Meta’s personal assistant, Muse, has gone viral.
On September 8, Meta officially launched Muse, which rose to number one on the U.S. App Store’s free apps chart ten days after its release. Muse is a personal agent that can read authorized emails, calendars, and files; use a browser to search for information, fill out forms, make purchases, book travel, or continuously process tasks in the background. It offers a free version, as well as paid tiers of $20 and $100 per month. According to market research firm Apptopia, Muse garnered approximately 2.8 million downloads in the first 12 days after its launch in the U.S. and Canada—a growth rate surpassing that of ChatGPT’s mobile app during its early release period. During the same period, Meta’s stock price rose more than 20%, adding about $200 billion to its market capitalization.
Comment: This surge in popularity came somewhat unexpectedly. Over the past year, Meta has rarely been at the center of discussions about cutting-edge models, as its models have underperformed in the market and its AI division has undergone multiple restructurings. Now, Muse is being hailed as Meta’s “ChatGPT moment.” Muse’s sudden popularity has also raised concerns about privacy, security, and data authorization. Amazon has already blocked Muse from making purchases on its platform, signaling that its business model challenges platform access points.
Yan Wentao, a 90s partner at Hillhouse, has been appointed CFO of DeepSeek.
On September 21, Yan Wentai, a 90s-born former partner at Hillhouse Capital Ventures, officially joined DeepSeek as Chief Financial Officer (CFO)—the company’s first-ever CFO appointment since its founding over three years ago. Yan will primarily oversee the company’s capital operations, financial management, and preparations for an IPO on the STAR Market. Born in 1991, Yan graduated from the Department of Environmental Science and Engineering at Fudan University. He began his career at Tencent Investment in 2013, later joining HCapital, and joined Hillhouse Capital Ventures in 2020, where he was promoted to partner. His representative investments include ByteDance, Zhipu AI, MiniMax, Xiaohongshu, and J&T Express.
Comment: Yan Wentaos joining comes at a pivotal time as DeepSeek accelerates its capitalization. In April this year, DeepSeek launched its first funding round, completing approximately RMB 51 billion in financing by June, with a post-money valuation nearing RMB 400 billion. In July, it initiated its second funding round, with a pre-money valuation rising to approximately RMB 500 billion. Meanwhile, DeepSeek has engaged CITIC Securities to prepare for an IPO on the STAR Market, planning to launch the process this year, with CITIC Securities already conducting due diligence.
Alibaba releases its next-generation AI chip, Zhenwu V900.
On September 22, at the 2026 Yunqi Conference, Alibaba unveiled its next-generation AI chip, Zhenwu V900. Alibaba CEO Wu Yongming stated that this is currently the most powerful Zhenwu AI chip in terms of computing performance, offering three times the performance of the M890. It is reported that the annual shipment volume of Pingtouge AI chips will significantly increase.
Review: Previously, the Pingtouge integrated AI chip Zhenwu M890 has been running Qwen3.8. As China’s first super node to successfully run a large model with over 2 trillion parameters, the Zhenwu M890 has been adapted for Qwen3.8 and is now live on Alibaba Cloud’s Bailian platform, offering model inference services. Following the M890, the Zhenwu V900 is scheduled for mass production and release in the first quarter of 2027, also targeting training and inference for trillion-parameter large models.
AMD's market capitalization has surpassed $1 trillion for the first time.
On September 21 local time, AMD's stock rose nearly 10%, closing at $615.52 per share, with its market capitalization surpassing $1 trillion for the first time, boosting sentiment across the global semiconductor sector. In the U.S. stock market, the Philadelphia Semiconductor Index rose over 4%, ARM climbed more than 17%, Intel surged over 12%, Meta increased over 11%, and Qualcomm rose over 9%. Meta, one of AMD’s top customers, further expanded its collaboration with AMD in February this year.
Comment: AMD's recent surge is most directly driven by demand from AI applications. Meta’s new Muse personal agent tool has risen to the top of the U.S. Apple App Store within less than two weeks of its launch. Widespread adoption of agents will significantly increase workloads for inference and infrastructure. AMD’s partnership with Meta, announced in February, includes deploying up to 6 gigawatts of AMD Instinct™ GPUs.
OpenAI and Anthropic release new models
On September 22 local time, OpenAI and Anthropic released new models. On the same day, OpenAI launched GPT-6 Sol and GPT-6 Luna, stating that both models inherit some capabilities from GPT-6 Astra and reduce costs through improved caching and inference efficiency. The company announced that the API pricing for both models is 50% lower than GPT-5.6’s previous promotional rate. Meanwhile, Anthropic released Claude Opus 5.5, claiming that Opus 5.5 performs at the level of Claude Fable 5.1 on most tasks while reducing operational costs by 40% compared to the previous Opus 5 generation.
Review: Unlike previous campaigns that frequently highlighted pushing the limits of capability, both companies this time are focusing on reducing inference costs, improving speed, and scaling real-world usage. OpenAI’s head of core products described Sol as “a better model across the board,” emphasizing overall experience and efficiency rather than chasing record-breaking scores on individual benchmarks.
Elon Musk's SpaceX AI releases Grok 4.7
On September 22, Beijing Time, SpaceXAI officially released its new model, Grok 4.7, which the company describes as its most powerful model to date for programming and knowledge processing. Elon Musk reposted and commented, “Grok 4.7 is a powerful combination of intelligence, speed, and low cost.” According to data from the AI evaluation firm Artificial Analysis, Grok 4.7 scores 46 on the intelligence index, ranking sixth on the model leaderboard, placing SpaceXAI among the top four AI labs globally. The official blog noted that Grok 4.7 can handle complex tasks for longer durations, more carefully review its own outputs, and is equipped with the most comprehensive safety measures. “It’s twice as fast as comparable models and costs only half as much.”
Review: However, based on the initial reviews and user feedback, Grok 4.7 appears more like a targeted effort to close gaps rather than a comprehensive generational leap forward. It shows significant improvement in long-horizon agent tasks and programming tests, but in real-world scenarios such as web navigation, 3D environments, and visual programming, there are numerous cases where strong benchmark scores fail to translate into reliable performance.
Alibaba CEO Wu Yongming: In the future, the total amount of machine thinking will be 1,000 times that of humans.
"Machines are becoming the primary force behind thinking, and intelligence is evolving into a scalable commodity," said Wu Yongming, CEO of Alibaba, during his keynote speech at the opening of the 2026 Cloud Summit on September 22. Alibaba believes the future will enter an era of machine intelligence. He noted that, just as the Industrial Revolution turned power into a capitalized commodity, machine thinking will become a scalable commodity in the AI era: "In the future, the total volume of machine thinking will be 1,000 times that of humans, while today this ratio is less than 3%, leaving at least tens of thousands of times room for growth."
Comment: Currently, machine capabilities are still focused on taking on more physical labor, with the goal of relieving humans of more physical tasks. After solving motion-related challenges, the next objective is to address cognitive tasks. Against this backdrop, Wu Yongming believes that future machine intelligence will become abundant and eventually evolve into a scalable, commoditized offering.
Anthropic signs a $11.6 billion computing contract with Akamai
On September 24 local time, internet infrastructure company Akamai Technologies announced an expansion of its partnership with Anthropic, entering into a seven-year contract commitment worth $11.6 billion. The collaboration will leverage Akamai Cloud’s distributed AI infrastructure and software to meet Anthropic’s rapidly growing CPU workloads. The partnership’s value could increase by up to an additional $9 billion in the future, bringing the potential total commitment to nearly $20 billion.
Commentary: Along with reaching a partnership, both parties have also established an equity relationship. Akamai has issued warrants to Anthropic, granting it the right to purchase convertible preferred shares of the company, equivalent to 7.7 million common shares upon conversion. As an AI company, Anthropic’s computing power demands continue to grow; in July of this year, it also partnered with AMD, planning to deploy up to 2 gigawatts of AMD MI450 series GPUs within AMD’s Helios rack-scale solutions.
Google TPU will be launched into space.
On September 24 local time, Google announced that after years of research, the Project Catcher will launch its first orbital test, sending a prototype satellite into space to evaluate the performance of Google’s Tensor Processing Units (TPUs) in orbit. The project will ride aboard SpaceX’s Transporter-18 mission, which is scheduled to launch on October 1. Project Catcher aims to explore whether space could one day support scalable machine learning infrastructure.
Comment: Last year, Google launched the "Sun Catcher" project, but it is not the only Silicon Valley tech company exploring the idea of building data centers in space. Space computing company Starcloud last year launched a satellite into space carrying an NVIDIA H100 GPU—the first time an advanced data center GPU has been deployed in outer space. The advantage of placing computing power in space is the continuous availability of solar energy; however, heat dissipation remains a challenge.
AI Evolution Update | Moore Threads Unlocks Full Pipeline of Protenix-v2
The team of authors exceeds a hundred, including Liang Wenfeng.
For AI phones to achieve large-scale adoption, they must identify scenarios that are frequent enough, complex enough, and clearly demonstrate value to users.
Muse has reignited market interest in CPU, but the long-term commercial prospects of Agent still depend on factors such as user retention, security and privacy, and platform ecosystem openness.
The global computing power infrastructure is shifting toward CPUs.