Moonshot pauses Kimi K3 sign-ups after demand overwhelms GPUs
Chinese lab Moonshot AI has paused sign-ups for its Kimi K3 model after user demand exhausted its computing power, exposing the global hardware bottleneck now limiting cheap AI deployment.
Moonshot AI has stopped accepting new subscribers to its Kimi K3 platform after just 48 hours, citing a lack of computing power. The Chinese startup said demand "pushed close to the limits of our current capacity," adding simply that "our GPUs are feeling it."
Existing users retain access, with the company directing all spare capacity toward them to maintain stability. New slots will reopen in controlled batches rather than all at once. Moonshot has also split its service into a general membership and a separate tier for programmers. Coding and agentic tasks require long reasoning chains and vast context windows, making them heavily compute-intensive. Isolating these users prevents them from draining resources from casual browsers.
The bottleneck is striking because Kimi K3 is marketed as the world’s largest open-weight model, with 2.8 trillion parameters. Open weights are designed to let third parties run models independently, spreading the server load. However, Moonshot will not release the weights until 27 July. Until then, all traffic remains on its own infrastructure.
Running the model independently is not practical for most organisations anyway. Estimates suggest serving K3 requires a rig of around eight H100 or H200 chips, coupled with a one-million-token context window for text and images. "So much for the idea that open source cuts computing needs," as The Decoder noted. The 27 July weight release will act as a pressure valve, allowing large cloud providers and enterprises to host K3 themselves.
For European investors and tech firms, the incident underscores a shifting market dynamic. Chinese models like K3 are arriving at prices that undercut US rivals, eliminating capability as the primary competitive barrier. Instead, the new constraint is physical infrastructure. European cloud providers looking to host these cheaper models face the same fundamental scarcity of advanced semiconductors. Moonshot is not alone in the squeeze: Anthropic recently cut limits on its Fable 5 model for similar reasons, and Alibaba immediately launched a discounted competitor.
Moonshot is converting its viral moment into tangible revenue. Annual recurring revenue reached $300 million in June, up from $200 million in April. The company is now preparing a Hong Kong listing that could value it above $30 billion. For markets betting on the future of AI, a model too popular to serve is an unusual but compelling pitch, provided the chips hold out.
