Menlo Ventures report: 89% of US enterprise spending on open-source models goes to Chinese models, with Kimi accounting for 61%
The "2026 State of Enterprise AI" report released by Menlo Ventures on November 3, based on a survey of nearly 400 heads of AI business at US enterprises, shows that 89% of US enterprise spending on open-source large models goes to Chinese models, of which Moonshot AI's Kimi accounts for 61%, DeepSeek 15%, Zhipu GLM 11%, and Qwen 3%, while Meta Llama accounts for only 7%. The report also notes that the average token price of open-source models fell from $0.45 per million to $0.13, usage rose 22.5-fold, and open-source models account for 19% of total platform tokens but only 4% of enterprise model spending. Data from OpenRouter for October 26 to November 1 shows that DeepSeek ranks first with a 29.8% share of text requests, and the token usage volume of Chinese models has exceeded that of US models for 27 consecutive weeks.
Why read this
AI score 75
The Menlo Ventures report shows that 89% of US enterprise spending on open-source models goes to Chinese models, and compares the divergence between Kimi and DeepSeek in spending and usage volume.
Topics
Body · AI translation
Nearly 90% of the money American enterprises spend on open-source large models flows into the pockets of Chinese companies.
On November 3, the "2026 State of Enterprise AI Report" released by Silicon Valley venture capital firm Menlo Ventures made this accounting very clear. Based on a survey of nearly 400 heads of AI business at American enterprises, the report gives a striking figure: 89% of what American enterprises spend on open-source large models is paid to Chinese models.
Breaking it down, Moonshot AI's Kimi takes 61%, DeepSeek accounts for 15%, Zhipu's GLM accounts for 11%, and Alibaba's Tongyi Qianwen accounts for 3%. Among Western models, the highest share, Meta Llama, is only 7%.
The money flows to China not because Chinese models are so cheap they are practically free, but because they have raised the capability ceiling of the open-source tier.
Menlo Ventures states bluntly in the report that "as Chinese models define the open-source frontier, they account for 89% of open-source spending." Over the past year, the average token price of open-source models did indeed fall by two-thirds, from $0.45 per million to $0.13 per million, but usage simultaneously rose 22.5-fold, with monthly token volume surging to 2.9 quadrillion.
With prices halved and volume up more than twentyfold, the conclusion is that open source accounts for 19% of total platform tokens but only 4% of enterprise model spending.
Cheap, but used heavily.
The latest weekly ranking (October 26 to November 1) from global AI model aggregation platform OpenRouter shows that DeepSeek ranks first among all providers with a 29.8% share of text requests, up 7% week over week; OpenAI falls to second at 18.8%, and Google drops the most, continuing to slide from 14.4%. Z.ai (Zhipu) ranks fourth at 11.9%, and Tongyi Qianwen ranks sixth at 5.6%.
This week, the token call volume of Chinese large models has exceeded that of American models for the 27th consecutive week; ranked by individual model, Chinese models occupy four of the top five in weekly call volume.
Ranked first, DeepSeek has recently been making one move after another.
On October 17, it upgraded its main online model to DeepSeek-V4.1-Flash-1017. The API calling method and price remained unchanged, but the web and APP ends were simultaneously switched to the new version. A month earlier, the V4.1 Flash released on September 10 was the smallest tier in this new model architecture, natively featuring multimodal visual understanding, with the official description being "higher capability ceiling, faster inference speed, greater throughput."
On OpenRouter, DeepSeek V4.1 Flash processed 114 trillion tokens over the past 30 days, up 95% week over week; the 1017 version updated on October 17 has already reached 6 trillion tokens in a single day.
More than just the model itself. Harness, the desktop agent product launched by DeepSeek on August 13, officially went live on the desktop on September 29, and the latest official figures disclosed are 1.54 million daily active users and an average daily token usage of 92 trillion. On November 3, DeepSeek's web and API experienced a partial outage of about 1 hour, which was restored the same day.
A notable contrast: in the dimension of "actual enterprise spending" counted by Menlo Ventures, Kimi surpasses DeepSeek's 15% with a 61% share; but on OpenRouter's usage ranking by actual request volume, DeepSeek overtakes Kimi to become the Chinese model most frequently called by developers.
Spending concentrates toward Kimi, calls concentrate toward DeepSeek, and behind this is the same group of American enterprises: they are paying for open-source models and also writing code for Chinese models.
This accounting is still continuing.
Menlo Ventures' data shows that in 2026, American enterprise AI deployment spending has surged from last year's $33 billion to $176.8 billion, a 5.4-fold increase in one year; of this, 55% of the new budget is net new addition, not cut from traditional software or labor budgets.
Source: Xinghe Business Observation
This article comes from "ZAKER Finance," published by 36Kr with authorization.