Why Chinese Tech Builds In-House AI Models
Chinese giants like miHoYo, Xiaomi, and Meituan fund in-house AI models instead of buying rivals, while distillation blunts tech leads.
Rather than acquiring independent startups or locking into third-party clouds, Chinese consumer groups are pouring billions into proprietary foundation models.
Across China’s technology sector, major consumer businesses outside the traditional cloud market have decided against acquiring competitors or relying on outside suppliers. From gaming studios to on-demand platforms, large corporations are designing and funding in-house AI models from the ground up. The shift marks a stark departure from the US market, where tech titans routinely park billions in a handful of frontier labs.
Billions Outside Traditional Cloud
While US players like Microsoft and Amazon anchor capital in exclusive partnerships with OpenAI or Anthropic, Chinese consumer giants foot infrastructure bills directly. Gaming giant miHoYo, best known for Genshin Impact, mapped out an internal budget allocating up to 100 billion yuan (roughly $15 billion) across three years for AI compute and internal research, according to company statements shared by co-founder Liu Wei.
Hardware and consumer brands followed suit. Device maker Xiaomi landed its open-weight MiMo-V2.5-Pro model in ninth place globally on the Artificial Analysis leaderboard, while food delivery and retail giant Meituan developed its own LongCat model family. None opted for full acquisitions. Instead, they bought hardware clusters and trained proprietary weights.
Startups Face Public Markets Without Buyers
This reluctance to pursue mergers and acquisitions altered the path for independent Chinese AI startups. Without buyout bids from Alibaba, Tencent, or ByteDance to take them off the board, standalone laboratories had to find liquidity in public markets.
The Hong Kong stock exchange became the primary exit. On January 8, 2026, Zhipu AI raised 4.35 billion Hong Kong dollars in its initial public offering. A day later, MiniMax pulled in 4.8 billion Hong Kong dollars in its market debut. Competitors like DeepSeek, Moonshot, and StepFun are preparing similar public offerings to build cash reserves without corporate takeovers.
Why Build In-House AI Models and How Distillation Works
The widespread push into parallel systems leans on an engineering reality common throughout China: model distillation. By training a smaller, lighter neural network on outputs generated by larger, frontier systems, engineering teams sidestep months of trial-and-error and cut compute budgets substantially.
That shortcut cuts both ways. Whenever one lab makes a benchmark jump, rivals reproduce comparable results within weeks. To feed distillation workflows with frontier US models restricted by export controls, engineering outfits also tap secondary token-reselling channels in tech districts like Beijing’s Haidian, an arrangement documented by trade reports and flagged in an advisory by the US Cybersecurity and Infrastructure Security Agency (CISA).
The ByteDance Exception
The major outlier in this distillation-heavy environment is ByteDance. Founder Zhang Yiming issued an internal directive barring research teams from distilling outputs from American competitors.
The ban serves two purposes. It shields ByteDance from heightened Western regulatory scrutiny around TikTok, and it forces its AI researchers to build core models from first principles rather than inheriting structural habits from foreign rivals.
No Easy Monopolies in Sight
Flush consumer corporate treasuries and fast distillation have produced a market with practically no lasting technical moats. Unlike Silicon Valley’s tight race among a few heavily backed labs, China’s dynamic makes state-of-the-art performance relatively quick to match, but nearly impossible to protect.