AI Weekly: Nvidia CEO replaced, Anthropic launches IPO roadshow
On November 12, Nvidia's board of directors announced that Jensen Huang has resigned as president, CEO and director, and that CFO Colette Kress will serve as interim CEO while a search for a successor begins. Anthropic launched its IPO roadshow the same day, planning to issue 825 million shares at $70 to $78 per share, raising up to about $64.4 billion, with a valuation cap of about $1.02 trillion, below the market's previous expectation of $2 trillion. Google released three models within six weeks, including Gemini 4 Carbon and Gemini 4 Hydrogen.
Body · AI translation
Nvidia CEO replaced
On November 12, Nvidia's board of directors announced that Jensen Huang has resigned from his positions as president, CEO, and board member, effective immediately. Nvidia's current executive vice president and CFO, Colette Kress, has agreed to serve as interim CEO, and until a successor is determined, Colette Kress will continue to perform her duties as executive vice president and CFO while serving as interim CEO. Nvidia's board also announced that it has initiated the CEO search process and will consider both internal and external candidates.
[Commentary] Management turmoil at the world's largest AI chip company has become one of the important variables affecting the direction of the AI chip sector in the near term. As the wave of in-house chip development by cloud providers and AI companies intensifies this year, and as old GPU competitors catch up, this should be a critical period for Nvidia to defend its territory and open up new markets. Whether Nvidia's innovative capacity can be maintained after the CEO change has become a question that is hard for outsiders to avoid when assessing Nvidia.
Anthropic launches IPO roadshow, valuation shrinks
On November 10 local time, Anthropic launched its IPO roadshow, planning to issue 825 million shares at a price range of $70 to $78 per share, raising up to about $64.4 billion. At the same time, the company also plans to raise about $26 billion through private placements to cornerstone investors including Google, Nvidia, and AMD. Based on the roadshow documents and outstanding shares, Anthropic's corresponding IPO valuation ceiling is about $1.02 trillion; on a fully diluted basis, the valuation ceiling is about $1.4 trillion.
[Commentary] Whether calculated on the basis of the roadshow documents and outstanding shares or on a fully diluted basis, Anthropic's valuation ceiling is lower than the $2 trillion the market once expected this year. In addition to Anthropic, reports also say that OpenAI has lowered its new round of financing target from about $30 billion previously to about $20 billion, with its valuation falling from about $1.4 trillion to $1.25 trillion. The two leading model companies are facing valuation pressure almost simultaneously.
Google releases three major models in six consecutive weeks
After Google officially released its flagship model Gemini 4 Carbon on November 11, Beijing time on November 13, Google released another new model, Gemini 4 Hydrogen, the third new model Google has released in six weeks. Unlike the previous two models, Google's new model Hydrogen no longer pursues the upper limit of intelligence, but instead focuses on faster response speed, lower usage cost, and more efficient AI agent execution capabilities. According to Google's explanation, Carbon and Argon are mainly responsible for complex reasoning, long-term planning, and highly difficult software engineering, while Hydrogen is better at quickly writing code, calling tools, and handling daily tasks that require continuous execution. According to the intelligence index ranking of third-party AI evaluation agency Artificial Analysis, Hydrogen currently ranks seventh on the list, while the flagship model Carbon, updated the previous day, ranks second.
[Commentary] In just over a month, Google has almost completed its product refresh from flagship to lightweight models. By contrast, the next-generation flagship model from OpenAI that the market has been expecting has not been released for a long time. "It feels like Google is back, the 'big three' live up to their name, and the one falling behind is now OpenAI," one user lamented. Fortune in the AI large-model industry changes hands, and the pressure now falls on OpenAI.
DeepSeek releases full-suite application API, developers can replicate the DeepSeek App 1:1
On the evening of November 13, DeepSeek announced the opening of its full-suite application API. This update directly integrates the various interactive functions of the DeepSeek App into the original API. After enabling the corresponding parameters, the relevant tool calls and context assembly are automatically completed by the server. In the future, developers can implement complex conversational functions in their own applications, and even replicate the complete DeepSeek App 1:1. When users call the DeepSeek API to use extended capabilities, in addition to token fees, they will also be charged tool invocation fees by category. For example, text/image search and web scraping are 5 yuan per thousand calls, text recognition is 1.5 yuan per thousand pages, and speech recognition is 0.39 yuan per hour.
[Commentary] Previously, in the conversational services of the DeepSeek App (including the web version), capabilities such as web search, file input, link reading, voice interaction, and security worked together with model inference, but for developers, if they wanted to implement similar functions in their applications, in addition to connecting to the original API, they often had to expend effort integrating external tools and handling invocation connections. Therefore, now that DeepSeek has opened this full-suite application API, developers can implement complex functions more conveniently.
OpenAI third-quarter revenue rises to $10 billion
On November 12, OpenAI Chief Financial Officer Sarah Friar said in a media interview that OpenAI's third-quarter revenue grew nearly 50% quarter over quarter to $10 billion, with annualized revenue reaching about $50 billion. That growth rate was significantly higher than the 18% quarter-over-quarter increase in the second quarter. Friar said third-quarter growth was driven mainly by frontier model research and development and the enterprise business. In OpenAI's current revenue mix, the enterprise business and the consumer business each account for half, and "the growth momentum comes from code and cybersecurity products, as well as the gradual rollout of enterprise-oriented intelligent assistants such as Dots." Friar also disclosed that OpenAI's consumer business continued to grow, with revenue up 20% quarter over quarter, paying subscribers exceeding 100 million, and weekly active users reaching 1.2 billion.
[Commentary] However, OpenAI still faces considerable profitability pressure. According to reports, the company's third-quarter operating loss widened from $10.1 billion in the previous quarter to $10.9 billion, although revenue growth has outpaced the expansion of losses. In the same period, Anthropic's third-quarter revenue was about $17 billion, still higher than OpenAI's. However, the two companies use different revenue accounting methods.
Zhang Yiming on AI: Catching up with OpenAI and Anthropic no earlier than 2029 or 2030
Reports say that in a recent internal sharing session, ByteDance founder Zhang Yiming discussed the AI business. He said ByteDance is capable of competing with OpenAI and Anthropic, but truly catching up with frontier technology will take "at the earliest 2029 or 2030." Compared with top overseas AI companies, the pre-training dataset scale of ByteDance's Seed series models is still smaller, which is the most urgent issue for ByteDance AI at present.
[Commentary] Zhang Yiming previously stressed at an internal meeting that ByteDance will not use model distillation technology to accelerate AI model development, even if that decision means the Seed models will temporarily lag behind some domestic competitors in performance and evaluation rankings in the short term. This time, his remarks focused on the gap between ByteDance and leading overseas AI companies. Recent market reports say ByteDance is pre-training a model with up to 10 trillion parameters, which may be a product benchmarked against overseas frontier models.
Li Yanhong responds to Baidu "rising early but arriving late at the market"
On the evening of November 12, Baidu released a full video of digital human Xijiajia interviewing Baidu Chairman Li Yanhong. Li Yanhong was asked many sharp questions, among which the most provocative was a classic outside assessment of Baidu: "rising early but arriving late at the market." Li Yanhong responded that he thinks there is nothing wrong with rising early, and the reason outsiders feel Baidu arrived late is that Baidu likes to innovate, and innovation has a high failure rate, with the possibility of success and the possibility of failure.
[Commentary] Li Yanhong said he psychologically accepts that 90% of attempts will fail; if an attempt fails, outsiders say Baidu arrived late, but in his view, this is the price that innovation requires.
He also disclosed that he still spends most of his time on application-layer issues, and he admitted that the industry has not yet found a true killer app. If it still cannot be found in a few more years, the industry will face a new round of reshuffling.
Google releases TPU to challenge Nvidia
On local time November 9, Google announced the launch of two new TPUs (Tensor Processing Units): TPU 8tx and TPU 8i Max, aimed respectively at security and confidential inference and ultra-large context processing. Unlike the past focus mainly on computing performance improvements, Google this time focused on demand brought by AI agents (AI Agent), and the blog said the new TPUs are "built for the agent era." Among them, TPU 8tx focuses on hardware-level data security and can run and train AI models while processing sensitive data; TPU 8i Max, through a dual-chip design, doubles computing power, memory capacity, and memory bandwidth compared with TPU 8i to support million-token-level context processing.
[Commentary] This is Google's second update to its TPU product portfolio this year, and the first time it has launched four TPUs in one year. Although Nvidia still holds an important position in the AI computing power market, Google is trying to expand market space through more products and sales channels, posing a challenge to Nvidia. However, to compete with Nvidia, Google cannot rely on hardware performance alone. Competition in AI chips is multidimensional, including software and hardware ecosystems, deployment flexibility, and overall computing costs. Whether Google can win more orders may still need market and time to verify.
Li Yanhong in dialogue with Yang Yuanqing, on intelligent organizations
At the Baidu World Conference, Baidu Chairman Li Yanhong and Lenovo Group Chairman and CEO Yang Yuanqing held a dialogue. Li Yanhong said that in the past few years, Baidu has also stepped into some pitfalls during its AI transformation, from business departments training models on their own to chasing agent application hotspots, and Baidu gradually discovered that employees' high enthusiasm for using AI does not necessarily mean organizational efficiency will improve. Yang Yuanqing said Lenovo's "fifth entrepreneurship" revolves around AI, and its three major businesses—intelligent devices, intelligent infrastructure, and intelligent solutions and services—will all change because of AI.
[Comment] Why are companies investing more and more in AI, yet their revenue and profits are not growing along with it? Li Yanhong borrowed a metaphor from the industrial age: companies have merely replaced the motor without redesigning the factory, so the benefits naturally are not fully released. He proposed three principles for organizational evolution: breaking down job boundaries, accumulating tacit knowledge, and designing closed-loop processes. As for the AI input-output ratio that is hotly debated in the market today, both believe that in the early stage of a revolutionary technology's development, it is still too early to discuss the input-output ratio.
MiniMax open-sources the H3-Image image model
On the evening of November 10, three months after the release of MiniMax H3, MiniMax (00100.HK) released the image generation model MiniMax H3-Image and simultaneously opened the model weights. It is understood that MiniMax H3-Image supports text-to-image, image reference, and image editing, and can generate a high-specification image of 4096×4096 pixels, about 16MP (4K), within 10 seconds, at a speed up to 5 times that of mainstream levels on the market; in terms of image quality, the model provides rich image details and a strong sense of realism, further narrowing the distance between AI-generated images and the real visual world.
[Comment] According to the actual test feedback of many netizens, the level of H3-Image can already go toe-to-toe with the world's top image generation models, reaching SOTA level among open-source image models. On the first day of open-sourcing, MiniMax, together with global chip companies such as Huawei Ascend, T-Head, Kunlunxin, Moore Threads, Biren Technology, and Nvidia, as well as multiple inference and cloud service platforms, advanced integration work.
CAS researcher says computing power rent has doubled over the past year
"At present, the biggest problem the institute encounters is computing power," said Li Guoqi, a researcher at the Institute of Automation of the Chinese Academy of Sciences, on November 11 at the Conference on Social Forces Supporting Basic Research held at the Shanghai Science Hall. Li Guoqi said that over the past year, the rental price of high-end computing power has risen by nearly 100%, from more than 20,000 per unit last year to more than 40,000 per unit this year. A large amount of computing power and data is highly concentrated in leading big companies, and it is difficult for non-mainstream, non-consensus original research to obtain long-term support from enterprises. He hopes that social capital can give scholars more support on the computing power foundation.
[Comment] Basic research is the source of scientific and technological innovation, but at a time when AI is reshaping the scientific research paradigm, how to invest money and to whom to invest it is becoming a difficult problem. Some people in the AI industry and scientific research community believe that the intervention of social forces in basic research usually occurs in the "middle ground" where public funding is insufficient, commercial capital is unwilling to enter, but the scientific value has not yet been fully recognized. With computing power rents soaring, social forces can support basic research.
On November 12, SMIC released its third-quarter financial report. In that quarter, SMIC's sales revenue was US$3.19 billion, up 6.1% quarter-on-quarter; gross margin was 29.9%, up 4.6 percentage points quarter-on-quarter. Another leading domestic manufacturer, Hua Hong Semiconductor, has also released its third-quarter performance report. Hua Hong's sales revenue in the same period was US$987.1 million, up 19.9% year-on-year and 9.9% quarter-on-quarter, a record high; gross margin was 24.6%, up 7.6 percentage points year-on-year.
[Comment] The overall improving performance of the two domestic wafer foundry "giants" benefits from the wave of AI development. Co-CEO Zhao Haijun said that the record-high revenue in the latest quarter was mainly due to customer demand for AI-related chips growing faster than expected. At Hua Hong, at least more than 50% of downstream end products are driven by AI development. Against the backdrop of continued spillover of AI demand, the capacity of both leading manufacturers in the quarter was close to or exceeded full load.