Overview
A new Menlo Ventures report finds that about 89% of global corporate spending on open-source large models flows to Chinese vendors: Moonshot’s Kimi leads with a 61% share, DeepSeek 15%, Zhipu GLM 11%, while the highest Western share—Meta’s Llama—is only about 7%. Crucially, the report measures actual enterprise paid expenditure rather than downloads, offering a real distribution of open models’ commercial conversion. Meanwhile, DeepSeek disclosed that its open-source agent app Harness reached 1.54 million DAU as of November 4, consuming roughly 92 trillion tokens daily with about 35,000 third-party plugins—signaling the open camp’s shift from “price advantage” to “paid-share advantage” and “ecosystem platformization.”
The data reflects a structural shift in global AI competition: Chinese open models, through rapid iteration and adaptation to Chinese and multilingual scenarios, are building moats in enterprise paid markets, squeezing the price space of Western closed APIs.
Background & Interpretation
1. Technical Background and Evolution
The open-vs-closed debate is long-standing. Over the past year, Chinese vendors—DeepSeek, Kimi, Zhipu, Qwen, ERNIE—have densely released weights, with capability quickly approaching the closed-source frontier. Menlo’s data shows open models’ edge lies not in “free” but in enterprise control and total-cost advantages from deployability, fine-tuning and privatization.
2. Drivers and Underlying Mechanism
Enterprise spend flowing to open source is driven by “sovereign AI” and cost discipline: sensitive sectors like finance and government need local deployment and data residency, which open weights naturally fit; meanwhile open models compress inference cost to a fraction of closed APIs via quantization, distillation and edge deployment. Plugin ecosystems and agent runtimes shift competition from “leaderboards” to “runtime and distribution.”
3. Market and Industry Implications
The eastward tilt means a transfer of MaaS distribution power and developer-ecosystem gravity. Domestic operators (e.g., China Mobile Cloud’s joint token-development center with Zhipu) are becoming key distribution channels, while Ollama and llama.cpp set de-facto deployment standards. Closed vendors are forced to respond with “price cuts for scale.”
Implications & Outlook
Near term, open models’ enterprise lead will accelerate closed-API price cuts. Medium term, the battleground moves from “parameters and benchmarks” to “plugin ecosystems, token consumption and runtime distribution.” Key variables: third-party re-testing of open-model capability, enterprise private-deployment penetration, and Western regulatory stance on open-model export and safety.
Takeaways for Industry Participants
Model vendors should move beyond the “open equals free” narrative, building sustainable revenue via plugin marketplaces, enterprise support and fine-tuning services. Application developers can cut inference cost and retain data sovereignty with open weights. Enterprises should evaluate total cost of ownership and compliance boundaries rather than comparing benchmark scores alone.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.