China's Kimi K3 sparks chip stock rout; US AI lead questioned

Moonshot, an Alibaba-backed startup, released Kimi K3 on Thursday, a model featuring 2.8 trillion parameters—making it the largest open model ever released. The company claims the model matches Anthropic's Claude Fable and OpenAI's GPT-5.6 in capability. Starting July 27, Moonshot will allow anyone to download Kimi K3 for free. The model was immediately tested on Arena's coding leaderboard and achieved 1,679 points, securing the top ranking and displacing Claude Fable 5 into second place.

The announcement sent shockwaves through financial markets. The Philadelphia Semiconductor Index, which tracks America's biggest chipmakers, fell 12.5% in a single week—its worst week in over 15 months. Nvidia, AMD, and Broadcom all experienced sharp declines, though the index remained up more than 60% for the year. Walter Bloomberg noted that "investors are questioning the strength of the AI-driven rally after China's Moonshot unveiled a powerful open AI model."

Price is a central driver of Kimi K3's competitive threat. Moonshot charges $3 per million input tokens for cloud access, while Claude Fable 5 costs $10 per million tokens and Anthropic charges $56—roughly 19 times more expensive. Investor Chamath Palihapitiya highlighted the disparity on CNBC's Squawk Box, noting that Chinese labs price at around 50 cents per million tokens. Kimi K3 can read a million tokens at once, sufficient to hold an entire codebase in a single prompt. Moonshot founder and CEO Zhilin Yang outlined the company's scaling strategy in a recent presentation, describing three approaches: squeeze more efficiency from each token, stretch context windows, and run agent swarms in parallel. He emphasized that "token efficiency is not just about efficiency. It's actually also about improving the upper bound of intelligence."

The market reaction mirrors the shock from January 2025, when China's DeepSeek model stunned Wall Street and triggered Nvidia's largest single-day loss in market history—$589 billion（約94兆円）. CNBC host Jim Cramer expects a repeat, though he frames the deeper issue as trust between the US and China. On July 17, 2026, Cramer posted: "There will always be another Deep Seek. But the one thing that has become consistent is the ingrained lack of trust between the two countries. I don't care who you hire–CrowdStrike, Palo Alto, you have to be a first class idiot to let China have any of your data, even your name." Chinese AI stocks themselves fell in Hong Kong trading on Friday, with Zhipu dropping 28% and MiniMax losing 16%, suggesting broader market uncertainty. Traders have begun betting on the competitive intensity itself: Bernstein notes that crypto-style derivatives now trade on AI computing power, with CME Group planning the first compute futures with Silicon Data and ICE announcing GPU contracts with Ornn.

The longer-term question hinges on whether US technological and trust advantages outweigh Chinese cost leadership. America still holds the top scores on most major benchmarks. However, July 27 marks a turning point: anyone can then run Kimi K3 locally for free. If trust keeps enterprise money in US models, Cramer's argument holds and the American lead persists. If price and capability prove decisive, the competitive race will tighten significantly and very quickly.

Kimi K3's release strikes at the core vulnerability of the US AI market: cost. While American models maintain performance advantages on many benchmarks, the pricing disparity—Moonshot's $3 per million tokens against Anthropic's $56 and OpenAI's $26—creates an opening that free open-source availability will amplify starting July 27. The market's reaction echoes the DeepSeek shock of January 2025, when a Chinese model's efficiency gains triggered Nvidia's single worst trading day in history, erasing $589 billion（約94兆円） in value. Yet the chip index remains up 60% this year, suggesting investors may be recalibrating rather than reversing course.

The strategic context matters. Washington has been squeezing chip exports to China, and Nvidia recently purged Asian buyers—yet Moonshot trained Kimi K3 on Nvidia's export-grade H800 chips anyway, highlighting the difficulty of enforcing technology controls. Chinese AI stocks themselves fell in Hong Kong trading (Zhipu down 28%, MiniMax down 16%), indicating broader market uncertainty rather than confidence in an imminent Chinese dominance. The real battleground, as observers like Jim Cramer suggest, may not be pure technical capability but trust: whether enterprises remain willing to pay premium prices for US models perceived as safer or aligned with their interests, or whether price and capability converge enough to tip competitive advantage.