Open‑source models are becoming global AI infrastructure — and China is cashing in

Cursor, K2.5 and a revealed supply chain

A developer trick exposed a key detail: the popular AI coding tool Cursor quietly routed requests to a model identified as kimi‑k2p5‑rl‑0317‑s515‑fast. Cursor subsequently acknowledged using K2.5 via an authorized partner, Fireworks AI, and Kimi confirmed the licensing chain. That admission turned a technical curiosity into a business story: open‑source models are not just free downloads. They sit at the bottom of a commercial stack — model licensors, hosting and fine‑tuning vendors, integrators such as Cursor — each layer extracting value.

From algorithmic breakthroughs to global tentacles

Chinese‑rooted innovations helped seed this shift. Companies like DeepSeek used MoE (mixture‑of‑experts) and MLA (multi‑modal learning architecture) plus FP8 training to slash training and inference costs, reportedly achieving GPT‑4‑level performance at a fraction of the expense. Kimi’s K2.5, released under a Modified MIT license, quickly climbed aggregator platforms such as OpenRouter and, it has been reported, became a top‑called model ahead of offerings from other Western providers. The upshot: Western tooling increasingly fine‑tunes and bundles Chinese open‑source models, reversing the three‑year narrative that Chinese teams were always “two generations behind.”

Tokens, supply chains and geopolitical angles

Think of model weights and inference tokens like steel and silicon in a factory. It has been reported that daily global token consumption jumped from roughly 100 billion in early 2024 to about 30 trillion by mid‑2025 and to roughly 180 trillion by February 2026 — driven in part by always‑on agent apps. Alibaba (阿里巴巴) has already responded: it created an Alibaba Token Hub (ATH) to build and distribute tokenized compute and services. That matters geopolitically. Open‑source licensing can diffuse software dependency around export controls and sanctions, but underlying chokepoints — high‑end chips, datacenter GPUs and cloud capacity — remain subject to trade policy and supply‑chain pressure.

Who wins, and what does China gain?

Short answer: China gains a positional advantage in the new AI “infrastructure” stack — not because of altruism, but because its models and engineering have become cost‑effective, deployable building blocks that global developers prefer. Revenue flows to those who host, optimize and operate the tokens and models; standards and platform footprints translate into long‑term bargaining power. Will this become a new form of strategic dependence, like batteries or semiconductors? Possibly. The shaping question now is whether Western policy will target compute and hardware more aggressively, or whether commercial interdependence around open‑source models will deepen — and who will ultimately control the taps.