After exiting the "horse racing" trial-and-error phase, big tech companies' AI development enters a unified integration era
In the past half month, major tech giants have been taking the same action in their AI businesses: unifying entry points and converging capabilities.
For example, WeChat's native AI assistant "Xiaowei" has launched a small-scale gray test, while Alipay has launched its AI assistant "Abao". These two national-level super apps are promoting the implementation of AI technology through integration. ByteDance's Doubao has officially launched paid features to achieve centralized operation of C-end AI products through market stratification. Baidu has integrated the entire Wenxin product line to create a one-stop AI service portal. In addition, Alibaba has consolidated scattered AI tools including QoderWork, Wukong, and MuleRun, preparing to launch a unified productivity AI platform.
As the wild growth period of the "Hundred-Models War" comes to an end, the once widespread and separately competing AI product matrices are being actively consolidated, integrated, and unified by major tech companies. With enterprises making concerted efforts to reduce costs and increase efficiency, and regulatory authorities introducing new policies to strengthen top-level design, a clear main line of AI "great unification" featuring "internal enterprise integration and unified industry standards" has emerged, leading the current AI industry into a new era of transformation.
Uniting Forces: The Integration Wave of Major Tech Giants' AI Businesses
In the past six months, adjustments made by major tech companies in their AI businesses have shown striking consistency—they are no longer keen on launching independent AI super apps, but instead deeply embed AI into existing national-level applications and integrate internal resources to build a unified service entry point.
On the mobile end, the two super platforms WeChat and Alipay have coincidentally chosen to conduct internal tests on their native AI assistants "Xiaowei" and "Abao". This strategy does not require users to download an unfamiliar independent app, nor is it a simple superposition of functions. Instead, it deeply embeds AI capabilities as underlying services into daily life scenarios, allowing ordinary users to complete the transformation from perceiving and contacting AI to actually using it.
In terms of specific capabilities, "Xiaowei" has covered multiple daily usage scenarios, enabling users to directly adjust WeChat settings, send messages, make calls, and call mini-programs for life services through voice or text instructions. "Abao" covers scenarios such as home management, transportation, preferential shopping, government affairs and certificates, wallet bills, and companion interaction. Users only need to give a verbal instruction to complete tasks including inquiries, handling procedures, coupon claiming, shopping, and planning.
On the web and tool ends, integration efforts are even more obvious. Baidu is attempting to unify scattered AI-native applications such as Wenxin Yiyan, Wenxiaoyan, AI Search, Baidu Netdisk, Maps, and Wenku to create a one-stop AI service entry point that covers multiple scenarios including learning, office work, life, and entertainment. Alibaba has revamped its internal productivity tools, integrating and reshaping products such as QoderWork, Wukong, and MuleRun. The new unified productivity AI product will most likely adopt a three-layer architecture of "desktop + cloud + organization" to concentrate resources and consolidate its AI-to-B strategy.
Doubao is no longer satisfied with only conversation and generation capabilities, and has started to add practical service functions. Its most differentiated move is launching a paid professional version for productivity scenarios. The latest news shows that Doubao has started to integrate with Feishu, with its accounts able to be associated with Feishu's web version and enterprise version protocols.
Even DeepSeek, which previously focused on the model layer, has started large-scale recruitment after completing a 50 billion yuan financing round, with positions covering technology, product, and operation. This means it does not want to stay limited to model capabilities alone, but also intends to supplement application, product, and scenario capabilities, moving closer to becoming a more comprehensive AI company.
Leveraging the advantage of "being able to get things done", AI agents have become a key focus for major tech companies in their AI applications over the past two years. However, the current AI agent industry faces a realistic dilemma: countless agents operate independently, without a universal neural pathway to enable them to collaborate with each other. The underlying reason is that each company's self-developed interfaces and communication protocols are incompatible, naturally forming closed "agent islands" with no unified identity verification or capability interoperability mechanisms for cross-domain interactions.
In this context, the Ministry of Industry and Information Technology of China recently officially released the "Artificial Intelligence — Agent Interconnection" standard. As the first systematic guiding technical document for agent interconnection in China, this means that in the future, agents from different vendors, with different architectures, and under different deployment forms will no longer need customized one-to-one adaptation interfaces. As long as they follow the same set of standards, they can achieve seamless cross-platform and cross-system collaboration. For the industry, the development cost of multi-agent integration will be greatly reduced, and project cycles will be significantly shortened. Complex collaboration solutions that could not be implemented before due to excessively high adaptation costs will now have the commercial foundation for large-scale promotion for the first time.
To sum up, whether at the enterprise level or the policy level, the AI industry is gradually bidding farewell to the era of "scattered development and independent operations" and entering a new stage of intensive development characterized by "unified entry points, unified underlying infrastructure, unified specifications, and unified commercialization".
Trial and Error and Elimination: From the Hundred-Models War to the Survival of the Fittest
To understand today's "great unification", we must look back at yesterday's "Hundred-Models War".
In early 2023, the explosive popularity of ChatGPT ignited China's "Hundred-Models War". According to incomplete statistics, by the end of 2024, more than 200 large models had been publicly released in China, with a new model launched on average every 1.5 days. During this competition, major tech companies generally adopted a "horse-racing mechanism" — encouraging various business departments to independently explore AI products, with the winner being whoever achieved breakthroughs first.
In recent years, almost every business department of major tech companies has launched its own AI applications. For example, Alibaba's large model applications include Tongyi Qianwen, Tongyi Wanxiang, Lingguang, Tongyi Tingwu, Huiwa, EMO, Lingguang, and Quark AI Browser. Baidu's AI applications are even more widespread, including Wenxin Yiyan, Wenxin Yige, Wenxiaoyan, AI Search, the AI version of Baidu Netdisk, Map AI Assistant, and AI features for Wenku. Tencent has M2UGen, AnimateZero, Yuanbao, ima, Hy Translation, and Miaosi. ByteDance has restarted its "App Factory" model, launching more than 20 AI products covering categories such as chat, social interaction, office work, education, image, video, and music. Its domestic products include Boximator, Dreamina, Doubao, Maoxiang, Xinghui, Xiaoyunque, and BuboGPT, while overseas counterparts include Cici, BagelBel, and PicPic.
During this three-year trial-and-error period of the "Hundred-Models War", the AI applications of major tech companies pursued "full coverage without omission". Every department wanted to prove its AI capabilities, and every team wanted to create the next hit product. This almost inevitably led to resource competition, redundant R&D, and scattered development rhythms, ultimately forming a massive, uneven, and homogenized matrix of AI products that fell into the awkward competition of "reinventing the wheel".
For example, Baidu's Wenxin Yiyan, as one of the first batch of domestic large model products, once occupied the public spotlight and was known as the "Baidu version of ChatGPT". However, subsequent features such as Wenxiaoyan, AI Search, and Baidu Netdisk AI received lukewarm responses due to vague positioning, numerous names, and poor user experience. Baidu Netdisk AI attempted to combine file management with AI, but users found that it could neither efficiently organize files nor provide valuable intelligent recommendations. Map AI, due to the particularity of navigation scenarios, even increased operational complexity through AI interactions. Baidu essentially reworked almost all its business lines, attempting to reconstruct search, Wenku, Maps, and even Netdisk with AI. However, the common problem with these products is that they pursued AI for AI's sake, ignoring the real needs of users.
Alibaba's situation is similar. Tongyi Qianwen, as the base model, should have been the core of the AI ecosystem, but Alibaba later incubated multiple products simultaneously including Lingguang (for designers), Quark AI Browser (for young users), and DingTalk AI (for enterprise office work). Although these products have different focuses, they are highly homogenized in underlying capabilities, leading to redundant investment in R&D resources. More seriously, due to the lack of unified planning, data cannot be interconnected between different products, resulting in fragmented user experiences.
Objectively speaking, the essence of the "Hundred-Models War" was a trial-and-error behavior by major tech companies in an uncertain environment. When technical routes were unclear, multi-point deployment could reduce risks. However, as technology maturity increased and the market landscape took shape, scattered resources only dragged down competitiveness. After nearly three years of wild growth and market testing, the number of real players has sharply decreased. Many of the previously piled-up AI products have either been quietly removed from the market due to poor data performance or vague positioning, or have been incorporated into larger parent platforms for resource reuse.
The current AI industry has gradually transformed from the "Hundred-Models War" a few years ago into a market landscape of "giant ecosystems leading the way and emerging technical elites breaking through". Each enterprise only retains 1-2 core unified entry points to carry all AI capabilities for both the C-end and B-end. This is itself a result of competition: the Hundred-Models War focused on quantity, while the era of great unification emphasizes concentration. Widespread trial and error is an unavoidable stage, and centralization and unification are the inevitable destination of the industry.
The Threefold Logic Behind "Great Unification"
Major tech companies actively consolidating scattered AI businesses and the Chinese government introducing unified industry standards are not short-term strategic follow-ups or simple organizational restructuring. Instead, they are the result of multiple overlapping factors. We can understand the reasons behind this "great unification" from the following three aspects.
First, resource intensification and cost control under the background of reducing costs and increasing efficiency, bidding farewell to the era of burning money. AI is an extremely capital-intensive business. When various departments within an enterprise develop their own large models separately, computing power resources are fragmented and wasted, talent teams are redundantly established, and market promotion efforts conflict with each other.
From the perspective of computing power, each independent AI product requires a separate deployment of inference clusters. The parallel operation of multiple product lines means multiple sets of computing power running idle 24 hours a day, leading to a multiplicative superposition of inference and training costs. Tongyi, Wenxin, and Doubao are all large models with hundreds of billions of parameters, resulting in high costs per call. Scattered layouts directly push up enterprises' monthly computing power expenses. From the perspective of organizational manpower, independent AI teams on different business lines have a large amount of redundant R&D: basic functions such as conversation capabilities, image generation, document processing, and code tools are repeatedly developed by dozens of teams, leading to inconsistent product standards and extremely high difficulty in cross-departmental collaboration. From the perspective of market promotion, multiple independent AI products require separate traffic investment, operation activities, and user acquisition, resulting in scattered brand awareness and making it difficult for users to form stable perceptions.
This model of independent operations was tolerable in the early and rising stages of the industry, but under the general background of reducing costs and increasing efficiency, enterprises must concentrate resources to accomplish major tasks. Mergers and upgrades are implemented to invest limited computing power and talents into the most promising core products. Capability integration and collaboration have become the only rational choice. For example, the C-end fully takes over daily trivial matters, the B-end realizes cross-functional multi-Agent collaboration (such as the division of labor and collaboration between customer service and supply chain), and vertical industries (finance, government affairs, manufacturing) will form customized solutions. After unified integration, enterprises only need to operate one set of products and one membership system according to market demands, greatly reducing costs in marketing, customer service, and operation and maintenance.
Second, improving product experience and realizing a closed commercialization loop. The development of AI has now gone beyond just competition in technology and products, and is more about ecosystem and market recognition, especially how to achieve a closed commercialization loop. A scattered product matrix often means fragmented user experiences where membership rights are not interoperable. A unified entry point, on the other hand, can amplify AI capabilities through synergy effects and build a complete payment path.
Taking Doubao's paid service as an example, before integration, Doubao, as one of ByteDance's AI products, may have faced problems of overlapping or conflicting functions with other internal products, requiring users to switch between different products during use, resulting in fragmented experiences. After integrating all AI capabilities, Doubao has built a unified system of free basic service plus three levels of paid subscriptions: basic conversations are permanently free, while high-level capabilities such as office Agents, long contexts, and professional image generation are subscribed in a layered manner. A single product serves three types of paying users: light users, professionals, and enterprise clients, fully opening up the user conversion path. For Baidu Wenxin, its unified portal adopts an inclusive approach to expand the base of free users through a unified entry point, and realizes B-end monetization relying on its enterprise Qianfan platform. Both paths rely on a unified product matrix to complete the commercial closed loop.
Third, the inherent requirement of compliance supervision guides the industry toward standardized development. With the release of the country's top-level design for artificial intelligence, "compliance and unification" have also become mandatory requirements, which determine the future survival of products, and a large number of AI applications need to adapt to changes in a timely manner.
For example, the release of the series of national standard guiding technical documents "Artificial Intelligence — Agent Interconnection" mainly solves three major pain points: "agent islands" between vendors, security supervision of digital identities, and cost reduction to realize the large-scale implementation of agents.
The Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interactive Services, which will be officially implemented on July 15, put forward strict requirements for filing, review, and minor protection for AI anthropomorphic companions, custom roles, and virtual personas. Recently, Doubao and Tongyi have uniformly removed scattered agent functions, consolidated relevant demands into compliance-controlled products, and greatly reduced content compliance risks, which is a pre-emptive response to the implementation of this regulatory policy. Just imagine, if a vendor wants to maintain dozens of scattered AI products, each of which needs to independently complete security assessment and filing, the manpower and review costs for compliance may increase exponentially.
Conclusion
From the "Hundred-Models War" to "great unification", China's AI industry is undergoing a profound reshuffle. The future industry landscape will present a three-layer structure of "a small number of general-purpose large model bases + unified integrated entry points + customized solutions for vertical industries". Head tech giants will guard the unified C-end and B-end entry points, while small and medium-sized vendors will focus on the implementation of solutions in segmented industries. The disorderly competition of the past, where everyone was building large models and independent AI apps were everywhere, will no longer exist, and the industrial division of labor will become clearer.
Enterprises integrating their product matrices inward and the country unifying industry technical rules outward resonate with each other. This is not only a natural choice of market competition, but also an inevitable result of the law of industrial development.
This article is from the WeChat official account "Business Research Institute", written by Shang Jing, and published with authorization from 36Kr.