| 📒 | - miHoYo announces AI investment of up to 100 billion yuan (approx. 22.2 trillion KRW) over three years. - Official adoption of a full-stack in-house R&D strategy, including plans for efficiency and technical implementation. - Liu Wei: "If it fails, it's just like setting off a big firework"—potential for success noted through AI investment and proprietary model registration. |
Liu Wei, co-founder of miHoYo (global brand name HoYoverse), held a private technical sharing and recruitment session in Beijing on May 15. Chinese media outlets, including Yicai, reported that Liu announced an AI investment of up to 100 billion yuan (approximately 22.2 trillion KRW) over the next three years, alongside a strategy for full-stack in-house research and development.
During the presentation, Liu emphasized the principle that "the founder must be the technical lead," explaining that if decision-makers rely solely on reports, they cannot keep up with the rapid pace of paradigm shifts. He indicated that he would remain deeply involved in the field. Under his direct leadership, the organizational structure is also being reshaped. He emphasized a flat organizational structure focused on young, like-minded individuals rather than senior personnel who have already established their own domains.
In particular, the full-stack AI R&D signals a grand plan that goes beyond miHoYo simply fine-tuning external models. The company intends to develop everything in-house, from infrastructure to AI application architecture. During the presentation, specific plans regarding R&D efficiency and technical implementation were shared.
miHoYo highlighted the principle of "AI for AI, Model with Model." The core task is to build an in-house tuning pipeline where AI autonomously analyzes learning bottlenecks and writes GPU kernel code. They emphasized the importance of collaborative design for communication, computation, and data within a 10,000-GPU cluster. Furthermore, they defined "meaningful AI" as agents that not only understand context—such as codebases or conversation logs—but also possess the authority to modify them directly.
The application of AI in games is also expected to become more sophisticated. To support environments with tens of millions of concurrent users, AI-based natural language responses for in-game characters will utilize: ▲ a multi-module collaboration system where several small models cooperate to keep costs and latency at realistic levels; ▲ inference logic embedding, which trains the model's internal weights rather than just outputting token sequences; and ▲ the application of FP8 mixed-precision training to significantly reduce memory and computational requirements compared to traditional BF16. Additionally, the company plans to actively utilize AI NPCs in its new life simulation game, *Petit Planet*.
While miHoYo has stated a maximum investment of 100 billion yuan, the actual scale and speed of execution may vary. However, as the company is privately held and does not disclose detailed financial information, the specifics of the investment are unlikely to be fully transparent. Nevertheless, miHoYo's massive AI investment and full-stack in-house development strategy demonstrate that Liu Wei views AI not merely as a tool for R&D efficiency, but as the foundational technology for the next generation of gaming platforms.
Meanwhile, Liu Wei stated that if the project ultimately fails, he would "accept it" as "just having set off a big firework." This has been interpreted both as a remark made possible by the use of self-capital without pressure from external shareholders, and as a sign of his determination to pursue long-term investment without being swayed by short-term performance pressures. Furthermore, his early angel investment in the Chinese LLM startup MiniMax, which yielded high returns, and the registration of miHoYo's proprietary AI model, *Glossa*, serve as grounds for optimism regarding the success of their AI business.
Comments (0)