CXMT Loses 1 Trillion Yuan in Market Value in Two Months: The Industrial Shifts in AI Hardware and Apple AI's Entry into China

Alibaba Ramps Up AI Hardware Push, Links Tmall and AliExpress for Global Expansion

Alibaba is entering the AI hardware arena in a dual role as both investor and channel partner. According to Leiphone reports, Alibaba has recently approached a batch of AI hardware brands, offering not only angel-round investment but potentially also linking Tmall and AliExpress to provide global market expansion support. AliExpress will hold its annual and Double 11 merchant conference on October 19, and the public agenda unusually features Tmall. Multiple sources suggest Alibaba may announce a coordinated domestic and overseas plan for AI hardware brands at the event.

The significance of this news lies in the fact that it marks Alibaba's attitude toward AI hardware upgrading from "selling goods" to "building an ecosystem." Previously, the biggest bottleneck for AI hardware startups was not technology but channel access and overseas expansion capabilities—fierce domestic price competition and lack of overseas pathways are exactly the gaps that Alibaba's e-commerce system is best positioned to fill. If this chain of investment + Tmall domestic launch + AliExpress overseas expansion proves viable, the appeal of aligning with Alibaba will rise significantly for AI hardware entrepreneurs.

Apple AI for China Nears: Baidu and Alibaba "Dual Partner" Structure Finalized

The path for Apple Intelligence's entry into China is finally clear. According to foreign media reports, Apple plans to officially launch Apple Intelligence through an iOS 27.3 system update in December this year. The first batch of features includes real-time translation, writing tools, and image removal, with Siri AI not included in the initial rollout. Cook recently revealed that Apple Intelligence has received approval in China.

More noteworthy is the partner structure. According to foreign media reports and verification through iOS 27 Beta 2 firmware teardown, Apple will embed Baidu's software services in Siri voice assistant and visual intelligence features, with a new module called "Baidu Visual Search" added to the system. Meanwhile, Alibaba will integrate AI capabilities into Apple Intelligence, providing text and image understanding, content generation, and other experiences for Chinese users across iOS, iPadOS, macOS, and visionOS, with licensing covering Tongyi Wanxiang and Qwen.

This dual-partner model is unusual in Apple's global strategy. Baidu handles search and visual entry points, while Alibaba handles model capability output. The division of responsibilities actually reflects Apple's pragmatic choice under China's AI regulatory framework—rather than binding to a single partner, letting two companies each focus on different areas satisfies compliance requirements while preserving negotiating leverage.

Apple Intelligence's device-cloud collaborative architecture is the prerequisite for understanding this partnership. Not all of Apple's AI features run on-device: lightweight tasks (such as some writing tool capabilities) are handled by on-device models, while complex requests are forwarded through "Private Cloud Compute" to Apple's custom server chips running large models. In the Chinese market, due to regulatory requirements for data export and model registration, this cloud segment must be handled by licensed domestic large models—which is exactly where Baidu and Alibaba come in.

The generative AI service registration system explains why Apple must choose domestic partners. According to relevant Chinese regulations, providing generative AI services to the public requires algorithm registration and security assessment, and overseas models cannot be directly deployed. Alibaba's Tongyi series (Qwen, Tongyi Wanxiang) and Baidu's Ernie have both passed registration and have compliant output capabilities. Apple's dual-partner structure has an implicit benefit: when one party's model capabilities or regulatory status changes, Apple can adjust task allocation, avoiding lock-in to a single supplier.

CXMT Market Value Drops 1 Trillion Yuan in Two Months, A-Share "Leader" Changes Hands

CXMT's market value rollercoaster is the most dramatic event in the semiconductor sector recently. On October 15, CXMT's stock price closed at 43.59 yuan per share, with total market value falling below 3 trillion yuan, and ICBC reclaiming the top spot in A-share total market value. On August 17, CXMT's closing price was 61.8 yuan per share, with market value once reaching 4.2 trillion yuan. From August 17 to October 14, the stock price fell 25.57%, with market value evaporating by 1.07 trillion yuan.

Interestingly, during the market value plunge, CXMT's technological progress did not stop. On October 10, CXMT Storage President Cao Kanyu stated that the company expects to launch a new generation of DDR5 RDIMM products using the 4F² DRAM architecture by the end of the year. The sharp contrast between technological advancement and capital market performance indicates that this round of decline stems more from valuation-level correction rather than fundamental deterioration—the initial market value peak at IPO itself contained substantial emotional premium.

4F² DRAM architecture is the term most worth explaining in CXMT's technological progress. DRAM memory cells consist of transistors and capacitors, with traditional designs being 6F² (F being the minimum feature size), meaning each cell occupies an area of 6F². The 4F² architecture, through designs such as vertical channel transistors, compresses cell area by about one-third, achieving higher storage density and lower unit cost on the same wafer area. This is one of the mainstream paths for DRAM to advance to higher density nodes, with Samsung and SK Hynix also pursuing similar directions. If CXMT mass-produces 4F² DDR5 RDIMM on schedule by year-end, it means it is approaching the international first-tier ranks in advanced DRAM architecture, which is also the technical basis for judging that its market value correction is valuation-driven rather than fundamental deterioration.

DeepSeek Experiences at Least 18 Outages This Year, Compensation Plan Sparks Discussion

DeepSeek^Ad issued a compensation notice for the API service outage on the morning of October 15: full refunds in the form of gift credits based on users' account consumption amounts over the past 7 days (10.8-10.14). After the notice was published, many users shared their credited amounts, with some receiving just a few cents and heavy users receiving over 3,000 yuan. Some users commented that "the compensation is quite sincere," while others joked that this was a "Liang-style quota reset."

But beyond the compensation, a set of data deserves mention: since 2026, the DeepSeek platform has experienced at least 18 large-scale service anomalies, averaging one to two per month, with September alone seeing at least 7 service fluctuations. For developers who integrate APIs into production environments, stability issues are more critical than price. DeepSeek has not yet explained the cause of this outage, and the connection between frequent downtime and its computing power supply capacity is a continuing focus of concern in the developer community.

Chips and Computing Power: The Infrastructure Race from ByteDance to Musk

ByteDance is in deep negotiations with MediaTek regarding the procurement of custom chips for data centers. Sources familiar with the matter revealed that MediaTek is expected to supply a certain quantity of ASIC chips to ByteDance. If the deal is reached, ByteDance will become MediaTek's second major ASIC customer after Google. Previously, ByteDance had reached a similar procurement arrangement with Qualcomm, planning to purchase millions of AI chips. The background for this series of moves is that ByteDance cannot obtain NVIDIA's most advanced AI chips, while domestic suppliers like Huawei cannot yet meet all demand.

Another direction comes from Musk. His chip project Terafab is being rapidly constructed, and according to people familiar with the matter, it may welcome its first ASML EUV lithography equipment by mid-next year. Musk has hinted that Terafab plans to use free electron lasers (FEL) to replace ASML's laser-produced plasma light source, but the move to purchase ASML equipment in advance indicates that the first-generation product will still rely on conventional solutions.

ASML CFO Roger Dassen explicitly opposed the U.S. Congress's proposal to restrict DUV lithography equipment sales to China during an earnings call, stating bluntly that "this is a bad idea" and arguing that a ban would accelerate China's related technology development. According to ASML's Q3 earnings report, the Chinese mainland market accounted for 23% of its total net sales, and is expected to stabilize at around 20% for the full year of 2026.

The difference between ASICs (Application-Specific Integrated Circuits) and general-purpose GPUs is key to understanding ByteDance's procurement logic. ASICs are chips customized for specific algorithms or tasks, with Google's TPU being a typical example. Compared to NVIDIA GPUs, ASICs offer higher energy efficiency and lower cost per unit of computing power for specific inference or training tasks, but lack flexibility—once algorithms undergo major changes, chip value decays rapidly. ByteDance's choice to procure ASICs from MediaTek and Qualcomm is essentially, under the constraint of being unable to obtain NVIDIA's most advanced GPUs, trading a customized approach for certainty in computing power supply, at the cost of bearing the mismatch risk between algorithm iteration and chip固化. MediaTek previously focused mainly on mobile SoCs, and winning ByteDance's order would be a key step in expanding its data center business from Google to a broader customer base.

The EUV lithography and light source technology route debate also deserves elaboration. EUV (extreme ultraviolet lithography) is currently essential equipment for mass-producing advanced processes below 7nm, with ASML being the world's sole supplier. Its equipment uses laser-produced plasma (LPP) light sources: high-power lasers bombard tin droplets to produce 13.5nm extreme ultraviolet light. The free electron laser (FEL) mentioned by Musk is an alternative approach that theoretically could provide higher power and more stable light sources, but is extremely complex to engineer and has no commercial precedent to date. Terafab's advance purchase of ASML equipment indicates that its first-generation product still follows a mature route, with FEL more likely being a medium-to-long-term reserve.

Talent Mobility and Organizational Restructuring

Several personnel changes are worth noting. OpenAI infrastructure head Sachin Katti announced he will leave on November 15, with Chief Technology Officer for Compute Uday Ruddarraju taking over to lead the overall computing department. This is the latest in a series of executive departures at OpenAI this year.

Domestically, world-class white-hat security researcher Gu He (Yuki Chen) has officially joined Zhipu, entering the GLM foundation model team, with the core mission of enhancing large models' cybersecurity capabilities. Zhipu disclosed that 23 Windows vulnerabilities discovered by Gu He and his team using GLM models have all been confirmed and acknowledged by Microsoft MSRC.

Additionally, Alibaba DAMO Academy's head of AI for Science Rong Yu has left to start a company called "Gezhi Emergence," targeting the AI for Science direction, and has received investment from leading investment institutions. Meanwhile, 17-year-old prodigy Chen Guangyu announced his departure from Kimi; he was one of three co-first authors of the paper "Attention Residuals," which received praise from Musk and later became one of the key technologies of Kimi K3. Talent in AI for Science and security is accelerating its flow from big tech to startups and specialized teams.

AI for Science is the context behind the direction Rong Yu's startup is targeting. This field refers to using large models, graph neural networks, and other methods to accelerate scientific discovery, with typical applications including protein structure prediction (AlphaFold), materials screening, drug molecule generation, and weather forecasting. Compared to general-purpose conversational models, AI for Science has higher requirements for data specialization, model interpretability, and computational precision, and its commercialization path typically involves collaboration with research institutions, pharmaceutical companies, and materials companies. A number of startups in this direction have emerged in China in recent years, with rising capital attention driven by: intensifying homogenized competition in general-purpose large models, while scientific scenarios have higher barriers and clearer willingness to pay. Rong Yu's departure from the position of head of AI for Science at DAMO Academy to start a company reflects precisely how this niche track is splitting from big tech research institutes into independent companies.

One-line Summary

Alibaba enters AI hardware with investment and channels, Apple's China AI solidifies Baidu-Alibaba dual partnership, CXMT's market value evaporates by a trillion in two months.

Paragraph Summary

This morning report focuses on multiple key developments in AI and semiconductors. Alibaba is entering the AI hardware track in a dual role of angel investment plus Tmall and AliExpress channels, with the Double 11 merchant conference potentially announcing a domestic and overseas coordinated plan to help entrepreneurs overcome channel and overseas expansion bottlenecks. Apple's AI entry into China is clear, launching with iOS 27.3 in December, with Baidu handling visual search and Alibaba providing Tongyi series model capabilities, forming a rare dual-partner structure. On the capital markets side, CXMT's market value fell from 4.2 trillion yuan in August to below 3 trillion yuan, but it will still launch a new generation of 4F² DRAM DDR5 products by year-end, with the decline being more of a valuation correction. DeepSeek experienced at least 18 service anomalies this year, with this instance offering full refunds in gift credits based on 7-day consumption. On the computing power front, ByteDance is negotiating ASIC procurement with MediaTek, and Musk's Terafab is advancing EUV equipment acquisition. On the talent side, OpenAI, Zhipu, DAMO Academy, and Moonshot AI all had important changes.