Moonshot's Kimi Takes 61% of Enterprise Open-Source AI Spending. Meta's Llama Has 7%.

A Menlo Ventures survey of nearly 400 U.S. enterprise decision-makers puts Moonshot AI’s Kimi models far ahead of DeepSeek at 15% and Z.ai’s GLM at 11% in the dollars companies pay for open models.

A company that adopts open models to cut its token bill is mostly choosing among Chinese suppliers, because Meta’s Llama, the best-placed Western option, holds 7%. The lower price comes with a second decision: whose model the company is willing to run.

How much of the AI budget is open-source?

Little. Open-source models take 4% of enterprise spending on models, down from 6% a year earlier, while frontier models climbed to 63% from 35%. Their share of usage, counted in tokens, moved the other way, to 19% from 11%.

The wider pool is larger still. Anthropic, OpenAI and Google take $122.7 billion, or 69%, of the $177 billion enterprises spend on AI deployment overall.

Why does a fifth of the usage bring 4% of the spend?

Price. The average open-source token fell 3.3-fold in price, to 13 cents per million, while the blended price of closed-model tokens fell 24%, to 70 cents. By The Denominator’s arithmetic, an open token now costs less than a fifth of the closed average, so its share of usage can nearly double while its share of spending slips.

Volume did the rest of the climbing. Open-source usage rose 22.5 times to an estimated 2.9 quadrillion tokens a month, much of it on inference clouds such as Fireworks and Modal.

Who builds on Kimi?

Cursor built its Composer 2 model on Kimi K2.5, according to Menlo, with Fireworks providing inference infrastructure for the reinforcement-learning rollouts used in training. Kimi’s share is almost nine times Llama’s.

Few companies go that far. Only 10% of AI builders, the companies building products on models, post-train open models, meaning they train them further on their own tasks, up from 6% a year earlier. Among those that do, 91% use outside partners.

Whose model runs in the company’s stack, and whose cloud serves it?