The US may find it hard to shrug off Moonshot’s AI shockwaves

More than a year after the DeepSeek shock, another “open” AI model from China has sent a severe tremor through Silicon Valley. This time, it may be harder for the US AI companies to shrug off the challenge. That has made a response from America all the more likely — with unpredictable consequences for both the AI companies and their customers.

The latest shockwaves were caused by l ast week’s launch of Kimi K3, a model from Chinese start-up Moonshot that has come close to matching the most advanced competitors from the US frontier AI labs.

The angst this has induced is reminiscent of the worries over DeepSeek’s R1. Its low training costs appeared to threaten the far more expensive models being developed in the US. Wall Street came to terms with that threat, believing that US frontier models still had a meaningful lead, even if it had been cut to less than a year. But Kimi K3 — along with other Chinese models like last month’s GLM 5.2 and the Qwen 3.8 model that Alibaba announced at the weekend — have all but erased the time advantage.

Some US officials have been quick to accuse Moonshot of copying US rivals (using a technique called distillation) and getting around export controls on advanced US chips. There are signs, though, that innovation is playing an increasingly important part.

Even a top executive at OpenAI conceded that Kimi K3’s advances did not look like the result of copying, while its architectural improvements over earlier models have drawn admiration in the US. This looks like real competition, not just emulation.

In economic terms, the competitive effects are not clear-cut. Open-weight models (a limited form of open-source software) are cheaper because the companies that develop them do not look to recover their training costs, and because they are run either on a customer’s own systems or hosted by cloud-computing companies that compete on cost.

Very large models like Kimi K3, though, are expensive to run. Also, the true cost to a user is not…