one by one by leading applications.

Open source trends:

11. NVIDIA plugs 120B open source VLM into AWS’s one-click deployment portal

1. NVIDIA Nemotron 3.5 Super VL is officially launched on Amazon SageMaker JumpStart, providing two variants: NVFP4 quantized version (checkpoint is about 75GB, running on ml.g7.24xlarge and other instances) and BF16 full-precision version (about 249GB, running on ml.p5en.48xlarge, H200); quantization reduces the volume by about 70%, and the accuracy gap of five key benchmarks is reduced. Within 1 point, some projects even surpassed it.

2. The model uses a hybrid architecture (Mamba-2+MoE+Attention+C-RADIO visual encoder) with 120B total parameters and only activates 12B per token. It theoretically supports up to 1 million token contexts; however, JumpStart actually only opens 520 to 15,600 tokens, and MTP speculative decoding is not enabled, indicating that the cloud experience depends on the serving configuration rather than the model card parameters.

3. In terms of performance, the first token time of a short request is about 0.45 to 0.53 seconds, and the throughput is 63 to 140 tokens/second; once the input reaches 10,000 tokens (summarize configuration), the first token time rises to 12.62 seconds, and the long context delay needs to be measured according to the business. The model is positioned in agent scenarios such as document intelligence, GUI understanding, and video question and answer. It is licensed under the OpenMDW-1.1 license. The benchmark results are reported by NVIDIA itself. There is still room for third-party reproduction.

12. AWS Quick Skills: Job Description for AI Employees

1. Amazon AWS Quick launches a skill mechanism: it encodes repetitive tasks such as weekly reports, release notes, and work orders into reusable "job descriptions" that users can teach once and execute repeatedly. The core is to convert one-time conversations into reusable assets; this capability is based on the Agent Skills open standard released by Anthropic on December 18, 2025. The Quick desktop version was previewed in April 2026 and generally available in September of the same year.

2. The architecture layer adopts a progressive disclosure design: skill metadata is first injected into about 100 tokens, and the complete instructions are loaded on demand when the agent needs them to cope with the scarcest resource of the context window; thus, enterprises can accumulate dozens or hundreds of skills and only wake up the relevant few at a time. The premise of scale is not a larger context but a smarter loading strategy.

3. The value of early customer data verification reusability: Propulse Lab uses Quick to simplify customer service workflow, reducing the average work order processing time by 80%, and is expected to save more than 24,000 hours per year after planned expansion; Jabil uses it to optimize account collection and quote requests, and is expected to save approximately US$400,000 annually; DXC Technology plans to deploy it among more than 120,000 users, and Vertiv plans to expand its user base in 2026. More than 25%.

(Comprehensive compilation from wide-angle observation, Edge AI Daily, etc.)

According to the real-name requirement of the Cybersecurity Law, please bind your mobile phone number before posting a comment.

Express