DeepSeek's release of its V4 model preview on April 23 knocked 9% off shares of Zhipu AI and MiniMax in a single session, triggering an investor rotation out of AI application developers and into chipmakers that signals a new and more brutal phase of China's domestic AI competition.
The selloff is the direct consequence of DeepSeek's positioning. The Hangzhou-based lab released V4 as a frontier open-source model competitive with leading closed-source systems from OpenAI and Google DeepMind, priced at inference costs significantly below anything Zhipu or MiniMax can match. On the same day Zhipu and MiniMax fell 9.1% and 9.4% respectively in Hong Kong trading, Semiconductor Manufacturing International Corp rose 10% and Hua Hong Semiconductor surged 15%. The market was not selling China AI. It was rotating within it , from companies that build models and sell access to companies that supply the compute those models require. That distinction matters.
The irony is sharp. Both Zhipu and MiniMax only recently IPO'd in Hong Kong, and both had extraordinary runs. Zhipu shares climbed 524% from their January debut before last week's selloff. MiniMax was up 488%. The IPOs themselves were driven partly by urgency: as Hello China Tech's December 2025 analysis revealed before both companies listed, each had roughly nine months of runway remaining. Gross margins of 56% and 69% were promising; cash burn against a DeepSeek-accelerated competitive clock was the problem. The capital raised in Hong Kong bought time. DeepSeek V4 just reset how much time they have.
DeepSeek's competitive strategy since R1 has been consistent: release open-weight models at price points that make commercial API alternatives look expensive, then cut prices further. The downstream effect on the Chinese AI model market has been a race to match those economics, which is a race most independent AI labs cannot win on their own. Alibaba, ByteDance, and Tencent can subsidise model losses with revenue from adjacent businesses. Zhipu and MiniMax cannot.
Counterpoint Research's Ivan Su told CNBC that V4 is unlikely to create the same market disruption as R1, given that traders have already adjusted to the reality that Chinese AI is both competitive and cost-effective. That framing, while accurate for global markets, misses the domestic picture. The competition among Chinese labs has materially intensified since R1, and V4 places DeepSeek's open-source model in direct competition with the commercial offerings of Zhipu and MiniMax in a way that R1 did not. Both companies built their commercial cases on access to frontier-quality models. V4 at open-source pricing undercuts that case at the product level.
What this means for AI startups globally
The dynamic playing out in China's AI market is an early preview of what happens when frontier-quality open-source models become freely available at scale. The companies that survive are not necessarily the ones with the best models. They are the ones with the deepest integration into enterprise workflows, the most proprietary data advantages, or the largest balance sheets to absorb losses while the price war runs. For international AI startups evaluating their positioning, the China market provides a live experiment: commoditised model access destroys application-layer margins unless those applications have something that cannot be replicated by plugging in a cheaper model underneath. Vertical depth, proprietary data, and switching costs are the defensible variables. Model quality alone is not.
Also read: John Ternus points Apple toward on-device AI and it could be the most disruptive bet in the industry • OpenAI's official Codex plugin for Claude Code turns the two biggest AI coding tools into a single workflow • Unsloth's custom kernels make LLM fine-tuning viable on consumer GPUs
