In 2025, the wave of AI continues to sweep across the globe, with the competition among tech giants evolving from a focus on individual model capabilities to a comprehensive stack competition encompassing chips, computing power, models, applications, and ecosystems. As a key player in the industry, ByteDance is accelerating its AI strategy, building a full-stack system from underlying hardware to upper-level applications, and actively creating an open ecosystem driven by both C-end super app traffic and B-end enterprise services. This article will delve into ByteDance’s full-stack AI strategy and explore its impact on China’s AI industry.
Full-Stack Layout, Computing Power First
ByteDance has been increasing its investment in AI infrastructure. According to a report by Huachuang Securities, ByteDance is actively constructing computing power centers both domestically and internationally, significantly enhancing cluster performance based on self-developed DPU GPU instances. Previous reports indicated that ByteDance plans to invest over $12 billion in AI infrastructure by 2025. This substantial capital expenditure is primarily aimed at building its own computing power centers and developing DPU chips, which will undoubtedly provide strong computing support for ByteDance’s AI model training and inference, as well as lay the foundation for its autonomy and control in the AI chipsector.
Multimodal Models and Application Ecosystem
ByteDance continues to innovate in model architecture, with its latest open-source Seed-OSS-36B model adopting the Apache-2.0 license and supporting a native context length of 512K, while introducing an innovative ‘controllable thinking budget’ mechanism to enhance inference efficiency. Multimodal technology has recently been a focus for ByteDance. The Waver1.0 architecture supports generating from text to video, images to video, and text to images, achieving seamless switching in multimodal generation and reconstructing the content creation process. Meanwhile, OmniHuman-1.5 brings characters to life with just a single photo and an audio clip through the concept of ‘comprehensive conditional training.’ From the product matrix perspective, ByteDance’s AI product system is led by Doubao, covering multiple scenarios. The Doubao family now includes over ten segmented models, such as General Pro/Lite, role-playing, speech synthesis/recognition, text-to-image, and video generation. The video generation product line has shown particularly impressive performance, with Seedance1.0Pro leading globally in both text-to-video and image-to-video categories according to ArtificialAnalysis. As of Spring 2025, data released by QuestMobile shows that Doubao users have exceeded 110 million, a year-on-year increase of 864.35%.
Enterprise Services and Industry Penetration
In the enterprise market, HiAgent2.0 and the Doubao Enterprise Edition are driving growth. HiAgent2.0 employs a ‘dispatch dialogue action’ triadic architecture, supporting three task orchestration methods: flowchart/natural language/API, with an inbuilt library of over 100 industry templates. ByteDance has also launched AI hardware products, such as AI headphones OlaFriend and other AIoT products. As of June this year, the shipment of AIoT products connected to Doubao has exceeded 1 million units, with expectations to surpass 10 million units by the end of 2025. The hardware products complement the software ecosystem, aiming to create a more complete AI experience. The Doubao large model has served 9 of the top 10 global smartphone manufacturers, 80% of mainstream automotive brands, 70% of systemically important banks, and over half of the 985 universities. As of the end of May 2025, the daily average token usage of the Doubao large model exceeded 16.4 trillion, a 137-fold increase from its release in May of last year. According to an IDC report, in 2024, Volcano Engine ranked first in China for public cloud large model service calls, with a market share of 46.4%. ByteDance is building its differentiated advantages, with Volcano Enginereducing costs through scaling, providing enterprises with cost-effective multi-cloud services. At the same time, Volcano Engine is actively promoting ecosystem development, collaborating with leading companies, and incubating AI-native enterprise service startups.
Future Outlook and Industry Reflections
ByteDance’s AI development path shows several clear trends: technological integration will deepen, with the combination of multimodal technology and VR/AR technologies becoming a new growth point; the application ecosystem will become more open, with Volcano Engine creating a ‘model supermarket’ to gather third-party large models and build a broader developer ecosystem; and human-computer interaction methods will undergo transformation, with ByteDance exploring new interactive devices. In this era of technological transformation brought about by large AI models, ByteDance is striving to evolve from a ‘technology company’ to an ‘innovative technology company.’
In the context of the accelerated implementation of AI-native applications, do you think ByteDance’s full-stack layout can help China’s AI industry achieve leapfrog development?
返回搜狐，查看更多
平台声明：该文观点仅代表作者本人，搜狐号系信息发布平台，搜狐仅提供信息存储空间服务。