China is winning the AI race

Open-weight models are reshaping the AI “race” map, and China appears to be winning that specific contest—largely because Chinese labs release weights in ways that let many third parties host and improve them, while U.S. labs face structural disincentives to do the same at scale. The result is a chart where U.S. names dominate only when the lens is narrow (closed or API-first models), but China’s “blue bars” surge once downloadable, locally usable weights are counted—especially around top performers such as Kimmy, DeepSeek, and Mini Maxm2.

The transcript draws a sharp distinction between open-source and open-weight. Open-source typically means the source code is available so others can reproduce the exact build and outputs; open-weight instead means the model’s parameters (weights) are provided so users can run the model and generate tokens on their own hardware. The argument is that the training data and the full training pipeline are the real scarce assets—like the “engineers” in a development process—so withholding them doesn’t make open-weight meaningless. What matters for practical users is that the weights can be reused, modified, and executed, often under open licenses.

Why China’s open-weight strategy seems to dominate comes down to incentives and infrastructure. Chinese labs release weights even when it limits direct revenue, because hosting them through a wide ecosystem of providers gives them mindshare in markets that otherwise would not adopt their models. The transcript claims that if Chinese models are hosted in China, U.S.-based security teams and regulators would likely block them, and even GPU access restrictions can make local deployment difficult. Open-weight becomes the bridge: it lets U.S. and global providers host the models, and it lets enthusiasts and developers run them themselves.

That ecosystem effect is illustrated with provider competition. When multiple companies can host the same open-weight model, throughput and pricing become competitive, which can squeeze smaller hosts’ margins but also improves availability and performance. Kimmy is cited as taking this further by publishing a “vendor verifier” and tool-calling benchmarks, pushing hosts to meet reliability targets; the transcript frames this as a reputational defense mechanism, since poor tool-calling behavior can damage the model’s standing.

The transcript also argues that open-weight is harder to do responsibly. Once weights are released, labs can’t easily retract or firewall unsafe capabilities; with API models, safeguards can be inserted before or after inference. U.S. labs, facing higher legal and investor scrutiny (security, copyright, reliability), have stronger reasons to keep control via closed models or tightly managed deployments. Meanwhile, Chinese labs are portrayed as less constrained by those liabilities, releasing when they can win benchmarks.

Finally, the transcript claims the U.S. may not be able to win the open-weight “chart” again—because the largest open-weight models (hundreds of billions to a trillion parameters) require massive VRAM and aren’t realistically runnable on consumer hardware. It points to OpenAI’s strategy as an exception: releasing smaller open-weight models like GPT OSS120B and GPT OSS20B that fit within consumer constraints, enabling local inference and fast adoption. The broader hope is that as personal hardware improves and more platforms support local model execution (e.g., browsers and on-device inference), U.S. firms could compete on open-weight where local usability matters. But for now, the transcript’s bottom line is blunt: China’s open-weight approach is the only one that reliably buys relevance in the U.S. market, and that advantage may persist unless incentives and hardware realities change.