Chinese AI models account for 89% of global enterprise spending on open-source systems, with Moonshot AI's Kimi family commanding a 61% share, according to Menlo Ventures' "2026 State of Enterprise AI" report released on November 31,2.
The report, based on a survey of nearly 400 US enterprise AI business leaders, found DeepSeek at 15%, Zhipu's GLM at 11%, and Alibaba's Qwen at 3%, while Meta's Llama held just 7%. Proprietary US rivals, however, continue to capture the vast majority of overall revenue.
Proprietary leaders and the $176.8 billion market
Global enterprise AI deployment spending has reached an annualized $176.8 billion, expanding roughly 100-fold in three years. Anthropic, OpenAI, and Google together captured 69% of that enterprise deployment spending. Anthropic's share of the enterprise large language model API spending market rose from 40% to 55%, OpenAI's rose from 27% to 31%, and Google's fell from 21% to 9%.
AI coding led the enterprise AI application market at $34 billion in annualized spending, with Anthropic's Claude Code and OpenAI's Codex together accounting for 70% of the AI coding market. Enterprise chatbots and productivity assistants accounted for $25 billion. The global enterprise AI application market reached $77.3 billion in annualized spending in 2026.
Menlo Ventures estimates annualized spending on AI model development in 2026 is $102.5 billion.
Token traffic and the agent surge
OpenRouter data shows DeepSeek led with a 29.8% share of text requests in the week of October 26 to November 1, and Chinese models' token usage has exceeded that of US models for 27 consecutive weeks. Agents now contribute 80% of OpenRouter's token usage, roughly seven times direct human usage, and agent token usage has grown 37-fold since agent traffic surpassed human traffic in February.
Separately, an evaluation firm's overhaul of its testing system — making several tests non-public and adjusting question content and weighting — widened the performance gap between open-source and closed-source models from 3 points to 12 points, sparking debate over the true capabilities of open-source models.
ANALYSIS The split between open-source spending flowing overwhelmingly to Chinese models and proprietary revenue concentrating with US labs presents a two-tier market: enterprises rely on Chinese open-weight models for cost-sensitive or customizable workloads while routing higher-value API spending to Anthropic, OpenAI, and Google.
Moonshot AI released the full weights of Kimi K3.1 Pro on November 4[2] and has held talks with European neocloud providers about leasing Nvidia Blackwell GPU clusters[3]. Fifty-five percent of new enterprise AI budgets are net new rather than carved out of traditional software or headcount budgets.