The competition in the field of AI is far from the moment when the gold medal is closed. In the past week, from OpenAI’s acquisition of Jony Ive, the former chief design officer of Apple, in an all-stock deal at a valuation of $6.50 billion, to Anthropic’s release of Claude 4, which has taken another step forward in programming capabilities, the industry has still been busy. However, even in the noisy update, the Google I/O developer conference is still the one with the greatest “stamina”. In the AI circle, people are still trying to understand the transformation of Google and the entire AI wave through this year’s new I/O conference. In the past year or so, Google has seemed to stumble in the competition of generative AI. But this year’s I/O conference, whether it is the all-round upgrade of the Gemini 2.5 series, another generational iteration of Veo 3 sound and painting synchronization, or the official entry of Project Astra into real products, serving all users in the form of Gemini Live, as well as the “AI Mode” that announced that AI will fully take over search, and the penetration of the ambitious AI Agent mode in the whole product line, all clearly sent a signal: Google not only stabilized its position, but also began to occupy a favorable position in the new stage of competition. The reason why Google I/O has triggered continuous discussion is far more than the power of these products themselves, but also because it occurs at a subtle “potential energy transition” moment. Why Google, which was once considered to “get up early and catch up late”, can turn the situation around in a short period of time? One direct reason is the rapid response and focus of the entire organization. A noteworthy signal is that compared with the executive lineup who sat in front of the Silicon Star people after the Google I/O meeting last year, the “scope of responsibility” of the Google executives sitting in front of the Silicon Star people this year has been clearly divided, and the responsibilities of the main business and technical departments have been sorted out more clearly internally. The deeper reason behind this is rooted in Google’s full-stack thinking on AI and the long-term accumulation of “old books”. As Google CEO Sundar Pichai mentioned in his exchange with Silicon Star people, the development of TPU began nearly two decades ago, and Waymo’s accumulation is not a day’s work. This full-link control from chips, models, platforms to applications, and the continuous investment of technology and company trough according to its understanding of technology, allows Google to quickly mobilize resources and optimize the efficiency of the entire technology stack today. This potential turnaround stems from a deep understanding of the core business and corresponding technologies over a longer period of time. Google has proved two things with its actions: First, no matter how iterative AI search is, its foundation is still inseparable from the massive data accumulated by the traditional search business, user understanding and crucially – infrastructure. New AI experiences, such as AI Mode, essentially superimpose a layer of intelligent summarization and guidance driven by large models on top of existing search results. Second, no matter how strong the “sense of generation” brought by large models is, it is still part of the development process of AI technology. The past accumulation in natural language processing, machine learning, Knowledge Graph and other fields is not a write-off, but has become the cornerstone of rapid iteration and performance improvement of new models. The essence of the search business is to transform algorithms in the laboratory into products that meet the real needs of hundreds of millions of users. This process has taken Google for more than two decades. From the original PageRank to the later search, to the revolutionary improvement of BERT’s understanding of search, to today’s Gemini. Google is well-versed in how to turn cutting-edge engineering work of algo, productization, and build a sustainable business model around it. This experience makes Google better able to grasp the rhythm from technological maturity to practical application when facing the new technological wave of big models, and promote the continuous rotation of the innovation flywheel. Google’s return to the top has made many people in the AI industry curious about who can match Google’s full stack and long-term accumulation capabilities in the next competition. And obviously, such a company needs to be there from the beginning of the story, like Google, and has been on the table, until today. In their conversation with the Google search team, they mentioned the process by which Page Rank technology eventually became a valuable landing application for search, which also brought a strong reference to the decision-making process in the landing process of Google’s big model today. Another company that completed the same process in search during the same period was Baidu. The direct competition between Google and Baidu more than a decade ago, although in the name of search, actually started the beginning of today’s AI – search was the rare or even the only business that required a large number of algorithmic talents at that time, and AI talents completed the initial formation and division. The famous auction between Baidu and Google for Geoffrey Hinton happened under this logic. Afterwards, in the early key application of AI – machine translation, Baidu and Google also had a catch-up competition. Baidu once released machine translation papers in succession, demonstrating the performance of Google’s internal equivalent or even surpassing each other. The anxiety within Google was finally relieved by the timely application of TPU, and Baidu released a powerful translation service before Baidu. After that, the two companies also deployed autonomous driving at the same time. In the field of autonomous driving, Google’s Waymo and Baidu’s Apollo (Radish Run) were almost the earliest and most determined players in the world, and today they have finally become the only two platforms to operate on a large scale. As Sundar Pichai has emphasized many times before, even in the period when the autonomous driving industry is facing many doubts and is generally looked down on by the outside world, Google still chooses to continue to increase its investment in Waymo. Baidu has also invested in Apollo and Radish Run for more than a decade, and has experienced the complete cycle of the industry from fanaticism to calm to gradual commercialization. Both sides deeply understand that autonomous driving is the crown jewel of AI technology. Its complexity and extreme requirements for safety determine that this must be a “hardcore game” that requires long-term and continuous investment. Whether it is Waymo’s continued investment in sensor fusion, simulation test platform, or Baidu Apollo’s exploration of vehicle-to-road collaboration and L4-level autonomous driving, they all reflect the same technical strategic thinking: that is, to carry out a full-stack layout in key technology areas, and to overcome technical difficulties one by one with great patience and resources, and ultimately promote the maturity and commercialization of the technology. These intriguing similarities in core technology genes and strategic cognition of AI make people look forward to Baidu again today. Picture Baidu and Google started from a similar starting point (search), and also experienced the process of transforming algorithms into large-scale applications. Moreover, Baidu is not just a follower. In the long development of AI, Baidu has almost never missed a key node: from the early investment in deep learning, to the research and development of voice and image technology, to the ten-year cultivation of autonomous driving, and in recent years, the all-out commitment to the big model. Baidu has always been sensitive to the evolution of AI technology priorities, and continues to apply the algorithm understanding and engineering experience from businesses such as search to real products. This accumulation, and the unique knowledge formed in processing massive amounts of data and understanding user intentions, is also the opportunity for Baidu in the second half of the big model today. “I think innovation cannot be planned, you don’t know when innovation will come, all you can do is create an environment conducive to innovation.” Robin Li, Baidu’s founder and CEO, who has experienced the whole wave of AI firsthand, once expressed his views on today’s technological anxiety. This also reveals Baidu’s strategic patience in the field of AI to some extent. He stressed, “When the technology is developing so fast, you must continue to invest to ensure that you are at the forefront of technological innovation. We still need to make continuous investments in chips, data centers and cloud infrastructure to train better and smarter next-generation models.” Baidu’s “four-layer AI architecture” includes the cloud infrastructure layer with Wanka cluster, the Flying Paddle open-source framework layer widely used by Chinese developers, the iterative Wenxin large model layer, and the application layer such as Baidu Search and Baidu Library for AI refactoring. Recently, Baidu announced at the Create conference in April 2025 that it had lit up the first fully self-developed 30,000-card cluster in China, which provided a solid computing power base for its large model training and inference. In the process of transforming AI technology into practical applications and business value, the cloud platform plays a crucial role. For Baidu, Baidu Intelligent Cloud is not only a window for its AI technology output, but also one of the core engines of its AI commercialization strategy. According to Baidu’s 2025 Quarter 1 financial report, cloud revenue grew by as much as 42% year-on-year. Driven by AI, the AI contribution revenue related to cloud business has reached triple-digit growth, and the operating profit margin has exceeded 10%. According to statistics, 65% of central state-owned enterprises are currently using Baidu Intelligent Cloud. At the model level, the Wenxin model is iterating rapidly. From the Wenxin 4.5 Turbo to the Deep Thinking Model X1 Turbo, Baidu not only emphasizes multi-modal processing and powerful logical reasoning capabilities, but also continues to optimize costs. Compared with the Wenxin 4.5 Turbo, the Wenxin model 4.5 Turbo is faster and the price is reduced by 80%. Compared with the Wenxin X1, the Wenxin model X1 Turbo has improved performance and reduced the price by 50%. The average daily usage of Wenxin model exceeds 1.65 billion, and the user scale of ERNIE Bot has reached 430 million. “We live in a very exciting era. According to Moore’s Law, every 18 months, the performance doubles and the price halves. But today, when we talk about large language models, the cost of inference can basically be reduced by more than 90% in 12 months. “Robin Li said at the recent Baidu earnings conference.” Not only in the field of AI or the IT industry, but looking back over the past few hundred years, most innovations are related to cost reduction. If the cost is reduced in a certain proportion, the productivity will also increase by the same proportion, which is the essence of innovation. Today, the speed of innovation is much faster than before. “This is also Baidu’s thinking on ecological construction. Baidu Intelligent Cloud’s Qianfan Big Model platform plays a key role. The platform has been connected to hundreds of mainstream big models at home and abroad, providing developers with a rich selection of models and highly competitive prices. At present, Qianfan has helped customers fine-tune 33,000 models and develop 770,000 enterprise applications. It is also the first cloud vendor in China to be compatible with MCP. It hopes to promote the sharing and call of AI capabilities through standardized interfaces. In Quarter 1, 2025, Baidu Intelligent Cloud won the bidding market with 19 win the bidding projects and 450 million yuan, ranking first in the bidding market of general large model manufacturers. Moreover, in some specific AI product concepts and implementation rhythms, Baidu began to show a more sensitive sense of smell. Taking the agent as an example, Robin Li regarded it as “the hottest track for AI applications”. Baidu has launched a general-purpose “super agent” product “Heart Sound” App, as well as a no-code generation application development platform “Seconda”, which aims to lower the threshold for the development and use of AI applications. A detail worth playing with is that in the design of the “Wen Xiaoyan” product, Baidu has automatically selected the most suitable model according to user requests to handle specific tasks. According to previous exchanges between Silicon Star people and the Google Gemini team, Gemini has initially realized the automatic call of different capabilities of the same model according to user requests, and the longer-term goal is also the automatic selection across different models. In part, this reflects Baidu’s advanced thinking on the combination of user experience and technology in specific application scenarios, and even a bit of the original machine translation of you competing with me. In the field of autonomous driving, “Radish Run” has begun to expand to the international market. Baidu announced a strategic cooperation with Dubai Road Transport Authority (RTA) and Autogo of the United Arab Emirates to provide driverless travel services in Dubai and Abu Dhabi. The price of its sixth-generation driverless car is only one-seventh of Google’s Waymo model. In addition, Apollo ADFM, the world’s first L4-level end-to-end autonomous driving model released by Baidu Apollo, is also beginning to explore the next technology node. Today, AI is gradually moving from a single model to a deeper ecological construction and value implementation. Although the simple chatbot form has attracted a lot of attention in the early stage, it has shown limitations in terms of user retention and business model sustainability. It lacks a model of strong application scenarios and distribution channels, and its “stamina” may be insufficient. In contrast, those companies that are committed to building “full-stack services” show stronger resilience and development potential. From the comprehensive AI reconstruction application “family bucket” displayed by Google I/O, to the simultaneous flowering of Baidu’s four-layer structure – especially on the application side, the old tree is blooming with new flowers of AI, Baidu Library AI MAU users reach 97 million, Baidu Netdisk MAU users exceed 200 million, AI MAU users exceed 80 million, and the average daily storage file exceeds 1 billion. Looking at today’s competition in the long history of AI technology development, we will find that it has always been a process of one after another, alternately leading. The final test is endurance and foresight, as well as the accumulation of various “attachments” along the way. Baidu and Google have always been “paranoid” about technology, which was not sufficiently recognized by the outside world during the period when ChatGPT brought FOMO. Today, when everyone realizes the importance of long-term technology accumulation, full-stack strategic layout, and long-term adherence to core business, the potential of companies that have recently accumulated the deepest and deepest understanding of this technology has finally been rediscovered.
发表回复