Moonshot AI's Kimi and Silicon Valley's new bout of nerves

In its July 26 episode, Equity, TechCrunch's podcast, devoted the conversation to a question that has been chasing the industry for a year and a half: why does every competent Chinese model trigger a bout of nerves in Silicon Valley and on Wall Street? The trigger this round is Kimi, the model from Beijing startup Moonshot AI.

According to the episode summary, the hosts discuss why Kimi "seemed to panic" American investors and labs, and to what extent that panic responds to something real or to a conditioned market reflex.

Context matters here: this is not an isolated case but a steady drip. Alibaba with the Qwen family, DeepSeek with its successive versions and Moonshot with Kimi have spent two years releasing open weights that climb to the top of coding and reasoning leaderboards, while American labs keep their best models behind a paid API. That asymmetry, more than any specific launch, is what makes markets nervous.

Who Moonshot AI is

Moonshot AI is no stranger. Founded in Beijing in 2023 by Yang Zhilin and backed by Alibaba, it landed on Western radar in July 2025 with Kimi K2, an open-weights model with a mixture of experts architecture of around one trillion parameters (32 billion active per token) that performed particularly well on agentic and coding tasks, with API prices far below American labs.

That move followed the playbook DeepSeek debuted in January 2025: publish open weights, boast about training efficiency and let the benchmarks do the marketing. The outcome of that first episode is stock market history: Nvidia lost close to 589 billion dollars of market capitalization in a single session, the largest one-day drop recorded until then.

What is rational about the panic

There are legitimate reasons for concern. First, price: if an open, cheap model matches the proprietary model of the moment on the tasks that matter, premium API margins suffer. Second, the capex narrative: American tech companies justify investments of hundreds of billions in data centers on the premise that advantage can be bought with compute, and every efficiency leap by someone else erodes that premise. Third, speed: the gap between the closed frontier and the open alternative is now measured in months, not years.

And what is a conditioned reflex

The other half of the analysis is less alarmist. A good benchmark result is not the same as a good product: distribution, trust, regulatory compliance and support weigh as much as leaderboard points. Adoption of Chinese weights in Western companies also runs into internal policies and a regulatory context (export controls included) that is not going to relax. And the cycle itself repeats: after the January 2025 crash, the market recovered its positions within weeks, which suggests the panic is more a problem of expectations than of fundamentals.

Practical takeaways

For development teams, open competition means increasingly capable self-hostable models and falling inference prices, wherever the push comes from.
For companies, the lesson is to run your own evaluations: deciding based on headlines (or stock market panics) is the worst way to choose a model.
* For investors, the metric to watch is not the benchmark of the month but what happens to inference margins and actual data center utilization.

At ElephantPink we work daily with models from different providers, and the conclusion we draw from each of these episodes is the same: panic is bad advice and worse technical criteria. The useful question is not whether "China is catching up with the United States", but how each model performs against your own tasks, while competition does what it does best: lower prices.