At the Cloud Summit, I Finally Understood miHoYo's 100-Billion AI Ambition

[Core Events and Technology Overview]

At the 2024 Cloud Summit, miHoYo founder Cai Haoyu (Brother Da Wei) rarely publicly elaborated on the company's strategic blueprint in the AI field, stating bluntly, 'If we can't do it, come and slap me in the face in a year or two.' Behind this statement is miHoYo's AI R&D system with a cumulative investment scale reaching the 100-billion level. miHoYo's AI layout is not limited to shallow applications like in-game dialogue enhancement, but builds a complete technology stack from underlying computing power scheduling, self-developed multimodal large models, to upper-layer AI Agent game engines. The core goal is to comprehensively reconstruct the content production pipeline in traditional game development—which heavily relies on human labor, including plot arrangement, character performance, scene construction, and interaction systems—into an AI-driven generative pipeline.

Based on public information, miHoYo's AI technology matrix contains at least three major pillars: first, a large language model system for NPC behavioral decision-making and natural language interaction, supporting long-context memory and character personality consistency modeling; second, a multimodal generation pipeline, covering 3D asset generation, animation motion synthesis, voice TTS, and music generation, aiming to compress the labor costs of art and audio pipelines; third, an AI Agent framework, enabling non-player characters in games to possess autonomous planning, multi-turn dialogue, and emotional evolution capabilities, rather than relying on pre-written scripted finite state machines. In his speech, Cai Haoyu emphasized that AI should not merely be a cost-reduction tool, but should become the core engine for creating new interaction paradigms. This judgment clearly distinguishes miHoYo's AI strategy from peers who remain at the 'AI-assisted art outsourcing' stage.

Notably, miHoYo's investment in computing power is equally significant. The company has established deep cooperation with cloud vendors like Alibaba Cloud, conducting engineering optimizations in GPU cluster scheduling, elastic inference, and training co-location, while internally forming a dedicated AI Infra team responsible for the self-developed iteration of model training frameworks and inference engines. From recruitment information and technical sharing, it can be inferred that its training cluster scale has reached over a thousand cards, supporting the pre-training and continuous RLHF alignment of multimodal large models. This asset-heavy model is extremely rare in the gaming industry, closer to the infrastructure configuration standards of top-tier AI labs.

[Technical Principles and Core Breakthroughs]

From the perspective of technical architecture, the core breakthrough of miHoYo's AI system lies in the deep coupling of large language models with the runtime of game engines. Traditional game NPCs rely on behavior trees or finite state machines, and dialogue systems use retrieval-based or template-matching solutions, where interaction depth and freedom are limited by the scale of pre-written scripts. miHoYo's approach is to use LLMs as the 'cognitive hub' of NPCs, constraining model behavior within the safe boundaries of game logic through structured outputs—meaning the model not only generates natural language replies but also outputs action commands, emotional state tags, and goal planning sequences, which are then parsed by the execution layer on the engine side to drive character animation and AI behavior. This architecture requires the model to have extremely low latency during inference (usually controlled within 200ms to maintain conversation fluency) and highly deterministic outputs, thus it is speculated that they adopted inference optimization strategies such as quantization compression (INT8/INT4) and KV Cache reuse.

At the multimodal generation level, miHoYo's tech stack involves image and 3D asset generation models based on the DiT (Diffusion Transformer) architecture, motion generation networks based on flow matching, and a self-developed voice TTS system. 3D asset generation is the most engineering-difficult part—it requires generating mesh assets with reasonable topology, standard UV unwrapping, and directly importable into engines directly from text or concept art. The current best academic solutions still struggle to stably meet production-grade quality standards. miHoYo's investment in this field suggests it may have adopted a two-stage strategy: the generation model produces rough assets, which then go through an automated post-processing pipeline for topology reconstruction and texture baking, ultimately with manual quality control. This 'AI generation + automated post-processing + manual refinement' hybrid pipeline is currently the most engineering-viable path.

Regarding NPC personality consistency modeling, the technical challenge lies in making characters maintain stable personality traits, memory coherence, and emotional development arcs over thousands of rounds of interaction. This requires solutions beyond simple System Prompts—possibly involving encoding character settings into persistent vector memory, introducing hierarchical memory architectures (short-term dialogue buffer + long-term event summary + core personality vector), and decoding interventions for personality constraints during inference. Cai Haoyu's mention of 'if we can't do it' largely points to this technical bottleneck: currently, even the most advanced commercial large models still experience personality drift and memory contradictions after ultra-long multi-turn interactions, which requires fundamental innovations in model architecture and training paradigms.

[Industry Background and Competitive Landscape]

Examining miHoYo's AI strategy within the global competitive coordinate system, its positioning is quite unique. Unlike pure AI labs like OpenAI, Anthropic, and Google, miHoYo possesses massive game content consumption scenarios and vast user interaction data, providing a natural RLHF data flywheel for its AI system—every conversation and behavioral choice by players with NPCs can serve as feedback signals for model iteration. Compared to gaming giants like Tencent and NetEase, miHoYo's advantages lie in its short decision-making chain, aggressive technical route, and lack of legacy tech stack baggage; its disadvantages are that its computing power scale and data diversity still cannot compete with top AI companies, and the data distribution of gaming scenarios may be biased, making it difficult to directly transfer to the general intelligence field.

In the domestic game AI track, NetEase has launched a self-developed large model for game scenarios, while Tencent is piloting its Hunyuan large model in game NPC and content moderation scenarios. However, both companies' AI investments are viewed more as business empowerment tools rather than core company-level strategies. miHoYo's differentiation lies in Cai Haoyu elevating AI to the strategic height of 'defining the next generation of interactive entertainment forms,' meaning it is willing to bear higher technical risks and longer return cycles. From the perspective of industry evolution trends, gaming is becoming the ideal testing ground for AI Agent technology—closed environments, clear goals, rich interactions, and instant feedback, these characteristics may make the iteration speed of game AI surpass that of general-purpose conversational assistants.

Looking globally, Nvidia's ACE (Avatar Cloud Engine) and companies like Convai are also building game NPC agent platforms, but mostly as general-purpose middleware solutions, lacking deep customization with specific game worlds. miHoYo's choice to develop a full stack in-house rather than purchasing third-party solutions indicates its pursuit of deep coupling between NPCs and the game worldview, plot systems, and combat mechanics, an integration level difficult for general middleware to provide. This choice also makes its technical barrier higher, but it also bears all R&D risks.

[Developer and Industrial Implementation Insights]

From the engineering implementation dimension, the biggest challenge of miHoYo's AI strategy lies in inference cost control. If every NPC interaction in an online game with tens of millions of active users requires calling a large model for inference, based on current GPU computing power prices and token costs, the inference cost for NPC dialogue alone could reach millions or even tens of millions per month. This places extreme requirements on the model's inference efficiency—requiring an engineering optimal solution among model parameter size, quantization precision, batching strategies, and edge deployment. It is speculated that miHoYo is exploring solutions including: small model distillation (distilling large model capabilities into sub-7B specifications to reduce single inference costs), speculative decoding acceleration, and rule-based pre-filtering layers (routing high-frequency simple dialogues through rule systems, calling LLMs only for complex interactions).

For developers and the industry chain, if miHoYo's AI pipeline is successfully validated, it will provide a transferable methodology for the entire gaming industry. The 3D asset generation pipeline can reduce art outsourcing costs by 40-60%, the AI-driven NPC system can increase plot content production efficiency several-fold, and the multimodal generation pipeline enables small teams to produce AAA-level content volumes. However, migration costs cannot be ignored—the engineering complexity of self-developed AI pipelines is extremely high, involving multiple stages like model training, inference deployment, engine integration, and quality assessment, making it difficult for small and medium teams to replicate. A more realistic path is relying on engine manufacturers like Unity and Unreal to package AI capabilities into standardized tools, or for miHoYo to open some capabilities externally in a platform-based manner.

At the hardware level, miHoYo's AI strategy also places new demands on edge-side computing power. If NPC inference is entirely completed in the cloud, it puts pressure on network latency and bandwidth; if some inference is offloaded to the edge side, it requires mobile devices to have the ability to run quantized small models. Current flagship smartphone NPUs can support real-time inference of INT4 models under 3B, but multimodal generation (especially 3D assets and animations) still relies on cloud computing power. This means miHoYo's AI-enhanced content may present differentiated experiences on different devices, requiring a tiered strategy to be preset in the technical solution.

[Comprehensive Review and Key Points]

Overall, miHoYo's 100-billion AI ambition is not a capital narrative, but a technical judgment based on the underlying logic transformation of the gaming industry. Cai Haoyu's statement of 'come slap me in the face if we can't do it' is both pressure on the team and a signal to the industry—game interaction paradigms are migrating from 'limited choices of pre-written scripts' to 'infinite possibilities driven by AI.' If this migration succeeds, it will redefine the form of content consumption: games will no longer be products with fixed experiences, but dynamic worlds where unique content is generated with every interaction. In the next 1-2 years, the key observation metrics will be the actual implementation scale of AI NPCs in miHoYo's products and player experience feedback, which will become the core basis for judging the feasibility of its technical route.

From a broader industry perspective, breakthroughs in game AI may feed back into the general AI field. Game environments, as the best training grounds for AI Agents, accumulate long-term memory, multimodal planning, and real-time decision-making capabilities that can be transferred to scenarios like robot control, virtual assistants, and automated office work. If miHoYo can achieve a technical closed-loop on game NPC agents, its accumulated Agent framework and multimodal reasoning capabilities will have cross-domain reuse value. Regarding future trends, we expect the emergence of 'AI-native' game studios in the gaming industry—new categories that use AI generation as the core pipeline and AI Agents as the core interaction mode from project inception, which could be the biggest paradigm shift in the gaming industry since the free-to-play revolution.

Ultimately, the success or failure of miHoYo's AI strategy depends on one essential question: can AI-generated game content achieve the emotional depth and artistic quality of human creators? Technology can reduce costs and improve efficiency, but the ultimate competitiveness of the content industry still lies in whether it can reach users' emotions. If AI can only produce 'qualified but mediocre' content, its value remains at the level of a cost-reduction tool; if AI can surpass human creators in specific dimensions—such as interaction freedom and personalized experiences—it will open up a new era of content consumption. Cai Haoyu's bet is precisely on the latter.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.