China's Kimi K3 AI Model May Trigger a DeepSeek Moment. Expect a Lot of FUD.

It looks like China just made a major advance in the AI model space.

On Thursday, Moonshot released Kimi K3, which performed well on many major benchmarks, approaching the frontier, ahead of Anthropic’s Opus 4.8 and OpenAI’s GPT 5.5, though generally behind the most advanced models, Anthropic’s Fable and OpenAI’s GPT 5.6.

Moonshot said in a blog post that it plans to release the full model weights by July 27, 2026. An open-weight release allows users to download the model, run it locally, fine-tune it, and customize it more easily.

I’m writing this because I expect Kimi K3 to spur much discussion about the AI model race between the U.S. and China. It will likely affect stock market sentiment, given the prior history with DeepSeek.

First, a word of caution. This is all based on early impressions from less than 24 hours after release. The technical report with details on architecture and training is not out yet and Moonshot is obviously going to paint itself in the best light possible.

Still, Kimi K3 was widely praised and well received by many AI luminaries online. I experimented with the model yesterday afternoon. It seemed very good at least until Moonshot’s servers became so overwhelmed with demand that the site stopped working.

The AI skeptics may use Kimi’s advance to spread fear and uncertainty, just like the worries around DeepSeek. If you remember, the market and AI stocks sold off last year on misplaced predictions of an impending compute glut from DeepSeek’s efficiency gains. Instead, compute demand surged from rising reasoning models use including, ironically, from DeepSeek itself and stocks recovered to new highs.

I expect Kimi’s outcome to be the same. It will drive MORE demand for AI compute. Kimi K3 shows that scaling laws are still intact. The model’s massive size, at nearly 3 trillion parameters, is driving the increase in capabilities and will require gobs of compute to serve.

SemiAnalysis said Kimi K3 is “so large” it won’t fit on a single Nvidia DGX B200 and requires higher-memory, higher-powered AI servers such as the GB300 NVL72 and B300.

FundaAI noted Kimi K3 is NOT a cheap model to serve: “Given the 2.8-trillion-parameter model size, the current max-reasoning-only mode, and Moonshot’s relatively small user base and lower compute utilization, K3 most likely consumes more underlying compute to complete the same task, and its inference gross margins are very likely well below those of Anthropic and OpenAI.”

How should investors digest the Kimi news? My take: