MiHoYo bets 100 billion yuan on AI, with the founder saying: If it blows up, just consider it a fireworks display.

Liu Wei's original words: Even if we fail, it's just a big fireworks show.
The power of this statement lies in — he's not saying "We are confident," but "Even if it blows up, we won't regret it."
When the boss of a tech company publicly admits "we might mess up," it's more honest than any strategic announcement.
What does 100 billion yuan mean:
Kimi has raised about 10 billion, MiniMax 8 billion, Zhipu 7 billion.
ByteDance invested 20-30 billion.
MiHoYo directly flipped the table, investing more than all other large model startups combined.
Where does the money come from?
Genshin Impact and Star Rail's peak annual revenue exceeded 30 billion, with profit margins of 50%-60%.
They accumulated 40-50 billion yuan in cash over three years.
100 billion yuan is almost all the past earnings wagered, plus the projected profits for the next three years.
Why would a game company develop foundational large models?
There are three solid reasons:
First, the game engine itself is a world model.
Open-world physics simulation, NPC behavior, dynamic storylines — at the core, it's AI that "understands how the world operates."
These engineering capabilities can be transferred to large model training without starting from scratch.
Second, core capabilities cannot be outsourced.
If future games rely on LLMs to drive NPCs, generate content, and adjust difficulty, handing over the lifeblood to OpenAI?
API price hikes, policy changes, service interruptions — for a company with annual revenue of 30 billion, these are unacceptable single points of failure.
Third, the rules of content production are changing.
3D asset generation has been compressed from weeks of modeling to minutes, and AI-generated animations have entered Cannes.
As the toolchain shifts from "human-made" to "AI-made," the moat for game companies is no longer the size of their art teams, but who has the best generation models.
This is similar to Tesla developing its own chips and Apple designing its M-series chips — vertical integration.
Background is also crucial.
In the first half of 2025, the large model battle was a mess, and Zhu Xiaohu said "not investing in large models," industry sentiment hit rock bottom.
In the second half, DeepSeek proved Chinese teams could create world-class models with fewer resources.
Pessimists left, optimists placed their bets.
Liu Wei's fireworks were launched at this moment.
He knows it might blow up. But he accepts it.