# China's Structural Advantage in Open Source AI

Kevin Xu wrote this for his newsletter Interconnected on 25 June 2025, five months after DeepSeek R1. It builds on a podcast interview with Ion Stoica, the Berkeley professor behind the Sky Computing Lab and co-founder of Databricks, Anyscale and LM Arena, who described several structural advantages China has in open-source AI. Xu's question is why nearly all the leading Chinese labs release open weights under permissive licenses (Apache 2.0 or MIT rather than a corporate license like Llama's) while the leading American labs stay closed. Lambert's [[open-source-ai-reading-list]] files it under "China's structural advantages in open-source", next to Xu's later and longer [[chinese-open-source-a-definitive-history]].

## The DeepSeek moment, told from Hangzhou

Xu opens by dismissing the theory that R1's 20 January release was timed to upstage Trump's inauguration and the Stargate announcement. The deadline that mattered was Chinese New Year a week later, when every engineer in China wants to ship and go home. The lab whose holiday R1 actually ruined was Alibaba's. Chairman Joe Tsai told VivaTech that the Qwen team cancelled travel and slept at the office until it had a competitive model, and Qwen 2.5 Max shipped on 28 January, New Year's Eve. By June the top three open models on Artificial Analysis's intelligence index were DeepSeek, Qwen and MiniMax, the last an Alibaba- and Tencent-backed startup rumored to be heading for a Hong Kong IPO. "One's a dot, two's a line, three's a trend."

## Three advantages

The first is expertise. Jensen Huang's line that China has 50% of the world's AI researchers comes from MacroPolo's talent tracker; Damien Ma, who ran that project, puts it as "47% of the top 20% talent". Xu adds a less obvious point. China's manufacturing and biotech base produces domain experts, and domain expertise matters again now that reinforcement learning is the main way to improve models on real tasks, because someone has to design the reward.

The second is data, and Xu is careful to separate his claim from the cliché of a big country with weak privacy. Chinese internet data is a poor corpus: censorship keeps deleting it, and walled gardens like ByteDance and Tencent cannot be scraped. He had previously argued this could keep China permanently behind. The advantage lies elsewhere. Chinese industry is heavily 5G-connected and IoT-instrumented, so it produces large volumes of vertical data from real problems, which Xu calls the best kind of training and evaluation data. And OpenAI chose not to offer ChatGPT in China (it was not banned at launch, contrary to the usual story), with the other American closed labs following, so they have no use for Chinese internet data. Chinese labs have that corpus to themselves while still reading the global web.

The third is open source as a default. Academic research is open by nature, and the line between academia and commercial labs is more fluid in China than in the US, so labs inherited academia's habits. Xu does not claim this was inevitable. Chinese labs were behind when ChatGPT launched, and open-sourcing is the standard strategy for a follower trying to erode a leader's moat, the same reason he gives for Meta releasing Llama. What began as a defensive move turned into a flywheel. A corporate-backed open lab pays industry salaries while letting researchers keep publishing, releasing current rather than old models builds a developer community, and free customizable weights spread to more use cases, globally by default. ByteDance is the exception he names.

## Two different flywheels

Both countries connect universities to industry, but Xu argues the connection runs on different fuel. In Silicon Valley the intermediary is money: seed checks, giant pay packages, and in 2025 Mark Zuckerberg's $100 million offers and startup buyouts. In China pure financial VC has mostly disappeared, replaced by funds with corporate or government backers, and poaching happens with smaller numbers. What remains is an academic promotion system that requires real-world impact. Zhejiang University, Liang Wenfeng's alma mater, has researchers do tours of duty in companies held by the Zhejiang University Holding Group before they can advance. Xu judges the American version more lucrative and less collaborative, and the Chinese version more efficient, perhaps at the cost of discoveries with no obvious use today.

## Shared artifacts versus silos

The second contrast is Stoica's. An artifact in his sense is a reusable building block: a set of weights, a parallelism library, a deployment toolchain. Open development accumulates them in public, so everyone can see what is solved and aim at what is not. Closed labs work in silos and, in Stoica's estimate, all do roughly the same thing, sharing knowledge only when people move between them. Even Meta was catching up by spending rather than by building on shared work. Xu cites Coatue's estimate of $365 billion in 2025 AI data center spend by Amazon, Microsoft, Google, Meta, Oracle and Tesla, not counting all of Stargate, and suspects much of it is redundant.

He does not claim Chinese labs avoid duplication. Alibaba, Huawei, Baidu and DeepSeek (which runs its own data center to squeeze export-controlled GPUs) will keep building their own infrastructure, and he calls the idea that Beijing will force labs to pool resources another dumb hot take, since competition among the labs is what drives them. He closes with a question he does not answer: what it means that a closed society prefers open source for AI while an open one prefers secrecy.

## Against later sources

Some of the argument holds up in what came after, and some does not. Lambert's [[notes-from-inside-chinas-ai-labs]] (May 2026) confirms the student-heavy teams and the fluid academia-industry boundary, and the [[zai-playbook]] interview (November 2025) says many active PhD students work on GLM. On data the later sources complicate the picture. Lambert found China's commercial data industry poor enough that labs build RL environments in-house, and Z.ai's Zixuan Li says the company leans on synthetic data because it cannot scrape WeChat, so the vertical industrial data Xu describes does not show up in either account of how models are actually trained.

On motive, Li gives a blunter reason than an academic default. Z.ai's flagship was closed in 2024, and the company opened it after R1 showed that American customers will download a Chinese model but will not send data to a Chinese API. Lambert likewise describes the labs as practical rather than committed to openness in principle. Xu's own history nine months later leans more on overseas growth and a two-decade open source culture than on the academic link.

The piece says little about compute. Lambert's visits found every lab short of Nvidia chips and limited by it, which is at least as structural as anything here. Later commentary argues the advantage runs partly through distillation from American models; see [[how-much-does-distillation-matter-for-chinese-llms]] and [[the-distillation-panic]].