AI Weekly: Meta's Muse surpasses 10 million downloads; CXMT and Micron diverge on technology

DeepSeek Open-Sources Inference Framework and Discloses Its Books

On October 20, DeepSeek announced the open-sourcing of its inference framework DeepInfer, the core technology behind the high-throughput, low-latency online inference services for DeepSeek's large models. This open-sourcing is of great significance to the AI industry. However, what drew more attention than the open-sourcing was that, in the technical report released simultaneously, DeepSeek disclosed throughput and key revenue data for its cluster services. DeepSeek noted that, based on this inference framework, the team has achieved high-throughput, low-latency large-scale inference services for DeepSeek V4.1 Flash. Currently, a single cluster's service throughput is 155 trillion tokens per day, and overall revenue is $38 per GPU per hour (at valley pricing).

[Commentary] Industry analysts believe that the continued improvement of open-source models and Agent toolchains is expected to lower application development costs and drive AI Coding's expansion into scenarios such as game development, 3D content production, enterprise software, and automated workflows. Platforms with developer entry points and application ecosystems are expected to benefit first.

Meta's Muse Tops 10 Million Downloads

On October 21, data released by market intelligence firm Sensor Tower showed that since its launch in September, Meta's AI assistant has reached 10.9 million cumulative downloads, of which 10.2 million came from the United States. Muse has become one of the mobile apps to reach 10 million downloads in the United States within 40 days of launch. A Sensor Tower analyst said that so far, only 11 other apps have ever reached this milestone after launch. The most recent app to achieve this was Threads, launched by Meta in July 2023.

[Commentary] Meta's success is closely tied to its ability to use its other services to promote new apps and rapidly grow user scale. For example, Meta previously used this strategy to promote Threads, and now it is applying it to Muse. The total daily active users across Meta's apps have reached 3.6 billion, and at the same time, Meta is spending heavily on advertising to attract new users to Muse.

NVIDIA Launches DSX POD to Shorten Data Center Deployment Time

In NVIDIA CEO Jensen Huang's GTC Berlin keynote on October 21 Beijing time, NVIDIA released the DSX POD AI factory reference design. The product is built inside modular containers and is integrated and validated before delivery, with global deliveries set to begin in early 2027. Huang said that compared with traditional data center deployments, DSX POD can shorten the time from the start of deployment to token generation from six months to as little as 14 days—a transformation of the supply chain.

[Commentary] Shortening the installation time of computing nodes in data centers is aimed at enabling data centers to start making money sooner. According to Huang's calculations, a data center costs about $50 billion per gigawatt to build; assuming it generates $40 billion in revenue per year, a three-month delay in assembly would mean $10 billion less revenue. Launching DSX POD is also one of NVIDIA's measures to stimulate continued global AI infrastructure construction.

CXMT and Micron Diverge on Technology

ChangXin Memory Technologies (CXMT) earlier this month launched a 4F² DRAM (dynamic random-access memory) architecture. At a technology seminar held on October 21 local time, Micron Technology President and Chief Technology and Product Officer Scott DeBoer responded that Micron will follow the 6F² path and develop 3D DRAM, and will not pursue 4F², citing increased complexity and cost. He said Micron has studied 4F² for 20 years and believes there is no viable scaling path in that direction, with no practical benefit to pursuing it.

[Commentary] CXMT's 4F² architecture is said to offer a higher transistor count per unit area than the current mainstream 6F² architecture and to have escaped process bottlenecks to some extent, drawing overseas attention after its launch. But Micron has taken another path, following the 6F² route, adopting more advanced process nodes, and developing 3D DRAM. Which of the two technology routes is superior remains to be verified.

Anthropic Launches a Personal Agent

On October 21 local time, Anthropic launched Claude Duet, a personal agent app, opening research preview waitlist applications to Claude Pro and Max subscribers in the United States. The product is based on Claude Opus 5.5 and can connect to more than 6,000 apps and plugins, proactively organize to-do items, draft replies, prepare payroll and financial materials, and continue executing tasks while the user is offline; for sensitive operations such as sending emails and making payments, it requests confirmation. Scenarios Anthropic demonstrated include distilling important items from email and calendar every morning, identifying unpaid invoices and drafting payment reminders, and preparing financial models, presentations, and business dashboards.

[Comment] Before Claude Duet, products related to leading models such as Meta Muse, xAI GrokBot, and OpenAI Dots had already begun entering the personal agent track, attempting to turn model capabilities into products that continuously execute tasks. Leading domestic manufacturers are also accelerating. Alibaba's Qwen Go is embedded in the Qwen app and has opened experience applications, covering tasks such as claims processing, appointments, price comparison, and food ordering; Tencent Yuanbao is developing an independent app, "Yuanbao Assistant," to carry long-cycle tasks and long-term memory; ByteDance is reportedly set to launch an independent personal assistant, "Xiaodou" (the Spell project), planned to go online as a standalone app and currently still in beta testing.

Is 55% of domestic computing power used for video and short dramas? Doubao spokesperson refutes rumor

Recently, a topic about "more than half of domestic AI computing power is being used to create short dramas" attracted attention and discussion online and once trended on Weibo. In response to the rumor, on October 19, Liu Xing, head of public relations at Doubao, posted on Weibo to refute it, pointing out that the relevant data was inaccurate. "As far as ByteDance is concerned, only about 10% of our company's total computing power is used for training and inference of video generation models." A previously circulated "investor relations activity record" disclosed information showing that 55% of domestic AI token consumption is used for short dramas and video generation, and that token consumption for AI short dramas and video generation has reached more than twice that of software development. Liu Xing said that the computing power data released by a certain company was unreliable.

[Comment] This year, AI short dramas have become many people's "electronic pickles," while also bringing a lot of controversy, especially triggering discussions about "short dramas lowering intelligence." This rumor once again focused public attention on the topic of short dramas. In essence, this is a collision of views on how AI should be applied. Some believe AI should be deployed in core technology to benefit the people, while others believe AI can entertain the public and must first survive.

Report says DeepSeek is about to complete 100 billion yuan financing

On October 22, market news said that DeepSeek plans to raise about 100 billion yuan (about $15 billion) in its latest funding round, with Tencent and CATL as the largest investors in this round, and the deal is expected to be completed within the next few days. DeepSeek, Tencent, and other relevant investors have not publicly responded. A media outlet cross-verified with multiple sources close to the company and investors and confirmed that the above information is true.

[Comment] Previously, reports said that DeepSeek's fundraising target for this round was initially set at 50 billion yuan and later raised to about 100 billion yuan, but the final amount is still not fully determined. The funds will continue to be used for model training, computing power construction, and talent recruitment to support DeepSeek's participation in global competition. Some analysts also believe that this round of financing will pave the way for the company's planned IPO on the STAR Market.

Five optical module companies respond collectively: operations normal

At noon on October 23, five A-share optical module stocks issued announcements. Zhongji Innolight said the company's production and operations are normal, the supply chain is stable, and order demand is strong; TFC Optical Communication said the company's production and operating order is normal, orders on hand are sufficient, and no major abnormal changes have occurred in the internal and external operating environment; Eoptolink said production and operations are being carried out in an orderly manner, orders on hand are sufficient, and shipment pace is expected to accelerate. In addition, Yuanjie Technology said the overall development trend of the industry is normal and the company's production and operating activities are proceeding normally. Accelink Technologies said the company's production and delivery order is normal. Multiple companies said industry prosperity is relatively high or there have been no major changes in fundamentals.

[Comment] Recently, against the backdrop of market doubts that AI infrastructure construction may slow, the stock prices of multiple U.S. AI computing power-related stocks have fluctuated. The stock prices of multiple A-share optical module manufacturers have also declined to varying degrees. From September 22 to October 22, the stock prices of Zhongji Innolight, Eoptolink, and TFC Optical Communication fell by 37.19%, 43.51%, and 29.46%, respectively, while the stock prices of Yuanjie Technology and Accelink Technologies fell by 49.93% and 34.15%, respectively. This collective response reflects optical module manufacturers' hope to restore market confidence.

Jensen Huang says he is open to cooperation with ChangXin

At the close on October 23, ChangXin Technology rose 7.93%, with a market value of 2.91 trillion yuan, an increase of 213.8 billion yuan. On the news front, when asked whether Nvidia plans to cooperate with ChangXin Technology, Nvidia CEO Jensen Huang said in an interview in Germany on October 22 local time that Nvidia holds an open attitude and that the company cooperates with Chinese companies, and that almost the entire global supply chain cooperates with Chinese companies. Huang also said that memory oversupply is unlikely in 2029, and that memory can always be used more and more.

[Comment] From September 22 to October 22, ChangXin Technology's stock price fell 30.2% in one month. After its stock price showed weakness recently, ChangXin Technology's stock price rebounded significantly on October 23. In addition to Huang's remarks, Goldman Sachs recently said in a research report that the median revenue forecast disclosed by ChangXin Technology for the third quarter was 30% higher than the institution's previous forecast, and the median net profit forecast was 17% higher than the institution's previous expectation. Goldman Sachs slightly adjusted ChangXin Technology's target price from 129 yuan to 130 yuan.

Cambricon's revenue and net profit both double in first three quarters

On the evening of October 23, AI chip maker Cambricon released its third-quarter financial report. In the quarter, Cambricon's revenue was 3.531 billion yuan, up 104.47% year on year, and net profit was 1.404 billion yuan, up 147.74% year on year. In the first three quarters of this year, Cambricon's revenue was 9.526 billion yuan, up 106.76% year on year, and net profit was 3.715 billion yuan, up 131.49% year on year. Cambricon said the year-on-year revenue growth was mainly due to continued market expansion and efforts to help AI applications land.

[Comment] In the third quarter, both Cambricon's revenue and net profit hit record quarterly highs since listing. However, in the quarter, Cambricon's net cash flow from operating activities was negative, at -83.678 million yuan. As of the end of September, total liabilities were 7.236 billion yuan, up 56% from the end of June. In addition, as of the end of September, Cambricon's inventory reached 11.665 billion yuan, which may mean Cambricon is actively stocking up, but high inventory also carries risks.