Moonshot AI hit $1B ARR — the cheap model tier grew up

In July, Moonshot AI paused new Kimi K3 subscriptions about 48 hours after launch because it ran out of GPUs. Two months later, Bloomberg reports the company told investors its annualized revenue topped $1 billion in August, up from $300 million in June, and that it is targeting $2 billion by year-end. The cheap tier of the model market stopped being a curiosity somewhere in there.

What actually happened

The revenue curve Bloomberg describes is steep: roughly $100 million annualized in March, $300 million in June, past $1 billion in August. The inflection is July's Kimi K3 release — a 2.8-trillion-parameter open-weight model with a million-token context window that landed at the top of the coding and agentic benchmarks while costing a fraction of US frontier pricing.

The second-order evidence is stronger than the revenue claim. Cognition post-trained SWE-2 on K3 and shipped it on September 10 at 50.0% on FrontierCode 1.1 against Fable 5.1's 50.9% — within a point, at 64% lower cost. When a US coding-agent company builds its flagship on your base model, that is a durability signal no investor deck provides.

One caution on the numbers: annualized revenue reported to investors is not audited revenue, and "annualized" means one strong month multiplied by twelve. We would treat $1 billion as a direction, not a balance sheet. Reports of a pre-IPO valuation around $50 billion are circulating; we are not going to repeat a figure we cannot trace to a filing.

Why a scaled low-cost vendor matters for your business

Most teams treat the cheap model as the thing you fall back to when the bill hurts. That framing assumed the cheap vendor might not be there in six months. Revenue at this scale buys capacity, and capacity was the exact reason K3 was unusable in July.

So re-run the routing math. If a model within one point of frontier on your task class costs 60-some percent less, the default route should not be the expensive one. Route by task difficulty: a dependency bump, a support-macro rewrite, and a catalog translation do not need the same model as a payment-flow refactor.

Two things we would not do. We would not sign anything long. A vendor tripling ARR in two months is repricing, and every price you see this quarter is provisional in both directions. And we would not put a China-hosted endpoint in a path that carries customer data without reading your own compliance posture first — the open weights are a genuine hedge here, because self-hosting or a Western host is a real option with K3 in a way it is not with closed frontier models.

Sources: Bloomberg: China AI Star Moonshot Eyes $2 Billion Annualized Sales in 2026, Cognition: SWE-2.