What's this about?
Three days after launching Kimi K3, Moonshot AI paused new subscriptions: 48 hours of demand "pushed close to the limits of our current capacity," so compute is being reserved for existing subscribers. The launch-week victory lap turned into a capacity crisis.
What exactly is the constraint — can't they just buy more servers?
Serving a 2.8T-parameter model takes enormous GPU fleets, and inference capacity (compute for answering queries, as opposed to training) can't be conjured overnight — especially for a Chinese lab under US chip export controls, which restrict access to the top NVIDIA hardware. That's why demand spikes hit Chinese labs harder: the ceiling isn't money, it's physically obtainable compute.
What's the community saying?
HN 232 points, 96 comments
- The moat nobody benchmarks — the top comment wonders if Anthropic/OpenAI stay relevant simply because they're the only ones who can serve frontier demand at scale; enterprises won't tolerate outages.
- Quality praise, speed complaints — users report K3 is genuinely good at code review, but currently "SO slow" under load.
- A subscription pause is the honest move — pausing signups to protect existing users' quality of service gets read charitably, vs. silently degrading everyone.
- Jevons question — is the K3 moment net-new token demand, or just users migrating from pricier models? Nobody has cross-lab consumption data.