Chinese Models Take 89% of Enterprise Open-Source AI Spend, Menlo Ventures Finds

The figures come from "2026: The State of Enterprise AI", the venture firm's fourth annual report, published on November 3. Menlo puts DeepSeek at 15% of open-source spending, GLM at 11% and Alibaba's Qwen at 3%. Meta's Llama is the most popular Western open-source model, with 7%.

This means the main alternative to the big US closed-model labs is now largely Chinese. Companies that want models they can run, inspect or fine-tune themselves are mostly choosing systems built in China, even as Washington tightens controls on the chips those labs depend on. The report names Kimi K3, DeepSeek V4 and GLM-5.3 as open models that come close to frontier performance at a lower price.

The open-source segment is still small in money terms, and it is shrinking. Menlo estimates that open-source models handled 19% of tokens in September 2026, up from 11% a year earlier, but took only 4% of spending, down from 6%. "Open-source is gaining ground in usage, but losing spending share," the authors wrote. Near-frontier and open-source models together account for just under 5% of spending on large language model APIs, the report says.

That gap is mostly about price. The average open-source token fell 3.3 times in price over the year to $0.13 per million tokens, according to the report, while blended pricing for closed models fell 24% to $0.70. Open-source token volume rose 22.5 times to an estimated 2.9 quadrillion tokens a month, compared with 12.1 times growth for closed models.

At the top of the market, Menlo estimates that enterprise spending on deploying AI reached $176.8 billion by September 2026, 5.4 times last year's $33 billion. Anthropic's share of enterprise LLM API spending rose to 55% from 40%, while OpenAI's rose to 31% from 27%. Google's share fell to 9% from 21%. Anthropic and OpenAI now account for 86% of that spending between them, up from 67% a year earlier.

Menlo also found that 10% of companies building AI products now post-train models to fit their needs, up from 6% in 2025. This is one of the main reasons buyers choose open models, because they can adapt the weights themselves.

The Chinese share has been reported in different ways. Some coverage, including the South China Morning Post's, described the 89% figure as a share of global enterprise spending. The underlying survey covered nearly 400 US enterprise decision-makers who are responsible for more than $4 billion in AI spending. Morning Consult ran the survey from September 8 to 26, 2026, and Menlo combined the results with OpenRouter token data from September 1, 2025 to September 28, 2026. The market sizes are Menlo's own estimates, and the firm says it reclassified prior-year figures to fit a new taxonomy.

Menlo is an investor in Anthropic and in OpenRouter, whose usage data feeds the report's token figures. The report's authors are Tim Tully, Joff Redfern, Deedy Das and Derek Xiao.