Moonshot's $50bn mark prices Kimi at 25 times a run-rate it has yet to reach

Twenty times in twelve months is what the target describes: roughly $100m of annualized revenue a year ago, $2bn by December ^[2][1]. The nearer step is the steeper one. About $300m in June to more than $1bn in August is a bit over three times in eight weeks ^[1][2], and cryptopolitan.com reports those figures went to investors privately ^[1].

Most of the growth came after Kimi K3 arrived in July with 2.8 trillion parameters, finishing near the top of benchmark tests while costing far less to run than big US models ^[5]. The roughly $50bn mark on the money Moonshot is now raising is fifty times the August run-rate and twenty-five times the December target ^[4][3][4]. Annualized revenue is a rate measured at a moment, not a year of collected cash, and the gap between the two widens when the rate has tripled in eight weeks.

Anthropic's account of the same months describes the product from the cost side. The ten-day window it documented works out at 30,000 Kimi prompts a day sent to Claude ^[5]. Set that against the more than 23 million discussions Anthropic attributed to Moonshot between May and July ^[10], an average near 250,000 a day ^[6], and the customer passthrough is about 1.3% of the traffic ^[7]. Moonshot's 23 million is roughly a seventh of the 151 million exchanges Anthropic says it tracked to Alibaba over the same period ^[11][10].

The per-account rates differ by two orders of magnitude. Moonshot's 300,000 prompts moved through 5,380 accounts, under six a day each ^[8]. Anthropic's Alibaba peak of almost 3 million exchanges a day ran through more than 3,500 accounts, about 850 a day each ^[11][9]. Anthropic also said Moonshot kept some of the transcripts and collected them for training ^[9].

"Some of these exchanges included sensitive information, including from individual users, major multinational companies, and state-affiliated actors ... These practices are likely inconsistent with privacy laws and the labs' own terms of service," Anthropic said in the report ^[12]. It could not establish whether Kimi customers were told their prompts went to an outside provider ^[13].

Whether any of this reprices US model vendors cannot be settled from this material, which carries no revenue figure for one of them ^[11]. What it supports is narrower. A lab selling consumer subscriptions and hosted models was, on Anthropic's telling, buying some of its answers from a competitor ^[8], and the counter-thesis holds up: if Kimi's paying users are paying, the model behind the answer is a cost-of-goods line, and 30,000 prompts a day is too small to explain a $1bn run-rate ^[5]. What would break that read is a bigger passthrough figure. If the 300,000 prompts in ten days are a sample of a much larger flow, the cost side of that run-rate looks different, and the report as described does not say ^[8].