Moonshot AI’s Kimi family now takes 61 per cent of all enterprise spending on open-source AI models, according to Menlo Ventures’ annual State of Enterprise AI report, published this week. Chinese models together account for 89 per cent of that spending, leaving Meta’s Llama — the leading Western open model — with just 7 per cent. The figures mark a striking reversal: the open models winning enterprise budgets are mostly built in China, while American labs still take nearly all the money spent on closed models.
The numbers behind Kimi’s lead
The report breaks the open-source market down by model family. Kimi leads with 61 per cent of enterprise spending, followed by DeepSeek at 15 per cent, Z.ai’s GLM series at 11 per cent and Qwen at 3 per cent. Meta’s Llama, the best-known Western open model, trails far behind at 7 per cent, with the rest split among smaller projects.
Menlo’s figures are estimates, not audited accounts. They rest on a September survey of nearly 400 US enterprise decision-makers managing more than $4 billion in AI budgets, alongside filings and third-party reports. The survey was fielded by Morning Consult between September 8 and 26, 2026. Menlo discloses that it invests in Anthropic and AI infrastructure providers covered in the report.
Open models win tokens, not dollars
The spending lead sits inside a wider paradox. Open-source models now handle 19 per cent of enterprise AI tokens, nearly double the 11 per cent of a year ago, yet they take only 4 per cent of model spending. The gap is price: the average open-source token fell 3.3-fold to $0.13 per million, while closed-model tokens average $0.70.
Volume keeps climbing even as prices fall. Open-source token volume rose 22.5-fold to 2.9 quadrillion tokens a month, outpacing closed-source growth of 12.1-fold. The usage is moving east faster than spending, and the gap between tokens and dollars keeps widening.
The pattern shows up in real products. Menlo notes that American coding platform Cursor built its Composer 2 model on Moonshot’s Kimi K2.5, which Cursor then refined on code data and reinforcement learning. A US company shipping a product on a Chinese open model shows capability, not origin, now decides what developers build on.
Our read
Kimi’s 61 per cent is the number that reframes the open-model race. A year ago the default choice was Meta’s Llama; today it is Chinese, and Kimi takes nine times Llama’s share. Moonshot earned it on two strengths: long context enterprises need for documents and code, and a release cadence that keeps developers building on Kimi rather than trying it.
The result also exposes the limits of the open-model boom. Open models take 19 per cent of tokens but only 4 per cent of spending, so Kimi’s dominance is dominance of a small pond. Anthropic and OpenAI together still take 86 per cent of enterprise API spending. Kimi is winning the open race; the closed race is still America’s to lose.
Why It Matters
For enterprises, the message is that the cheapest model is increasingly Chinese, and Kimi is the name buyers spend on. For Moonshot, the 61 per cent share turns open weights from a strategy into a revenue engine. For Meta, the 7 per cent is a warning: Llama’s lead has evaporated, and catching up means beating Kimi on capability, not price.
Background
Kimi is Moonshot AI’s assistant and model family, and its flagship, Kimi K3.1 Pro, launched in October as an API model before its weights were published on November 4. The model has a one-million-token context window and 2.8 trillion parameters, with 104 billion active per token. It landed on Amazon Bedrock this week, and Moonshot launched hosted agents for customers.
Source: Menlo Ventures report and SCMP report.
0 Comments