Moonshot halts Kimi K3 subs as GPU demand maxes out in 48 hours

So much for the idea that open source automatically means lean infrastructure. Moonshot, the AI startup behind the recently released Kimi K3 model, just paused new subscriptions after watching its GPU capacity get swallowed whole in under two days. The company posted on X that “demand has pushed close to the limits of our current capacity,” a remarkably frank admission that their compute planning didn’t match reality.

Existing users won’t see any interruption, and Moonshot says new slots will roll out gradually as capacity allows. But the more interesting move is structural: the company is splitting its offering into two distinct subscription tiers. A general “Kimi Membership” will cover web, app, and productivity features, while a separate “Kimi Code Membership” will handle programming workflows exclusively. The stated goal is to spread computing power more evenly across different use cases and keep the experience from degrading under load.

The timing puts Moonshot in an awkward spotlight. While they’re scrambling to add capacity, competitor Alibaba is already pushing its rival Qwen 3.8 model, which will be open weight for the first time in quite a while. A paid but steeply discounted preview version is live now. That’s a sharp contrast — one company can’t keep the lights on, the other is handing out weights.

For the open-source AI crowd, this is a reality check. Releasing model weights doesn’t magically solve the inference cost problem when actual users show up. Moonshot built something people clearly want, and the infrastructure buckled. Whether the two-tier subscription model actually balances the load or just fragments the user base is an open question — but either way, it’s a sign that the economics of serving open models are still very much unsolved.