US Companies Send 89% of Their Open-Source AI Budget to Chinese Models: Kimi Takes 61%, DeepSeek 15% and GLM 11%

US enterprises paid for open-weight AI models at a ratio of nearly nine dollars to Chinese labs for every one dollar to everyone else in the past year. That is the headline finding of Menlo Ventures' 2026: The State of AI in Enterprise report, published in early November and based on a survey of 400 US enterprise AI decision-makers conducted between September 8 and 26, whose organizations collectively oversee more than $4 billion in AI spending.

The top-line economics are staggering. Annualized enterprise AI deployment spending reached $176.8 billion in 2026, up from about $33 billion in 2025. Within closed models, Anthropic, OpenAI and Google capture 69% of enterprise spend - the frontier trio still owns the premium tier.

The open-source line inverts entirely. Of US enterprise spending on open models, 89% goes to Chinese companies. Moonshot AI's Kimi leads with 61%, DeepSeek follows at 15% and Zhipu's GLM takes 11%, with Alibaba's Qwen at 3%. Meta's Llama, the most popular American open model, holds just 7%.

Price explains most of the shift. As of September 2026, the average open-model API price had fallen to about $0.13 per million tokens versus roughly $0.70 for closed models - a gap of more than five times - down from about $0.45 a year ago. Open models' share of enterprise token volume rose from 11% to 19% over the year, while their share of spending actually fell from 6% to 4%: usage is climbing faster than the bill.

Agents are the consumption engine behind those numbers. The report finds agents generate about 80% of token volume on OpenRouter, and the platform's own monthly data puts the figure at 76%. On OpenRouter, Chinese models have out-consumed American models for 27 consecutive weeks, and in the week of October 26 to November 1 they held four of the top five slots by call volume.

Buyer behavior backs the survey data, though the company claims cited here are self-reported. Lindy, a 25-person San Francisco startup, said it moved its traffic from Anthropic's Claude to DeepSeek-V4 in June, with its CEO telling Chinese media that inference costs fell about 95% - the company's engineering blog puts the figure near 90% - with success rates flat and latency only slightly up. Coinbase's CEO has publicly said the company set GLM and Kimi as default models for engineers, cutting AI spending nearly in half through routing and caching even as token volume grew exponentially.

The application layer concentrates in two categories. Menlo estimates the AI application market at $77.3 billion annualized, up from $19 billion in 2025, with AI coding the largest slice at $34 billion - up 7.7x year over year - and enterprise chatbots and productivity assistants at $25 billion. Claude Code and Codex together take $24 billion of the coding market, a 70% share from roughly $1 billion combined last year. Judgment-heavy categories lag far behind: HR AI is a $0.9 billion market and marketing AI $1.2 billion.

Menlo partner Deedy Das argues that the strongest models capture most enterprise spending because building frontier models is extraordinarily hard, but no lab's lead is permanent as new releases keep resetting the comparison. The report's structure suggests enterprises have settled on a layered strategy: premium closed models for high-stakes core tasks, cheap Chinese open models as the workhorse underneath.

For the industry, the $176.8 billion figure lands amid a brutal market selloff and investor doubts about AI capital expenditure, and it cuts both ways: the money is still flowing and growing more than fivefold year over year, but an increasing share of it is flowing to model vendors that US labs cannot charge for. The open-source battleground has a clear winner for now, and it is not American.