文 | Tech商业
今年夏天，马化腾在内部将腾讯的人工智能项目形容为“一艘漏水的船”。这句罕见的严厉批评，叠加同期资本市场对腾讯AI前景的反馈，以及上半年高达847.2亿元人民币的资本开支带来的财务压力，让这家社交巨头在AI赛道上的处境显得颇为微妙。
在C端，面对字节跳动旗下“豆包”在用户规模上的压制，腾讯推出的AI应用“元宝”在增长上面临着巨大的挑战。
然而，据彭博社近期报道，腾讯云与智慧产业事业群（CSIG）CEO汤道生提供了一套不同的解法：将重心向B端转移，并将内部代码工具改造为名为WorkBuddy的AI助手。数据显示，该产品月活跃用户已达到3235.3万。
从C端的扩张受阻，到B端工具的规模化起量，腾讯的AI战略正在经历一次从“追求流量入口”向“深耕企业场景”的务实转向。这不仅是对技术路线的调整，更是对其庞大商业生态如何与AI结合的重新审视。
C端的流量焦虑与算力成本
要理解腾讯AI的战略收缩，必须客观审视C端战场的投入产出比。
2025年到2026年初，国内大模型C端应用经历了一场流量圈地战。字节的豆包依托抖音的超级流量池，通过大规模投流和极低的使用门槛，迅速占据了市场领先地位。截至2026年第一季度，豆包的月活跃用户数达到3.45亿，日活稳定在1.4亿左右。而同期，腾讯元宝的月活仅为5735万，约为豆包的六分之一。腾讯虽然也推出了元宝，并在微信生态内给予了入口扶持，但在AI对话场景中，社交关系链的导流优势并未立刻转化为绝对的用户粘性。
与此同时，大模型是一项极度消耗算力的重资产游戏。2026年上半年，腾讯的资本开支达到847.2亿元人民币，其中第二季度单季资本开支高达527.84亿元，同比暴增176%。受AI业务驱动，当季服务器费用同比实现翻倍，AI相关成本的净影响从第一季度的约88亿元扩大至第二季度的105亿元。这种高强度的投入直接导致了自由现金流的恶化，第二季度自由现金流录得-138亿元，罕见地转为负值。
马化腾的“漏水船”之喻，正是对这种高投入、低转化状态的警示。在缺乏明确变现路径的情况下，盲目扩大C端规模只会加速消耗。腾讯必须寻找新的突破口，将AI能力转化为实实在在的企业收入。
WorkBuddy：向B端场景下沉 面对C端的阻力，腾讯选择了在B端寻找突破口。
WorkBuddy并非凭空诞生的新产品，而是脱胎于腾讯内部工程师用来提升写代码效率的辅助工具。其核心逻辑并非提供通用的聊天对话，而是嵌入具体的工作流，做企业的“AI打工人”。
据报道，WorkBuddy结合了腾讯现有的即时通讯（企业微信）和办公工具（腾讯文档），并接入了能够处理数据、生成报告的AI能力，类似于Claude Cowork的定位。这种做法的出发点在于：企业客户需要的不是能写诗的AI，而是能切实提升业务效率、且数据安全的工具。
根据QuestMobile发布的2026年9月数据，WorkBuddy以合计月活3235.3万稳居办公类Agent榜单首位，季度环比暴涨647.1%。其月活规模约为第二名百度搭子（977万）的3.3倍，第三名阿里千问办公（790.5万）的4倍。
这一数据很大程度上得益于腾讯在企业服务市场的长期积累。当企业客户已经习惯了使用企业微信和腾讯文档时，WorkBuddy作为一项附加的AI能力，能够以较低的迁移成本接入。此外，腾讯在B端市场提供的私有化部署能力，也契合了大型国央企和金融机构对数据安全的刚性需求。在数据不出内网的前提下，WorkBuddy能够通过定制化部署，解决企业在合规层面的顾虑。
B端战场的真实阻力
尽管WorkBuddy在数据上取得了阶段性进展，但B端AI市场的竞争远未到分出胜负的时刻。
首先，是竞品的快速追赶。 阿里和字节在B端同样有着深度的布局。阿里的钉钉和通义大模型在企业协同上具备极强的竞争力，字节的飞书也通过扣子（Coze）等平台在企业级Agent上发力。值得注意的是，虽然WorkBuddy目前领先，但百度搭子以1567.5%的季度环比增速成为榜单中增长最快的产品，千问办公和豆包工作也处于早期用户积累阶段。WorkBuddy面临的并非空白市场，而是竞争激烈的存量博弈。
其次，是B端AI商业化的“定制化悖论”。 大型企业客户往往需要高度定制化的私有化部署，这导致项目交付成本极高，很难像SaaS软件那样实现低边际成本的规模化复制。每一个大客户都需要单独的模型微调、数据清洗和系统对接，这使得WorkBuddy目前的高月活，究竟能转化为多少真实的B端付费订阅或项目流水，仍是一个未知数。在整体企业IT预算收紧的宏观背景下，客户对AI的ROI要求极其苛刻。
最后，是技术底座的考验。 腾讯在10月29日的全球数字生态大会上宣布，混元Hy4大模型即将发布。相比此前的Hy4 preview，新版本在长程和真实生产力任务上实现了全面大幅提升，显著提升了思维链效率，并首次引入视觉理解能力。如果底层模型能力无法在代码生成、多模态理解和复杂逻辑推理上形成足够的竞争力，WorkBuddy很难在产品体验上建立绝对的壁垒。模型能力是B端Agent的发动机，发动机不够强劲，上层应用再花哨也难以留住客户。
腾讯AI的长期主义与内部整合
打持久战，是腾讯极其擅长的领域。在移动互联网时代的千团大战、短视频战役中，腾讯都曾依靠强大的现金流和生态韧性熬过寒冬。
但在大模型时代，B端市场的培育周期比以往任何时候都长。企业对于大模型的应用，正从最初的新鲜感走向务实的ROI考量。腾讯将AI资源倾斜向B端，意味着要接受一个相对缓慢的增长曲线。在这个过程中，如何平衡算力投入与商业收入的差距，如何避免内部不同业务线在AI工具上的重复造轮子，都是管理层需要解决的难题。
过去一年，腾讯内部已经开始了频繁的架构调整。从混元大模型团队的整合，到将AI能力下放至各个业务线，腾讯试图在保持组织灵活性的同时，集中力量办大事。WorkBuddy的阶段性起量，某种程度上验证了这种内部资源整合的有效性。但面对阿里、字节的强力阻击，单一产品的突围还不足以宣告胜利。
结语
马化腾的“漏水船”警告，客观上推动了腾讯AI业务的战略调整。
从C端的高举高打，到向B端场景的务实下沉，腾讯正在将AI战略与自身的产业互联网基因深度绑定。这条路径或许缺乏C端爆款的爆发力，但更符合腾讯当前的技术结构与客户基础。
在未来的竞争中，腾讯AI的考验不再是能否在日活数据上超越豆包，而是能否在企业级市场中，跑通一条投入产出比合理、且能够持续交付价值的商业闭环。这艘正在修补的船，需要在更为深水区的B端海域，接受真实的商业风浪考验。
本文来自微信公众号“Tech商业”，作者：Tech商业，经授权发布。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Tech Business
This summer, Maerten internally described Tencent's AI project as “a leak ship”. This rare and harsh criticism, combined with the feedback from the capital market on Tencent AI’s outlook over the same period, and the financial pressure on capital spending in the first half of the year, amounting to 84.720 billion yuan yuan, makes the social giant’s position on the AI track delicate.
On the C side, the AI application of Tencent faces enormous growth challenges in the face of the byte jumper's “Doubao” pressure on the size of its users.
However, according to recent reports by Bloomberg, Tencent Clouds and the CSIG CEO Tutson provide a different set of solutions: shift the focus to the B side and transform the internal code tool into an AI assistant named WorldBuddy. Data show that 32353,000 active users per month of the product.
From the blockage of expansion at the end of C to the scaling up of the B-end tool, Tencent's AI strategy is going through a pragmatic shift from “seeking entry” to “deep-farming enterprise scenes”. This is not only a re-routing of technology, but also a re-examination of how its vast commercial ecology is integrated with AI.
C-side flow anxiety and calculation costs
To understand the strategic contraction of Tencent AI, it is important to objectively look at the input-output ratio in the C-field.
From 2025 to early 2026, the domestic LLMC-end application experienced a traffic ring ground battle. Doubao of byte, supported by Douyin's super-flow pool, quickly gained market leadership through large-scale drops and extremely low use thresholds. By the first quarter of 2026, the number of Doubao monthly active users had reached 345 million, and daily life had stabilized at around 140 million. During the same period, the monthly life of Tencent yuan was only 5,735,000, about one sixth of Doubao. Although Tencent has also launched its beauties and has provided access within the WeChat ecology, in the AI dialogue scene, the channeling advantages of the social relationship chain have not immediately been transformed into absolute user stickiness.
At the same time, LLM is a heavy-asset game that is extremely deplete. In the first half of 2026, Tencent ' s capital expenditure amounted to RMB 84.72 billion, of which the second quarter ' s single-season capital expenditure amounted to RMB 52.784 billion, an increase of 176 per cent over the same period. Driven by AI operations, the cost of servers in the current season doubled over the same period, and the net impact of AI-related costs increased from approximately $8.8 billion in the first quarter to $10.5 billion in the second quarter. This high-intensity input has directly led to a deterioration of the free cash flow, which recorded -13.8 billion yuan in the second quarter and rarely turned negative.
The term “spill ship” in Maten is precisely a warning to this high input, low conversion state. In the absence of clear realization paths, blindly expanding the C-end size will only accelerate consumption. Tencent has to look for new breakthroughs to translate AI capacity into real business income.
WorkBuddy: Falling down to the B-side scene and facing the resistance of the C-side, Tencent chose to find a breakout at B-side.
WorkBuddy is not a new product born out of a vacuum, but rather a help tool that was used by internal engineers in Tencent to improve the efficiency of writing codes. The core logic is not to provide a universal chat, but rather to embed specific workflow as an “AI-TIP WORKER” for an enterprise.
WorkBuddy reportedly combines the existing Tencent instant messaging (enterprise WeChat) and office tools (Tencent documents) and has access to the AI capability to process data and generate reports, similar to Claude Cowork. The starting point for this approach is that what business clients need is not an AI that can write poems, but a tool that can actually improve business efficiency and data security.
According to September 2026 data from QuestMobilepublish, WorkBuddy was at the top of the Agent list, with a combined monthly income of 32,353,000. Its monthly activities are approximately 3.3 times the size of the second Baiduer (9.77 million) and four times the size of the third Ali ' s office (7.905 million).
This data has largely benefited from the long-term accumulation of Tencent in the business services market. When business customers are used to WeChat and Tencent documents, WorkBuddy is able to access them as an additional AI capability at lower migration costs. In addition, the privatization deployment capacity provided by Tencent in the B-end market also corresponds to the rigid demand for data security by large conglomerates and financial institutions. Without an Intranet of data, WorkBuddy is able to address business concerns at the compliance level through customized deployment.
The real resistance in the B-front.
Despite WorldBuddy ' s progress in terms of data, competition in the B-Ai market is far from the moment of success.
First of all, it's the race to catch up. Ali and byte have the same deep layout at B. Ali's nails and Qwen LLM are extremely competitive in terms of business synergies, and Z113Z's flying books work on Enterprise-level Agent through platforms such as Coze (Coze). It is worth noting that, while WorkBuddy is now in the lead, the Baidu hook is the fastest-growing product on the list at a quarterly round rate of 1565.5 per cent, and the work on office and Doubao is at an early user accumulation stage. WorkBuddy is not facing a blank market, but rather a competitive stock game.
Second, there is the “sorted paradox” of B-Ai commercialization. Large enterprise clients often require highly customized privatization deployments, which results in high project delivery costs and makes it difficult to scale up low marginal costs, as in the case of SaaS. Each large client needs a separate model fine-tuning, data cleansing and system interfacing, which makes it still unknown how much of the current high-moment work of WorldBuddy can be converted into real B-side paid subscriptions or project flow. In the macro-level context of a tight IT budget for the enterprise as a whole, clients are extremely demanding on AI ROI.
Finally, it is a test of technology. Tencent announced at the Global Digital Ecology Conference on 29 October that the hybrid Hy4LLM would be publish. Compared to the previous Hy4 preview, the new version achieved a significant overall increase in long-range and real productivity tasks, significantly improving the efficiency of the thinking chain and introducing visual understanding for the first time. WorkBuddy has difficulty in establishing absolute barriers to product experience if bottom model capabilities are not sufficiently competitive in code generation, multimodular understanding and complex logic. Model capability is a B-end Agent engine, which is not strong enough to keep customers in the upper layers.
Tencent AI long-termism and internal integration
Fighting for eternity is an area where Tencent is extremely good. Tencent had been able to survive the winter with strong cash flows and ecological resilience in the context of the mobile Internet era, the Millennium War and the short video campaign.
In the LLM era, however, the B-end market has a longer breeding cycle than ever before. The enterprise ' s application of LLM is moving from its original sense of freshness to pragmatic ROI considerations. Tencent tilts AI resources towards B, which means accepting a relatively slow growth curve. In this process, balancing the gap between calculus inputs and business income, and avoiding duplication of wheels on AI tools across different business lines in-house, are challenges that management needs to address.
Over the past year, frequent structural adjustments have been initiated within Tencent. From the integration of the mixed LLM team to the decentralization of AI capacity to the business lines, Tencent seeks to focus on major issues while maintaining organizational flexibility. WorkBuddy ' s milestones have to some extent validated the effectiveness of this internal resource consolidation. But in the face of the strong resistance of Ali, byte, the single-product breakout is not enough to declare victory.
Concluding remarks
The Mauten “spillboat” warning objectively facilitated the strategic readjustment of Tencent AI operations.
From C to B to practical sinking, Tencent is tying the AI strategy to its own industrial Internet gene depth. This path may lack the explosive power of the C-end explosion, but is more in line with the current technical structure and client base of Tencent.
In future competition, the test of Tencent AI is no longer whether it will be possible to go beyond Doubao in daily data, but rather whether it will be possible to run through a commercial closed loop in enterprise-level markets where input output is more reasonable and can deliver value on a sustainable basis. This ship, which is being repaired, needs to be tested by real commercial winds in the B-end area of the deeper waters.
This is from WeChatWeChat Official Account “Tech Business”, by Tech Business, authorized publish.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.