Decades ago, sitting in an undergraduate political theory seminar, I absorbed a lesson that has explained more of the world to me than almost anything I learned afterward. Strip away the ideologies, the rhetoric and the flags, and a remarkable number of conflicts—between nations, between classes, between companies—come down to the relationship between the haves and the have-nots. The have-nots want what the haves possess. They organize, they innovate, they undercut, they share. And then, once they finally acquire the thing they were chasing, something predictable happens: they discover they are no longer quite so enthusiastic about sharing it.
I have been thinking about that lesson all week, ever since the Wall Street Journal reported that officials in Beijing are weighing whether to restrict the export of China’s most advanced AI models. The instinct in much of the coverage has been to treat this as a China story—another chapter in the great-power technology rivalry, another tit-for-tat response to American chip controls. I think that framing misses the point almost entirely. This is not really a story about China. It is a story about incentives, and about a pattern that politics, business and technology have been repeating for centuries. China just happens to be the nation acting it out this season.
Openness as Strategy
Rewind two or three years, and China’s AI industry was unambiguously in the have-not position. OpenAI, Anthropic and Google sat atop the frontier. American labs had the compute, the talent concentration and the head start. Chinese companies were fighting for relevance, credibility and adoption, and they responded with a strategy that anyone who has spent time in the open source world would recognize instantly: they gave the technology away.
DeepSeek, Alibaba’s Qwen team, Zhipu, Moonshot and others released model weights openly, priced their APIs aggressively, and courted the global developer community. Whether this reflected genuine philosophical commitment to openness or cold strategic calculation is, frankly, the wrong question. What matters is that openness aligned perfectly with China’s interests at that stage of the race. When you are behind, openness is a weapon. It buys you adoption you could never purchase, goodwill you could never manufacture, and distribution that routes around every gatekeeper the incumbents control.
This is not a Chinese invention. It is not even an AI invention. It is one of the oldest plays in the challenger’s playbook, and nobody has run it more successfully than the open source community itself.
The Great Infiltrator
I have spent most of my professional life around open source, and if there is one thing its history teaches, it is that open source wins markets from below. Linux did not defeat proprietary Unix in a boardroom bake-off. It crept into enterprises through the server closet, carried in by sysadmins and developers who adopted it because it was free, capable and theirs to modify. By the time the incumbents noticed, the battle was over. Kubernetes ran the same play against proprietary cloud orchestration a generation later. Git displaced an entire market of commercial version control tools. Python, PostgreSQL, Apache, MySQL—the pattern repeats so reliably that it barely qualifies as a pattern anymore. It is simply how open source disrupts: developers first, enterprises second, incumbents last.
And here is the part the open source community sometimes prefers not to say out loud. Open source has never really been about giving away software. It has been about giving away enough software to build an ecosystem that ultimately benefits the creator. Red Hat understood that. So did HashiCorp, Elastic and Confluent, each in their own complicated way. The license was the marketing budget. The community was the sales force. Openness was the strategy, not the destination.
Nations, it turns out, are discovering the same calculus.
The Irony That Was Always There
I will admit that I was among those who found something deeply ironic in watching the Chinese Communist Party preside over the world’s most enthusiastic embrace of open source AI. Here was a political system built on information control, championing a development philosophy built on transparency, collaboration and the free flow of code. Plenty of commentators resolved that tension by accusing Beijing of cynicism—of merely pretending to believe in openness.
I think the more defensible, and more interesting, observation is simpler: openness and China’s strategic interests happened to align. There was no need for anyone in Beijing to believe in the Cathedral or the Bazaar. Open models served the national interest when the national interest was catching up. The tension between a closed political system and an open development model was real, but it was a tension China could comfortably live with as long as the strategy was working.
The Journal’s reporting suggests that the alignment may be ending—not because China’s philosophy changed, but because its position did.
When the Challenger Becomes the Incumbent
Look at the numbers and you can see why. The latest Artificial Analysis intelligence index puts Anthropic’s Fable 5 at the top of the leaderboard at 60, with Opus 4.8, GPT-5.5 and Grok 4.5 clustered just behind. Zhipu’s GLM-5.2 sits at 51—ahead of Google’s Gemini 3.5. Qwen matches Gemini 3.1 Pro. DeepSeek, MiniMax and Moonshot’s Kimi are all within striking distance. The gap between the best American model and the best Chinese model is measured in single digits on a hundred-point scale, not in orders of magnitude.
More telling than any benchmark is what the market is doing. The Journal notes that companies like DoorDash and Harvey—sophisticated American firms with real production workloads and real money on the line—are routing meaningful traffic to Chinese models because the price-performance tradeoff is compelling. Across Silicon Valley, startups quietly mix Qwen, DeepSeek and Kimi into their stacks the way an earlier generation mixed Linux into their data centers. That is the moment everything changes. Not when you become number one, but when the market decides you are good enough to matter.
China is no longer trying to prove it belongs in the race. It has produced technology the rest of the world genuinely wants. And the instant that happens, the have-not logic that made openness so attractive begins to invert. If frontier models are strategic national assets, then handing the weights to anyone with a download link starts to look less like ecosystem building and more like giving away the arsenal. You do not have to agree with the conclusion to understand it. From a geopolitical perspective, it is entirely rational.
America Is Running the Same Play
Before anyone in Washington indulges in too much satisfaction at Beijing’s apparent change of heart, it is worth noticing that the United States arrived at the same conclusion first. American export controls already fence off advanced semiconductors, and the machinery that makes them, behind national security walls—ask Nvidia about CUDA-capable chips, or ASML about where its most advanced lithography systems are permitted to ship. The debate now swirling around Anthropic’s Mythos-class models, available only to approved organizations while a safeguarded version serves the general public, and the broader evolution of U.S. government thinking about who should be allowed to access frontier capabilities, tell the same story from the other side of the Pacific.
The point is not that China is behaving differently from America. The point is that both governments have converged on the same view: frontier AI is not ordinary software. It is infrastructure with national security implications, closer in kind to enriched uranium and stealth technology than to a web framework. When the two greatest rivals on earth independently reach the same conclusion, it is usually worth taking seriously.
Founder’s Remorse, at National Scale
The open source community has a name for what happens next, even if we have never applied it to nations before. Call it founder’s remorse. Companies have repeatedly released software under permissive licenses, watched competitors build lucrative businesses on top of it, and then tried to claw back what they gave away. HashiCorp relicensed Terraform and the community responded by forking it into OpenTofu. Elastic and MongoDB walked similar roads. The lesson from all of those episodes was reassuring, in its way: you cannot un-ring the bell. Code that has been released remains released. The community forks, routes around the restriction, and life goes on.
Future releases, however, are another matter entirely. Nothing obligates DeepSeek’s next flagship, or Qwen’s, to ship with open weights. Nothing obligates Meta’s, for that matter. The existing generation of open models will remain available forever, but the frontier moves fast, and a frontier that stops being published is a frontier the open community can no longer see.
Which brings me to the question I find most interesting in this entire story, and the one almost nobody is asking.
Can You Actually Fork a Frontier Model?
The comfort we draw from the Terraform precedent rests on an assumption: that AI models behave like software. I am not sure they do. When HashiCorp changed its license, the community could fork the code because the code was the whole artifact. Everything needed to continue development was in the repository.
A model checkpoint is not a repository. It is a frozen artifact—billions of numerical weights, the residue of a training process the recipient never sees. Meaningful continued development of a frontier model requires the training corpus and the pipelines that cleaned it, the synthetic data generation methods, the reinforcement learning infrastructure, the evaluation frameworks, the optimizer choices, the hardware recipes, and above all the institutional knowledge accumulated across years of research by teams who learned a thousand expensive lessons the rest of us never hear about. You can fine-tune a released checkpoint. You can quantize it, distill it, deploy it. What you cannot easily do is carry it forward to the next frontier without the machinery that produced it.
We know how to fork software. Do we actually know how to fork a frontier AI model?
If the honest answer is no—and I believe it is closer to no than the “open weights” branding suggests—then the open AI ecosystem is far more fragile than the open source ecosystem it superficially resembles. Linux survived every attempt at enclosure because the community possessed everything required to keep building. The open model community possesses artifacts, not capabilities. If the laboratories that produce those artifacts stop publishing, the community inherits a museum, not a movement.
Intelligence as Infrastructure
Regular readers know I have been arguing for some time that we are watching intelligence become infrastructure—that AI factories are to this century what power plants were to the last, and that an Intelligence Grid is taking shape beneath the application layer the way the electric grid once took shape beneath the appliance boom. If that framing is even approximately right, then nothing about this moment should surprise us. No nation treats its grid as ordinary technology. Governments regulate who can connect to it, who can attack it, and who can export the equipment that builds it. Semiconductors, nuclear technology, advanced lithography, satellite systems—every technology that became strategic infrastructure eventually attracted the same apparatus of controls.
Every nation that develops frontier AI capability will confront the same uncomfortable questions, regardless of its politics. Which models should be export controlled, and who should be permitted to receive weights? Should APIs be restricted, and to whom? At what point does intelligence stop being software and become strategic infrastructure? China is merely answering earlier and more publicly than most.
We may look back on this period as the brief moment when frontier AI was treated like software. Before long, we may instead think of it the way governments think about enriched uranium, advanced semiconductors or stealth technology—not because the code changed, but because its strategic value did.
The Lesson Holds
Which returns me, all these decades later, to that political theory seminar. The story was never really about China. It was about incentives. The have-nots embrace openness because openness helps them compete; the haves discover the virtues of control the moment they possess something everyone else wants. Ideology bends to position far more often than position bends to ideology. Open source itself was never immune to that law—it was, in its most commercially successful forms, an expression of it.
Perhaps the biggest question isn’t whether China is reconsidering open source. It is whether any nation that develops strategically important AI capabilities will ultimately make a different choice. We may be witnessing the moment when artificial intelligence stops being merely software and officially becomes a geopolitical asset.
