Moonshot AI’s Kimi K3 Release Triggers a Nasdaq and Chip Stock Selloff

Chinese AI lab Moonshot AI released Kimi K3 on July 16, a 2.8 trillion parameter model the company is billing as the largest open-weight AI system yet built. A day later, US markets sold off sharply, with chip stocks taking the heaviest losses as investors weighed what a competitive open-weight model out of China means for the AI infrastructure spending that has driven the sector’s valuations.

What Moonshot released

Moonshot’s own benchmark disclosures say Kimi K3 “still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol,” but that it “demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models,” according to the company’s statement reported by TechCrunch. Independent evaluators have broadly leaned toward that framing rather than dismissed it as marketing. Artificial Analysis, a third party model benchmarking service, put Kimi K3 at an Elo score of 1,547 on its private evaluation, a result independent AI commentator Simon Willison noted places the model “behind only Claude Fable 5” among the systems Artificial Analysis tracks. TechCrunch also cited independent analyses from Arena.ai and Vals AI describing Kimi K3 as competitive with flagship frontier models, and Tom’s Hardware reported that Kimi K3 topped Claude Fable 5 in a Frontend Code Arena benchmark.

The model is available now through Moonshot’s website and API, priced at $3 per million input tokens and $15 per million output tokens, a notable jump from the $0.95 and $4 pricing of Moonshot’s prior Kimi K2.6 model, according to Willison’s testing notes. Moonshot has said the open weights themselves, which would let developers download and run the model on their own infrastructure instead of through Moonshot’s hosted API, are due by July 27, 2026.

Willison’s own testing flagged limitations that complicate the frontier model framing: a test prompt consumed 13,241 reasoning tokens to produce a 3,417 token response, and a simple query registered 95 input tokens against 10 to 30 tokens on comparable models, which Willison attributed to what looks like “an 85 token hidden system prompt.” He also cautioned that his own benchmark, a test that has models draw an SVG of a pelican, says little about the capability that matters most for current generation models: agentic tool calling.

Markets react

The release landed the same day Chinese president Xi Jinping addressed the World AI Conference in Shanghai, and US markets moved the next trading day, Friday, July 17. The Nasdaq Composite fell 1.4%, the S&P 500 declined 1%, and the Dow Jones Industrial Average dropped 407 points, or 0.77%, according to market data reported by KTEN. Nvidia shares fell more than 2%, and the company’s market value briefly slipped to $4.85 trillion, dropping below Apple’s valuation and handing Apple back the title of the world’s most valuable company; Apple shares rose 0.1% on the day. The selloff extended into Asian trading, where Taiwan’s benchmark index closed down more than 6% and Japanese markets fell 4%.

A familiar argument, louder this time

The market reaction revived a debate that first surfaced when Chinese lab DeepSeek released its open source R1 model in January 2025, this time amplified by an ongoing US-China tariff fight and repeated disputes over whether Chinese AI models pose a national security risk. David Sacks, who served as the Trump administration’s AI czar and now co-chairs the President’s Council of Advisors on Science and Technology, argued that US policymakers are tying the country “in knots” with data center bans, state level regulation, and proposals for a federal agency to pre-approve frontier models, calling it “how you lose the AI race.” Former Uber CEO Travis Kalanick echoed concerns that Kimi K3 was built in part by training on the outputs of American models, a practice known as distillation, writing that “if distillation isn’t enforced against, then everyone should be able to distill from everyone else.”

Not everyone in the discussion treated Kimi K3’s performance as a trick. OpenAI’s head of strategic futures, Dean Ball, said Kimi is “a very good model” whose results probably cannot be “explained away by distillation or anything like that,” adding that he was “personally surprised the Chinese state continues to allow the open sourcing of models this good, given potential risks.” Ball predicted regulators would respond not with an outright ban but with what he called soft law pressure: government bulletins and advisories designed to create enough uncertainty that regulated companies avoid Chinese open-weight models on their own. Shakeel Hashim, editor of the AI-focused publication Transformer, pushed back on the alarm, arguing Kimi “likely does not have dangerous cyber capabilities” and that Beijing faces its own incentives to restrict open Chinese models once they do gain those capabilities.

Why it matters

The market reaction was not really about Kimi K3’s specific benchmark scores. It reflected investor concern about what a credible, freely downloadable alternative could do to the subscription and API revenue that closed source AI vendors, and the chipmakers supplying their training and inference hardware, have built growth forecasts around. Friday’s selloff priced in that question well before enterprise buyers have had time to answer it for themselves.