Kimi K3 Pauses New Subscriptions After 48 Hours of Demand Maxed GPU Capacity

Moonshot AI paused new Kimi K3 subscriptions on July 19 after demand in the 48 hours since launch pushed GPU capacity to its limits. Existing subscribers keep their access unaffected. Kimi K3 launched July 17 as a 2.8-trillion-parameter model priced at $3 per million input tokens - cheaper than every comparable frontier model - and the Kimi K3 subscription pause announcement hit 2.5 million views within hours of going live on X, with 17,000 likes.

[X EMBED] URL: https://x.com/Kimi_Moonshot/status/[insert-tweet-id] Caption: @Kimi_Moonshot · Jul 19, 2026

New Spots Reopen in Batches - No Timeline Given

Moonshot said new subscription slots will open in batches as capacity scales up. No date was given for when that happens. Compute for current members gets priority during the pause, and batch-releasing new spots rather than a flat waitlist gives Moonshot control over how quickly load grows during the scale-up - though it also leaves potential subscribers with no way to queue or estimate when they can sign up.

2.5 Million Views in Hours Shows Demand Beat Moonshot's Own Projections

Moonshot's own language in the announcement - "far more love than we expected" - confirms demand exceeded internal projections. No specific figures on new subscriber volume or GPU utilization have been shared. Kimi K3's combination of a 2.8-trillion-parameter architecture, open-weight promise on July 27, and pricing 70% below Claude Fable 5 on agentic tasks created a demand curve Moonshot had not fully prepared infrastructure for.

Kimi Splits Into Two Tiers - Kimi Membership and Kimi Code Membership

Alongside the pause, Moonshot confirmed membership will split into two dedicated plans: Kimi Membership covering Kimi Web, App, and Work; and Kimi Code Membership covering coding workflows. No pricing for either tier has been announced. Splitting compute pools by use type lets Moonshot allocate GPU budget more precisely rather than letting general chat and heavy coding workflows compete for the same capacity - and with K3 attracting developers specifically for coding automation, the coding tier is likely where the load is concentrated.

July 27 is when Kimi K3 weights drop. A subscription pause six days before that date is probably not a coincidence - teams building actively on the API before self-hosting becomes possible represent exactly the high-compute users that strain capacity most, and Moonshot needs infrastructure stable before running the API becomes optional for teams with enough GPUs to host a 2.8-trillion-parameter model themselves.