Edge AI Daily Morning News (November 12)

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

Edge AI Daily Morning News (November 12)

Edge AI Daily

2026.11.12 08:50

· From Beijing

Full text 8498 words

Silicon Valley Frontier:

1. Anthropic launches US$64.35 billion IPO road show, betting on AI coming of age ceremony with US$1.4 trillion valuation

1. Anthropic will launch its IPO on November 10, 2026. It plans to issue 825 million Class A shares at US$70 to US$78 per share, with a fundraising upper limit of US$64.35 billion. The fully diluted valuation upper limit is US$1.405 trillion, nearly half of the US$965 billion raised in May this year. It plans to price on November 18 and list on Nasdaq under the code "ANT" on November 19.

2. High financial growth and high losses: Revenue in the nine months as of September 30, 2026 was US$33.24 billion, a year-on-year increase of approximately 1,200%. The run-rate at the end of October was approximately US$81 billion, but the net loss during the same period was US$104. 0 billion (US$25.1 billion in the same period in 2025, mostly from changes in the fair value of convertible bonds and derivatives); the implied price-to-sales ratio based on annualized revenue is approximately 17 times, and the valuation expectation has been reduced from US$2 trillion to 1.4 trillion.

3. The IPO uses capital tools to bind the entire AI industry chain: Agreements were signed with AMD, Broadcom, Google, Micron, Microsoft, NVIDIA, and TSMC. Strategic investors privately purchased $26 billion in N-class shares at the issue price without being restricted by the lock-up, and the retail placement was about 20%. The article believes that the listing marks the AI industry's shift from "burning money to tell stories" to "let financial reports speak for themselves." High-priced takeovers in the secondary market face the risk of revaluation.

2. Anthropic enters the office hinterland: Claude Sheets challenges Excel, and targets Microsoft and Google on the eve of IPO

1. Product and market reaction: Anthropic launched the Claude Sheets, Whiteboards, Data three-piece set and the upcoming Claude Meetings on November 10, directly targeting the core office scenarios of Microsoft Office 365 and Google Workspace. After the news came out, Microsoft's stock price fell 3% intraday, and fell another 2% after the Bloomberg report came out.

2. IPO and performance: Anthropic submitted an S-1 on October 26 and plans to list on Nasdaq on November 19 (code ANT). It plans to issue 825 million shares at US$70 to US$78 per share and raise a maximum of 64.4 billion. US dollars, with an upper limit valuation of about US$1.4 trillion; an additional US$26 billion was raised through private placements, for a total of up to US$90.4 billion; as of October 31, annualized revenue reached US$81 billion (71 billion at the end of September and 65 billion in July).

3. Competition and cash-out: Microsoft and Google are both investors and challenged opponents. Microsoft has informed Microsoft 365 Copilot tenants in the UK to enable Claude by default from December 9; at the same time, the seven co-founders will sell a total of US$1.75 billion in stocks in the IPO, each accounting for approximately 1.5% of their total shareholding, and each person will cash out approximately US$250 million based on the midpoint of the issuance range.

3. Nvidia Board Opens Probe as Governance and National-Security Risks Reprice the AI Trade

1. On November 8, NVIDIA’s board of directors announced the launch of an independent investigation due to reports that founder and CEO Jen-Hsun Huang had become the target of foreign intelligence operations; on November 9, a shareholder class action lawsuit was filed in the California Federal Court, accusing the company of insufficient disclosure of its relationship with Singaporean customer Megaspeed during the 13-month collective period from October 9, 2025 to November 8, 2026, and allegedly violated the U.S. Securities and Exchange Act.

2. Market reaction: Nvidia closed down 4.27% to US$178.67 on November 9, and fell another 5.47% to US$168.89 on November 10. The market value is approximately US$4.07 trillion, approaching 5 The 2-week low is $164.27 (52-week high is $243.37); the market value has evaporated by about 26% in the past month. The article calls it an atypical AI cycle callback, but a repricing driven by governance and national security risks.

3. Core logic: The lawsuit focuses on "insufficient disclosure" rather than illegal behavior. The key to victory or defeat lies in the "materiality" of the information; the board of directors' statement of placing the United States before the interests of shareholders has been accused of raising capital costs (increased discount rates, restricted growth options), and the national security risk discount may spread from Nvidia to AMD, Broadcom, TSMC and other AI semiconductor chains.

4. NVIDIA is downgraded, and AI demand loses certainty first

1. On November 11, Baird downgraded Nvidia's rating from "outperform" to "neutral" and lowered the target price by 60% from US$500 (set in May 2026) to US$200; the reasons were limited visibility of data center demand in 2027, supply constraints, US chip export restrictions to China and the increased probability of CEO handover. It is worth noting that before the downgrade, the stock price had fallen by about 29.4% from the closing high of $239.24 on October 6, and closed at $168.89 on November 10. This revaluation is more like a rise in risk premium than a demand-side thunderstorm.

2. This time it is more like a "visibility shock" before a "demand shock": Nvidia's revenue in the second quarter of fiscal year 2026 was US$96.22 billion, a year-on-year increase of 106%, and data center revenue was US$89 billion, a year-on-year increase of 117%, and gave a third quarter 1 Revenue guidance of 08 billion US dollars (plus or minus 2%); Baird predicts that the growth rate of AI capital expenditures of hyperscale cloud vendors may slow down to about 40% or slightly lower in 2027, and reminds that 10% to 30% of advance orders have timing risks of being reduced.

3. Key indicators that need to be verified in the future are: utilization rate, inventory level and capital expenditure return after customer deployment (that is, whether the chips are continuously used instead of being backlogged), as well as risks at the policy and governance level - export controls may prompt customers to "design out" US chips. The company's next quarter guidance does not assume data center revenue from China. The investigation surrounding Huang Renxun has also brought issues of executive transition and strategic continuity into valuation considerations.

5. Nvidia’s AI craze has made power architecture the next major bottleneck

1. The bottleneck of AI infrastructure is shifting from "insufficient computing power" to "mismatch of power architecture": Nvidia's data center revenue in the second fiscal quarter of fiscal year 2027 (as of July 26, 2026) was US$89 billion, a year-on-year increase of 117%. Computing power is concentrated in large clusters at an extremely high speed, and the construction period of power generation, transmission, substations, and grid connection is longer, which has become a key constraint on whether AI projects can be launched on schedule.

2. Electricity demand shows structurally high growth but is highly concentrated in space: The International Energy Agency predicts that global data center electricity consumption will increase from approximately 415 terawatt hours in 2024 (accounting for approximately 1.5% of the world) to approximately 945 terawatt hours in 2030. hours (about 3%), with an average annual growth rate of about 15%; the United States may increase by about 240 terawatt hours (an increase of about 130%) by 2030 compared with 2024, and the United States and China together will contribute nearly 80% of the global increase before 2030. A report from the Lawrence Berkeley National Laboratory in the United States shows that the load of U.S. data centers has increased approximately three times in the past decade, and is expected to double or triple by 2028.

3. The focus of industrial value and competition logic has shifted to "power architecture": scarcity has changed from "how much electricity is there" to "where is the electricity, when is it available, and can it be stably connected and expanded?" The benefit weight of transformers, switchgear, power distribution, liquid cooling, energy storage, energy management software and power engineering services has increased; flexible arrangements such as workload transfer, energy storage, on-site power generation, as well as long-term power purchase agreements and interruptible loads are changing from optimization items to "tickets" for data centers to gain access, control financing risks and maintain asset value.

6. OpenAI’s quarterly revenue exceeded US$10 billion for the first time, and enterprise business doubled

1. Revenue and structural transformation: In the third quarter of 2026, OpenAI's revenue exceeded US$10 billion for the first time, a quarter-on-quarter increase of nearly 50%, approximately three times the same period last year, and an annualized scale of approximately US$50 billion; enterprise business doubled quarter-on-quarter, pushing the consumer and enterprise income ratio to 50/50. Consumer income in a single quarter alone exceeded the entire year of 2025.

2. Cost and gross profit margin bottleneck: The gross profit margin is only 33%, far lower than the 75% to 85% of typical enterprise-level SaaS, and also lower than Anthropic's nearly 85%; inference costs will rise from US$8.4 billion in 2025 to US$14.1 billion in 2026, and cash consumption is expected to be approximately US$27 billion in 2026 and US$63 billion in 2027, showing the marginal cost expansion characteristic of "the more you sell, the harder you burn".

3. IPO reconciliation and new track rules: OpenAI submitted a confidential IPO registration statement on May 22, 2026, and Anthropic has also submitted it. It aims to be listed from the autumn of 2026 to the beginning of 2027; with the unified adoption of GAAP standards, the focus of investors' attention has shifted from "run rate" to gross profit margin and corporate revenue ratio, that is, whether gross profit margin can rise from 33%, and whether corporate revenue can stabilize at 50%.

7. OpenAI puts Google’s 20 years of business into ChatGPT

1. Release and mechanism: OpenAI will launch "Businesses in ChatGPT" (beta) on November 10, 2026. Through Business Hub, companies can claim and update information, upload product feeds with price inventory, and view three types of aggregated data: number of mentions, website clicks, and "discovered topics"; at the same time, "Business" is open to qualified e-commerce and online service companies in the United States. Agent", covering its 1.2 billion weekly users, essentially transforming "how companies appear in the answers" from a black box into a measurable and purchasable commodity.

2. Monetization path: Metrics are the bait, advertising is the hook - OpenAI has been testing ChatGPT advertising in the United States since February 9, 2026. By August 31, the annual advertising revenue run rate reached 1 billion US dollars (it took less than seven months). As of October 21, more than 100,000 advertisers have been advertising in 89 countries; the free Business Hub is intended to introduce a large number of unadvertised companies into the paid advertising ecosystem, reusing Google's path of "free tools first, then paying for traffic".

3. Competition and impact: OpenAI directly targets Google. As of the September 2026 quarter, Alphabet’s search business revenue was US$63.4 billion, accounting for 49.7% of total revenue, which fell below half for the first time; at the same time, third-party GEO startups (OtterlyAI, Peec AI, Profound) relies on simulation and estimation, while the platform holds real usage data, and the information gap narrows. Its role has been reduced from "category definer" to "platform ecological supplement", and its living space has shifted to cross-platform monitoring and prompt word attribution that are not currently covered by the platform.

8. Google Gemini 4 Carbon unveiled: Specialized in gnawing hard bones, the longer you think about it, the more accurate it will be

1. Google officially opened Gemini 4 Carbon on November 10, positioning it as "the most powerful inference model so far", specializing in complex software engineering, multi-step agent tasks, scientific research and network security; it forms a "hard and one soft" division of labor with the efficiency model Argon - Argon is used for daily coding and document checking, and Carbon is used for hard tasks such as reconstructing legacy systems and auditing code bases. The model was first introduced by Koray Kavukcuoglu of DeepMind on October 28. The initial stage was only open to some network security defenders through the Fairwind Program, and has since been pushed to developers, enterprises and consumers.

2. Safety and scientific research are Carbon’s prestige scenarios: SEC-bench evaluation without guardrails 88.2% of Pro vulnerabilities were discovered, 54.8% were exploited in ExploitGym, and 71.0% were patched in CWE-bench. The internal security code generation success rate was more than twice that of Argon. The Chrome team said that the performance of generating high-quality patches was more than ten times higher than other laboratory cutting-edge models. The Wiz team used it to sweep out zero-day vulnerabilities in less than two hours, compressing weeks of vulnerability research into one afternoon. In terms of scientific research, Carbon mines distantly related homologous sequences in multi-agent mode and increases the number of homologous sequences retrieved by an average of 40 times on seven low-homology protein families in the CASP17 blind test set.

3. The price of deep thinking is equally clear: Artificial Analysis, a third-party organization, recorded that Carbon output approximately 200 million tokens in its intelligence index evaluation, which is approximately three times the median of 64 million for similar models. The average single task cost is approximately US$7.76 based on standard pricing. Pricing shares the same launch discount as Argon, with input tokens per million being US$2 and outputs being US$10. The cached input price is 95% lower than the input price, and will rise to US$4 and US$20 after the discount period. In addition, Google adopts a phased approach of opening up to defenders first, then to trusted partners, and finally to the public to deal with the double-edged sword risk that the model can be used to find vulnerabilities and write attack payloads; its security benchmark scores have not yet been fully reproduced by an independent third party.

9. Google releases the most powerful robot model Gemini Robotics 2.5: Not only can it learn to tie bow ties, it has also successfully worked on the Mercedes-Benz production line

1. DeepMind released the Gemini Robotics 2.5 Vision-Language-Action (VLA) model on November 10, 2026. In the evaluation of more than 200 tasks, the task success rate increased by 30% to 100% compared with the previous generation, and the completion speed increased by 20% to 30%. It also realized fine operations such as tying bow ties, picking up food with pliers, and pipetting, marking the transition of embodied intelligence from demonstration to deployment.

2. In terms of industrial implementation, Apptronik's Apollo 2 humanoid robot completed the sorting of car kits on two real production lines of Mercedes-Benz in Kecskemét, Hungary and Sindelfingen, Germany, with a success rate of 96% (8 errors out of 200 times); DeepMind Director Kanishka Rao said that filling the last 4% is a "steep road", which is still at the PoC level and is far from 99.9% industrial-grade reliability. It will take time for large-scale unmanned operation.

3. The performance jump relies heavily on the Gemini 4 Argon simulation engine released on October 13, 2026, which significantly compresses the training cycle (its pure inference on the ERQA benchmark reaches 87%); the current model is still running on the cloud, with risks of delays and network disconnections. The demonstration video is accelerated by 2 times. Google takes the "Android" route of the model layer and has not yet opened up public APIs, weights and training codes.

10. CoreWeave rebounded 17%, can AI orders pass through the financing black hole?

1. CoreWeave’s revenue in the third quarter was US$3.661 billion (a year-on-year increase of 168%, exceeding expectations by US$3.55 billion), with an adjusted EBITDA profit margin of 62%, and a backlog of orders to be delivered of approximately US$132.5 billion (continuing to expand from US$104 billion at the end of June). The above-expected performance alleviated market concerns about the slowdown in AI infrastructure investment, and the stock price rebounded by approximately 17% after the financial report.

2. The other side of good growth is financial pressure: net interest expenses in the third quarter were US$846 million (310 million in the same period last year), and net losses expanded to US$709 million; the 62% adjusted EBITDA profit margin cannot cover depreciation, capital expenditures and interest, reflecting its "asset-heavy, strong financing dependence" infrastructure operation nature, rather than pure high-margin cloud services.

3. The core indicator of valuation is cash creation ability after financing: the company's capital expenditure guidance for 2026 is US$37 billion to US$39 billion, and it is expected to receive approximately US$9 billion in customer advances and deposits in the next three months; investors need to track the conversion speed of order backlog, the conversion of adjusted profits into cash, whether financing costs can decrease with scale, and the volatility risk caused by the concentration of large customers.

11. Microsoft will take the AI grid bill back to itself first

1. On January 13, 2026, Microsoft released the "Community First AI Infrastructure" initiative, promising that the data center will bear its own electricity and related grid infrastructure costs, not seek local property tax exemptions, replenish water that exceeds water withdrawals, and train and hire local residents; in essence, it is trying to write "who benefits, who pays" into the very large user rate design in exchange for local social permission.

2. According to data from the International Energy Agency, global data center-related power generation will increase from 460 terawatt hours in 2024 to over 1,000 terawatt hours in 2030 and 1,300 terawatt hours in 2035. Renewable energy will meet nearly half of the new demand in the next five years; because new loads are concentrated in a few areas with land, network and access conditions, the impact on local power grids, approvals and residential electricity prices is much higher than the global average. Green power certificates purchased by enterprises cannot replace local instant power supply capabilities.

3. The biggest commercial impact of the initiative is to move public infrastructure costs back to Microsoft's balance sheet: upfront capital expenditures and long-term fixed costs will increase, and the cash recovery cycle will lengthen. The costs may be transmitted to Azure's utilization and computing power pricing; if the industry generally adopts "whoever increases the load, who bears the cost", some marginal projects will be postponed, and the AI computing power growth curve may shift from capital-driven to demand-driven. Success or failure depends on contract details such as minimum payments, long-term capacity reservations, and extension responsibilities rather than principle commitments.

12. Amazon Japan’s algorithmic price limit was found to be illegal

1. On November 11, the Japan Fair Trade Commission (JFTC) officially recognized Amazon Japan’s rectification plan: it is no longer allowed to use algorithms to exclude sellers whose selling prices are higher than the lowest price of the same model on other major platforms from the “Recommended Offer” (Buy Box); when there is no recommended offer that meets the price conditions, the purchase offer and entry with the lowest price including shipping must be displayed on the product page. This mechanism lasts for 5 years and is a behavioral correction rather than a penalty.

2. The investigation was launched on November 26, 2024. This is the third time that Amazon Japan has been targeted by the JFTC (in 2016 involving pricing terms, and in 2018 involving soliciting "cooperation funds" from suppliers); according to the Japan External Trade Organization (JETRO), Amazon Japan's share reached 28.2% (in fiscal year 2021). The difficulty in supervision lies in the platform's hiding of coercive intentions in algorithms and replacing contract terms with "automatic determination by the system."

3. Global platform antitrust is shifting from fines and threats of splitting to behavioral remedies: the US FTC will sue Amazon in 2023, the EU's "Digital Market Law" will impose continuous obligations on "gatekeepers", and China's State Administration for Market Regulation will issue fines for "choose one", all for a single behavior; the JFTC's 5-year verifiable and auditable long-term mechanism obligation to constrain the algorithm will continue to affect Amazon's traffic distribution, seller structure and logistics revenue.

13. Musk: SpaceX has ordered HJT equipment with an annual production capacity of 100GW from Singulus to build a "smart sun"

1. Order events and magnitude: On November 10, 2026, SpaceX signed a turnkey supply contract for the entire heterojunction (HJT) solar cell line with German equipment manufacturer Singulus. Musk specified on X that the annual production capacity is 100GW, and the first production line will be delivered in the first half of 2027 and landed at the Bastrop factory in Texas. According to the IEA, the world's newly installed photovoltaic capacity will exceed 600GW and China's will be nearly 370GW in 2025. The 100GW capacity limit is equivalent to about one-sixth of the global annual increase and nearly one-third of China's.

2. Starmind project and HJT selection: Starmind is SpaceX’s space AI computing power satellite network. It has applied to the FCC to deploy up to 1 million satellites in sun-synchronous orbit; the AI1 satellite has a wingspan of 75 meters, an average power consumption of 175kW, and a peak power consumption of 250kW. HJT was chosen because of its high conversion efficiency, radiation resistance, thinness, and adaptability to the stringent efficiency and weight requirements of space. SpaceX provided financial support to Singulus in March 2026 and added a subordinated secured loan in October, showing a vertical integration path of "invest first, place orders later".

3. Industrial impact: Singulus’ revenue in 2025 will be 48.3 million euros, and its first production line alone is worth more than 60 million euros; its announcement positions itself as a “leading Western supplier of next-generation solar cell technology”, reflecting SpaceX’s consideration of supply chain security in the context of the global photovoltaic supply chain being concentrated in China. Brokerage research reports believe that p-type HJT is most suitable for space scenarios due to its radiation resistance, thinness, and low silver consumption. Space data centers may become the second growth curve of HJT technology.

Domestic progress:

14. Cross-border e-commerce exports at the Hong Kong-Zhuhai-Macao Bridge Port exceed 100 billion yuan

1. Scale and growth rate: In the first 10 months of 2026, exports through the Hong Kong-Zhuhai-Macao Bridge Zhuhai Highway Port and the customs cross-border e-commerce management platform reached 112 billion yuan, a year-on-year increase of 33.1%, exceeding the full year of 2025 (2025) two months ahead of schedule. In 2025, the port’s cross-border e-commerce exports reached 110.98 billion yuan, an increase of 92.5%); the average daily export package increased from approximately 1.6 million in 2025 to over 3 million. The volume of packages grew faster than the value of goods, the value of single items of goods declined, and orders became fragmented and frequent.

2. Mechanism and geography: The core of growth relies on the 9610 "list verification and summary statistics" supervision model, which has reduced the customs clearance costs of enterprises by more than 30% on average and improved the timeliness by about 50%; the scale of the Zhuhai Cross-border E-commerce Comprehensive Pilot Zone has increased by more than 166 times in the six years since it was approved. The logistics in the Pearl River Delta has shifted from "goods from the east to the east" to "in and out from the west". Cainiao, Xiyin, Jingdong, Pinduoduo, etc. have deployed warehousing and logistics facilities in Zhuhai and Hengqin.

3. Pain points and prospects: Starting from April 1, 2026, the national customs will promote the return of cross-border e-commerce retail exports across customs zones, superimposing "TIR + cross-border e-commerce" transportation and "pre-supervised warehouses" to respond to difficult returns, slow transportation, and customs clearance respectively. However, the channel is highly dependent on the air transport capacity and route policies of Hong Kong and Macao airports, and faces competition from Shenzhen, Guangzhou Baiyun Airport and other regions. It is expected that the cross-border e-commerce exports of the bridge will reach about 130 billion yuan in 2026.

Open source trends:

15. Nvidia donated OpenShell to open source, not for the sake of agent security.

1. Events and strategies: At the KubeCon North America conference in Salt Lake City on November 10, 2026, NVIDIA launched three sets of actions at once - donating the agent's secure runtime OpenShell to the CNCF Sandbox, contributing code to the Google-led Agent Substrate and connecting with Dynamo inference scheduling and KAI Scheduler, and building the DSX AI Factory Platform has added three new open source components: NVKE, NVKR, and Snapshot, which are intended to move the main battleground of competition from the GPU chip to the orchestration layer of the AI factory.

2. Layered implementation: The governance layer OpenShell provides deterministic control of agent behavior, and officials say it has been implemented in hundreds of institutions; the scale layer Agent Substrate runs millions of agents on Kubernetes at about 10 times the density, and cooperates with Dynamo and KAI Scheduler forms a scheduling closed loop; base layer NVKE and NVKR are in public preview, and Snapshot compresses inference copy recovery to 3.6 seconds to 42.7 seconds, which is 8 to 36 times faster than cold start, and supports checkpoints spanning 8 GPUs and 1TB.

3. Impact and pattern: NVIDIA trades open source for ecology, changing the moat from "code closure" to "ecological dependence"; the cloud native ecosystem and developers benefit from the expansion of options, and manufacturers that rely on private protocols to "collect closure taxes" at the agent runtime and scheduling layer are under pressure; under Google-led Agent Substrate, both parties have consistent interests in making Kubernetes an AI base. The direction of competition depends on whether the profits ultimately flow to the chip layer or the orchestration layer.

(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