China’s Kimi K3 AI model has had to pause new subscriptions after demand battered Moonshot AI’s compute capacity.
According to the Associated Press, the launch has rattled the US tech crowd because Kimi K3 is being pitched as a powerful, lower-cost Chinese answer to top Western models.
Beijing-based Moonshot AI said the flood of interest had arrived faster than expected. It now has the awkward job of telling new users to jog on until more capacity appears.
“Kimi K3 has received far more love than we expected. Over the past 48 hours, demand has pushed close to the limits of our current capacity,” the outfit said.
The company said it would prioritise existing subscribers while pausing new ones. “We’re adding capacity as fast as we can and will reopen new subscription spots in batches,” Moonshot AI said.
The mess shows the catch in China’s AI sprint. Open-source, cheaper frontier models can grab global attention, but serving millions of punters still needs serious chips, power and plumbing.
Kimi K3 arrived late last week and aimed itself at advanced models from Anthropic, OpenAI and others. At 2.8tn parameters, it is being touted as the world’s largest open-source AI model.
That claim helped create another mini DeepSeek moment. DeepSeek’s V4 had already made markets twitch and forced investors to admit China was not just copying homework.
K3 is “very demanding” in terms of compute. That makes capacity allocation difficult and expensive, particularly when advanced chips are not exactly raining from the sky in China.
By front-end coding capability, Kimi K3 topped major advanced AI models on Arena after release. That sort of leaderboard win tends to send fanboys, investors and rival labs into a flap.
The release has put pressure on US technology stocks because cheaper Chinese models threaten pricing power. American AI firms are already trying to justify enormous bills for compute, talent and corporate theatre.
The timing is useful for Beijing. US-led export controls have blocked China from accessing some advanced technologies, yet Chinese labs keep churning out models that make Silicon Valley look nervous.
Alibaba previewed its Qwen3.8 Max over the weekend, claiming 2.4tn parameters, while Zhipu, or Z.ai, recently rolled out GLM-5.2. The awkward bit for Moonshot is that global attention is lovely until the servers start wheezing.
