Liberal Democracies Dominated Early Open-Source AI – Now China is Winning
Pull up any global leaderboard for open-source AI today and read the names: Qwen, DeepSeek, Kimi, GLM.
They are not from Silicon Valley, the United States, or even a liberal democratic nation with a capitalist economy. They are from Hangzhou, Beijing, and Shanghai, Chinese-born and bred.
Three years ago this would have been unimaginable. The technology America and its allies gave the world, the permissive licensing, the GitHub culture, the “release open weights to the world” ethos, has somehow been snagged by the Chinese Communist Party’s national champions. They are the ones (mostly) flying the open banner now.
This matters for people who are learning about AI, integrating it into their workflows, and building with tools. And it matters for every consumer who cares about the future of technology.
Why open-source matters
When LLMs are released open-source or open-weights, it enables tens of thousands of “code entrepreneurs” to build on top of these models and tweak them to create yet more value for others.
They can host them on their own machines and equipment for maximum privacy and security while not blowing their budgets on “token maxxing”. They are often available at much lower cost using AI harnesses or apps that allow much more experimentation and customization. Greater open-source benefits everyone (even your adversaries).
Agentic AI projects like OpenClaw, Hermes, and the dozens of others released everyday and popularized on X or GitHub are allowing people with little tech proficiency to use these products and find real value, hosted on their own machines or servers at low cost.
I use some of these models myself when I launch them to fix coding or docker issues on my servers (using a harness like OpenCode, VSCodium, or even Codex), or desire maximum AI privacy on apps that integrate their models into Trusted Execution Environments (TEEs) that are end-to-end encrypted. They work well enough for personal use, even if they’re not frontier. And they’re always getting better.
The frontier of innovation
At present, America is leading when it comes to frontier AI models. Nine out of the top 10 models ranked by AI benchmarks are American, made up of various releases by Anthropic, OpenAI, and even Google’s Gemini. These models are where the real leaps in agentic AI performance and intelligence are happening and represent the cutting edge of capabilities as we know them now.
Though GPT, Grok, and Gemma have all had some open-source/open-model releases, the frontier models are usually kept closed-source.
By keeping their training and weights closed off from competitors and charging a premium, the companies can dedicate more resources to cranking out smart models with more complexity, and be able to provide the ungodly amounts of compute needed to do that.
Both of these model types can and should exist, and each has their use case.
But we need more open-source champions from the Land of the Free (and allies) too.
The good news is that the open-source future is not lost. It is just unattended.
We were here first
The early architecture of generative AI is, almost in its entirety, focused on companies here. Google researchers wrote “Attention Is All You Need” in 2017, the paper that gave the world the transformer. Google open-sourced BERT in 2018 and T5 in 2019. Both became foundational. TensorFlow out of Google and PyTorch out of Meta are the libraries every modern AI lab on Earth still uses, including the Chinese ones.
Hugging Face, the platform that hosts more than a million open models, was co-founded in New York and runs on contributions from labs across the liberal-democratic world (though we’ll let the French have that one).
In July 2023, Meta released Llama 2 under a permissive commercial license. It was the first frontier-grade open-weight model that anyone could download, modify, and ship in a product. Llama 2 is what turned open-source AI from a research curiosity into a real industry, and it was a huge boon to anyone starting to build with these tools and understand how AI could be democratized.
Google followed in 2024 with Gemma, a smaller open-weight family aimed at developers, and Elon Musk’s xAI released open versions of Grok in the same year.
In 2025, OpenAI released two open-weights models, large and small versions of GPT-OSS, that are still available and used today.
The Canadian AI Company Cohere Labs has also recently released great benchmark-ranking open models that are gaining steam among enterprise and casual users, and the French company Mistral has also released solid models that see a similar uptake.
This was, until very recently, an American story with an American lead and some European (and Canadian) support. Liberal democracies unite.
The reversal
Then something shifted.
As the competition between the frontier models heated up, they have resulted in different tactics for picking up customers. Anthropic is asking for a globally coordinated pause on AI development the same week the company filed confidentially for a reported $965 billion IPO (my reaction here).
OpenAI, also soon opening itself up to the public markets, has similarly invoked safetyist language when discussing future policy efforts, even endorsing the terrible online safety bill known as KOSA.
These are companies with potentially trillions of dollars at play, and they’ve become key parts of the American economy in the 21st century. While they focus on improving their models and gaining every compute edge, though, the Chinese are focusing on opening theirs.
Alibaba’s Qwen family now captures more than 50% of global open-source model downloads. Roughly 30% of all global AI usage runs on Chinese open-weight models. DeepSeek’s V3 reportedly trained for around $6 million (though some analysts are skeptical of such a low figure), and its descendants now serve developers at $0.20 per million tokens while Anthropic’s Claude Opus runs $15 or more. American frontier companies are closing weights, raising prices, and asking for federal protection.
China is not winning open-source AI because it is more virtuous. It is winning because America’s best-funded AI companies and their most capable models are in a race to win paying customers and pay their bills for massive compute. That has left the door open for Beijing’s slew of tech companies to try a different path.
Why China leans in
Alibaba, makers of the high-performing Qwen models, treats open source as a customer acquisition engine. The company has committed roughly $53 billion to cloud and AI infrastructure over three years, and they offer steep discounts to the customers that lock-in to their cloud, where the actual revenue lives. Their model is simple: Give away the code, sell the compute.
DeepSeek is a different animal. It was spun out of the quantitative hedge fund High-Flyer, whose founder Liang Wenfeng stockpiled Nvidia GPUs in 2021, before US export controls landed. Funded outside the venture capital cycle, DeepSeek prioritizes research over quarterly revenue, and open source accelerates the feedback loop. The company pioneered efficient training techniques like Mixture-of-Experts and routinely hits massive marks on benchmark intelligence per dollar. The “DeepSeek” moment, when its first model was released, caused a collective shock in the tech and investing industries.
Other Chinese open models are also quite popular and frequently used by hobbyist AI enthusiasts, including Moonshot’s Kimi family of models and Z.ai’s GLM models. They’re cheap, relatively smart, and easy to run on local devices or modest cloud servers.
Then there is the geopolitical layer. The US-China Economic and Security Review Commission has noted that open source helps China “overcome constraints in compute.”
In 2022, the Biden Administration enforced early export controls on American chipmakers like NVIDIA and AMD to try to halt China’s AI progress, something many geopolitical hawks would like to see the Trump Administration enforce again.
But those early export controls only gave Chinese firms the incentive to hew close to their own domestic industry, handing tons of cash to Chinese semiconductor makers like Huawei, Cambricon, and Biren Technology to improve their technology, catch up in the GPU race, and find new customers. It was the ultimate policy backfire. And now, customers of Chinese AI firms are no longer just on the Chinese mainland.
Every developer in a foreign country who downloads Qwen or DeepSeek, running on Chinese technology in Chinese data centers, is now a free force multiplier for Chinese AI standards. If the global south runs on Chinese open weights and GPUs, then Beijing sets the agenda.
What US policy should look like
The better strategy starts with treating AI, whether closed or open source, as the strategic American asset it is, not as a national security threat.
Establish a federal regulatory sandbox. Startups need to ship without a 50-state compliance maze. Without federal preemption and smart rules from the top, every small AI company spends its first $10 million on lawyers and lobbyists instead of research. That shouldn’t need to be the case.
End the export-control overreach. AI is not a munition, despite what some frontier companies or political figures claim. The Crypto Wars of the 1990s tried that with encryption and it gave us a weaker American product while every other country built its own. I argued in 2024 that open source is for everyone, including adversaries, because code is speech and speech does not stop at the border. That has not changed. The same should apply to American-made and American-led AI technologies that should flood the world: chips, models, cars, and all technology. We cannot have Chinese-made goods and chips become the standard elsewhere.
Use procurement neutrality. Federal dollars should not quietly tilt the market toward whichever AI company lobbies the hardest. If a federal agency can run an open-weight Gemma or Llama model on its own hardware, it should be free to do so, and probably should prefer it for privacy reasons alone.
Permissionless innovation. Canada’s new “AI for All” plan hands federal officials authority to decide which models pass an ideological screen before they receive procurement preference or approval. The EU’s “stoplight” system tries to smother models before they’re even released. The US should do the opposite. Pick winners on capability and price, not on whether they reflect the right cultural priorities. Let builders build.
Match safety policy to harm, not capability. The frontier companies are already sharing their most capable models with federal researchers and regulators. This is strong self-regulation we should promote. Where AI companies voluntarily share their most capable models with federal researchers, that’s the right instinct, far better than freezing capability by decree.
Why open source still deserves American champions
America’s AI economy relies on our frontier companies that are breaching the borders of what we thought was possible. But open source is how America can keep pluralism in AI. It is how a startup in Austin, a hospital in Cleveland, or a researcher in Atlanta can compete, build, and improve without asking permission, and can create the tools that normal people can use, not just coders on MacBooks.
Open source is how consumers get a model that runs on their own machine, with their own data, on their own terms.
The Trump Administration has so far championed a permissionless approach to AI policy, and it’s one we should encourage because we know consumers benefit from it, whether it’s a frontier AI company or an open-source one.
China’s open-source advantage is real, but it is not destiny. America still has the talent, the capital, and the cultural DNA to lead this field. We wrote the open-source playbook, and we should have the best laws and regulatory environment for all of that to thrive so that consumers can benefit from this time of unfathomable technological innovation. Let’s go.
Published at the Consumer Choice Center.