文 | Tech商业
在AI大模型进入商业化落地的深水区后，百度正在试图通过一场内部的组织架构大手术，来重新寻找自己的节奏。
10月28日，据多家媒体及内部人士消息，百度人工智能云事业群组（ACG）启动了新一轮的组织架构调整。这次调整的核心动作有两个：一是产研阵型收缩为“基础设施+智能体”两层结构；二是百度副总裁阮瑜轮岗，担任小度科技组织负责人。
与此同时，市场上传出“小度科技CEO李莹已离职，由阮瑜接任，一见团队并入小度”的传闻。对此，有知情人士向媒体明确回应，李莹离职传闻不属实；但“一见团队并入小度”这一说法尚未获得百度官方确认。截至发稿，百度官方并未就高管的最终去留作出进一步说明。
在经历了大模型的狂热与财报的压力后，百度的这次调整，折射出了一家老牌巨头在面对AI时代“起大早、赶晚集”时的急迫与务实。这不仅仅是一次简单的人事换防，更是百度在重新思考：在算力成本高企、To B落地艰难的当下，究竟该以何种姿态参与这场残酷的淘汰赛。
做减法：告别分散，走向“两层”
要看懂这次调整，首先要理解百度ACG过去面临的痛点：战线过长，资源分散。
在此次调整前，ACG下属的智能应用事业部虽然看起来枝繁叶茂，包含了众多的AI应用和行业解决方案。但在实际的商业运作中，这种庞杂往往意味着重复造轮子、跨部门协同困难等大公司通病。尤其是在大模型时代，算力成本极其高昂，如果前端应用不聚焦，很容易陷入“什么都做，但什么都不精”的窘境。
百度此次的动作非常坚决：按业务属性对智能应用事业部相关业务进行拆解和划转。官方表述为“基于各业务属性，分别转入基础设施事业部、智能体事业部、服务交付与质量工程中心”。结合此前已披露的业务归属，大致对应关系是： * 偏底层的划入基础设施事业部； * 偏AI原生应用的划入智能体事业部； * 偏行业交付的划入更名后的行业应用与技术服务中心。
这意味着，百度ACG的产研阵型被压缩成了“基础设施+智能体”的两层架构。
在To B的商业逻辑里，这属于典型的“结硬寨，打呆仗”。基础设施（算力、云服务）是底座，智能体（Agent）是中间层的大脑，行业应用则变成了最上层的服务。百度不再奢望每个部门都能独立跑通大模型闭环，而是要求底层把算力和模型做好，上层去把客户场景吃透，减少跨部门的内耗与摩擦。
这套打法的本质，是百度试图深化“云智一体”的战略。在过去，云业务和AI业务往往是两张皮，云卖算力，AI卖模型。如今，百度试图用统一的AI底座去支撑所有应用，从而在成本端实现极致的压缩，在交付端实现更高效的协同。
人事轮岗：阮瑜的C端任务与李莹的去留
相比于架构的调整，人事变动往往更容易挑动市场的神经。
据公开信息，此次轮岗至小度科技担任组织负责人的阮瑜，是一位不折不扣的“老百度”。她2013年加入百度，先后负责过百科、百家号等内容生态，曾任百度知识垂类总经理；2020年起执掌百度网盘，任内将网盘用户规模推破8亿；2024年8月起任应用产品事业部（后更名智能应用事业部）总经理，负责百度一见（视觉智能）、Hogee、胜算及交通、金融等行业应用。
而传闻中的当事人——小度科技CEO李莹，则是一位纯正的技术研发出身的高管。她2004年加入百度，做过搜索、知识图谱、地图，2020年升任集团首席信息官（CIO），2023年10月才接手小度科技。
一个偏产品和商业化运营，一个偏底层技术研发。
阮瑜轮岗至小度，且市场传出“一见团队并入小度”的消息——需要说明的是，这一并入说法目前尚未获得百度官方确认。但若该传闻属实，其信号将十分明确：百度希望小度科技不再仅仅是一家卖智能音箱和智能屏的硬件公司，而是要将大模型能力、视觉多模态能力（一见团队）与硬件深度结合，打造真正能在C端落地的AI原生硬件。
尽管李莹离职的传闻被知情人士否认，但在大厂的组织调整期，高管职务与权责的重新分配是必然的。李莹是否彻底离开，还是转任其他职务，或许在不久的将来会有明确的官方定论。
但不可否认的是，小度科技作为百度重要的AI硬件入口（2020年独立融资，估值约200亿元），正处于一个极其关键的转型期。面对字节、小米等新老对手在AI硬件上的包抄，以及端侧AI爆发的前夜，小度需要更懂市场和用户的人来重塑打法。若视觉团队（一见）真如传闻并入小度，意味着百度希望在智能硬件的“眼睛”上做足文章，让硬件从“只听会说”进化到“能看会认”。
押注FDE：大模型To B的“最后一公里”
在此次调整中，百度特别提到了一个词：FDE（前置部署工程师）机制。
这并非百度独创。在过去的一年里，无论是硅谷的Palantir，还是国内的腾讯、字节，都在布局FDE人才。所谓FDE，就是能够深入客户现场，将大模型的技术能力与客户复杂的业务流、老旧的数据系统缝合起来的人。
大模型卖不动API了，因为企业客户发现，拿到API自己也搞不定复杂的业务落地。企业真正需要的是懂AI、又懂自身业务逻辑的专家，来帮他们把大模型装进具体的场景里。
百度强调通过FDE机制推动AI能力在行业场景中落地，说明百度云已经放弃了早期那种“卖算力、卖Token”的浪漫主义，转向了重交付、重服务的苦活累活。在当前的宏观环境下，企业客户的IT预算极其紧张，他们不再为PPT上的AI愿景买单，而是要求看到切切实实的降本增效。FDE就是那把破冰的铲子。百度试图通过FDE，把标准化的云和智能体，定制化地塞进客户的业务里，打通大模型To B落地的“最后一公里”。
聚焦与焦虑：百度的背水一战
回看这次百度ACG的组织调整，我们可以清晰地看到一种情绪的转变：从大模型早期的狂热扩张，转向了极度务实的聚焦收缩。
在文心一言刚发布的2023年，百度是国内最早吃螃蟹的巨头。但两年多过去，战场已经发生了剧变。字节凭借豆包极其狂野的流量分发占领了C端心智，DeepSeek凭借极致的工程优化和开源策略横扫了开发者生态，腾讯则依托微信庞大的生态在B端和短剧等领域闷声发大财。
百度虽然在底层大模型和AI云基建上仍有深厚积累，但在应用层和用户时长争夺上，却始终面临巨大的压力。将阵型收缩为“基础设施+智能体”，把懂商业的阮瑜放到小度，把行业交付团队统一归集。这一切都是为了缩短决策链条，把有限的钱和资源，精准地打在能产生真实商业收入的靶心上。
对于这家立志要“用AI重塑一切”的巨头来说，这次重整阵型，既是一次战略调整，也是一场生死攸关的背水一战。因为在这场大模型的淘汰赛中，留给参赛者调整姿势的时间，真的不多了。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Tech Business
After AZLLM had entered the commercialized deep-water zone, Baidu was trying to re-establish its rhythm through an internal structural surgeries.
On 28 October, according to a number of media and insiders, the Baidu artificial cloud enterprise cluster (ACG) launched a new round of organizational restructuring. There are two core actions for this restructuring: the contraction of the production and research formation into two layers of infrastructure plus AI agent, and the rotation of Baidu Vice President Nguyen Yu, who is the head of small-scale technology organizations.
At the same time, there is a rumor in the market that “Small technology CEO Lee Li has left, was replaced by Nguyen Yu, and the team was incorporated into small size”. There has been a clear response from informed sources to the media, and rumours of Li's departure are not true; however, the claim that “the first-hand team was integrated into small groups” has not yet been officially confirmed by Baidu. As of the submission, the official Baidu did not provide further information on the final departure of the executive.
After the pressure of LLM's frenzy and financial reports, Baidu's adjustment reflected the urgency and pragmatism of an old giant in the face of the "earthings and evenings" of the AI era. This is not just a simple change of personnel, but a rethinking of Baidu: what kind of attitude should we take in this brutal phase-out, in the face of high-cost, hard-pressed To B?
Subtract: Disconnect and move towards two layers
It is important to understand this adjustment, first and foremost, the painful point that BaiduACG has faced in the past: the length of the front and the dispersion of resources.
Prior to this readjustment, the ACG's Smart Applications Department, although it appeared to have a proliferation of applications and industry solutions. However, in actual business operations, this multiplicity often means that large companies are ill, such as repeat wheels and cross-sectoral synergies. Especially in the era of LLM, the cost of computing is extremely high, and if front-end applications are not focused, they can easily be caught in the trap of “doing everything but doing nothing”.
Baidu was very determined in this exercise: the operations of the Department of Smart Applications were dismantled and transferred according to business attributes. The official expression is “on the basis of business attributes transferred to the Ministry of Infrastructure, the Department of AI agent Enterprise, the Centre for Service Delivery and Quality Engineering”. In combination with previously disclosed business attributions, the general correspondence is as follows: * Bottomline placement in the Infrastructure Department; * Enterprise AI agent for primary AI application; * Industry Applications and Technology Service Centre for renamed industry delivery preference.
This means that BaiduACG's production formation has been compressed into two layers of “Infrastructure+AI agent”.
In To B's business logic, this is typical of a “scrambling, war and war”. Infrastructure (calculations, cloud services) is the base, AI agent (Agent) is the brain in the middle, and industrial applications become the top level of service. Baidu no longer aspires to run independently through the LLM closed loops in each sector, but demands that the bottom level do the math and model, the upper level eats up the client scene and reduces inter-sectoral internal consumption and friction.
The essence of this set-up is that Baidu is trying to deepen the strategy of “uniqueness”. In the past, cloud operations and AI operations were often two skins, cloud sales, and AI sales models. Today, Baidu attempts to support all applications with a unified AI base, leading to extreme compression at the cost end and more efficient synergy at the delivery end.
Personnel rotation: the mission C of Nguyen Yu and Li Xian's stay
Changes in personnel tend to drive market nerves more easily than structural adjustments.
According to publicly available information, Nguyen Yu, who was the head of the organization on this rotation to Minster Tech, is a “old Baidu”. In 2013, she joined Baidu, which was responsible for encyclopedias, encyclopedias and other content ecologys. She was the General Manager of Baidu Knowledge Probes. Since 2020, she has been in charge of the Baidu Webboard, which is designed to reduce the number of Internet users by 800 million. Since August 2024, she has been the General Manager of the Ministry of Applied Products (subsequently renamed the Ministry of Smart Applications) and has been responsible for the applications of Baidu in the sectors of viewing intelligence, hogee, winning and transportation, finance, etc.
And the rumored client, CEO Lee Li, is a top executive of pure technology. In 2004, she joined Baidu, performed search, knowledge mapping, maps, in 2020 she was promoted to the Group ' s Chief Information Officer (CIO) and only took over small technology in October 2023.
A product-based and commercial-based operation, and a bottom-up technology.
Tuen Yoo's rotation to a small point and the market has news of the integration of the team into a small scale. It should be noted that this incorporation has not yet been officially confirmed by Baidu. But if the rumour is true, the signal will be clear: Baidu hopes that small-scale technology will no longer be just a hardware company that sells smart soundboxes and smart screens, but rather will combine LLM capabilities, visual multi-modular capabilities (see team) with hardware depths to build a truly original AI hardware that can land on C.
While the rumours of Li’s departure were denied by the informed sources, the redistribution of executive positions and responsibilities was necessary during the organizational restructuring of the plant. Whether Li has left completely or transferred to other positions may be the subject of clear official findings in the near future.
It is undeniable, however, that small-scale technology, which is an important entry point to Baidu ' s AI hardware (with an estimated $20 billion in 2020 for independent financing), is in a critical transition. In the face of new and old rivals such as byte, Mi, and the night before the end of AI, there was a need for a smaller group of people with a better understanding of the market and users to reshape the method. If the visual team (at first sight) is true and is incorporated into a small story, it means that Baidu wants to write enough on the "eye" of smart hardware, so that hardware can evolve from "hear to speak" to "see to recognize".
bet FDE: “Final kilometre” of LLMto B
In this adjustment, Baidu refers specifically to the FDE (pre-deployment engineer) mechanism.
This is not an original Baidu. Over the past year, both Palantir in Silicon Valley and Tencent and byte in the country have deployed FDE talent. The so-called FDE is a person who is able to reach out to the client site and sew the technical capabilities of LLM to the complex business stream of the client, the old data system.
LLM can't sell API because business clients find that they can't get to API themselves to land complex operations. What businesses really need is experts who know AI and who understand their business logic to help them put LLM into a specific scene.
The emphasis placed by Baidu on promoting the capacity of AI to land in industry through the FDE mechanism suggests that the Baidu cloud has abandoned its early romanticism of “calculating and selling Token” and has turned to heavy service delivery. In the current macro environment, the IT budget of corporate clients is extremely tight, and they no longer pay for the AI vision on the PPT, but rather demand tangible downside efficiency gains. FDE is the ice-breaking shovel. Baidu attempts to customize the standard cloud and AI agent through the FDE into the client's business and access the “last kilometre” of LLMTo B landing.
Focus and anxiety: Baidu's fallback
Looking back at this organizational restructuring of BaiduACG, we can clearly see an emotional shift: from an early feverish expansion of LLM to an extremely pragmatic contraction of focus.
In 2023, Ernie was the first crab eater in the country. But more than two years later, the battlefield has changed dramatically. byte has taken over C's mind with the extremely wild flow of Doubao, DeepSeek has swept the developer's ecology with extreme engineering optimization and open-source tactics, and Tencent has made a huge fortune in areas such as B and short dramas thanks to the vast ecology of WeChat.
Although Baidu still accumulates heavily on the ground floor of LLM and AI cloud infrastructure, it continues to face enormous pressure on applications and user competitions. The formation has been reduced to “Infrastructure + AI agent” and the commercial yoga has been reduced to a minimum, bringing together industry delivery teams. All of this is aimed at reducing the decision-making chain and precisely targeting the limited money and resources that generate real business revenues.
For this giant who aspires to “restore everything with AI”, this reorganization is a strategic readjustment and a life-and-death war. Because in this phase-out of LLM, there is really not much time left for the contestants to adjust their positions.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.