Zhipu, MiniMax plunge 10% as Chinese AI model war heats up

Moonshot AI's Kimi K3, a 2.8-trillion-parameter model, tops the Arena coding leaderboard at 1,679 points while charging $3 per million input tokens — less than a third of Claude Fable 5's price. The model's 1-million-token context window allows it to process an entire codebase in a single prompt, a capability that previously required complex chunking strategies. On the Arena leaderboard, Kimi K3 pushed Claude Fable 5 into second place, marking the first time a Chinese open model has held the top spot.

"Token efficiency is not just about efficiency — it's actually about improving the upper bound of intelligence," Zhilin Yang, founder and chief executive officer of Moonshot AI, said in a recent session. Yang outlined three strategies for scaling AI: extracting more from each token, stretching the context window, and running agent swarms in parallel. Kimi K3's 2.8 trillion parameters make it the largest open model ever released, and Moonshot plans to make it freely downloadable starting July 27 — a move designed to accelerate developer adoption.

The sell-off extended beyond Zhipu and MiniMax. In Hong Kong trading Friday, Zhipu dropped 28% and MiniMax lost 16%, according to market data. The broader selloff came as DeepSeek released its V4 series and Alibaba launched Qwen3.8-Max-Preview, with the Qwen3.8 official version expected soon. The rapid cadence — three major model launches in a single week — shows that China's AI labs are locked in an accelerating arms race where each new release must outperform the last. Zhipu's market value has now fallen by more than a third this week, while MiniMax has lost nearly a quarter.

US chip stocks also suffered collateral damage. The Philadelphia Semiconductor Index fell 12.5% in its worst week in 15 months, as investors questioned whether America's AI lead is safe. Nvidia, AMD and Broadcom all fell sharply, though the index remains up more than 60% this year. The selloff echoed the DeepSeek shock in January 2025, when Nvidia lost $589 billion in a single day — the biggest single-day loss in market history, according to CNBC. CNBC host Jim Cramer said the real issue is trust, as the market struggles to price the competitive threat from Chinese AI.

The Pricing War

The pricing gap is the central battleground. Investor Chamath Palihapitiya said on CNBC's Squawk Box that a million tokens costs $56 from Anthropic, $26 from OpenAI, and 50 cents from Chinese labs. Moonshot charges $3 per million input tokens for Kimi K3, compared with $10 for Claude Fable 5. The 70% discount puts pressure on every Chinese AI company to match or justify higher prices through superior performance.

The market is already adapting to the new competitive dynamics. Bernstein says crypto-style derivatives now trade on AI computing power. CME Group plans the first compute futures with Silicon Data, and ICE announced GPU contracts with Ornn — financial instruments that allow investors to hedge or speculate on the cost of AI compute.

For investors, the question is which Chinese AI companies can sustain their moats. Zhipu and MiniMax, once considered frontrunners, now face an existential pricing and performance challenge from Moonshot, DeepSeek, and Alibaba. Moonshot, backed by Alibaba, has the financial runway to sustain its aggressive pricing strategy, while smaller players may struggle to keep pace. The rapid iteration cycle suggests the window for differentiation is narrowing, and the market has already begun repricing the laggards. The next test will come when Qwen3.8's official version launches, potentially adding another competitor to an already crowded field.

This article is for informational purposes only and does not constitute investment advice.