When the Have-Nots Become the Haves: Why China’s Open-Source AI Restrictions Aren’t About China
TL;DR: China open-source AI reshaped the market. China spent years releasing open-source AI to compete with OpenAI, Anthropic, and Google. Now Beijing is reportedly weighing restrictions on its most advanced models. The shift isn’t ideological — it’s the oldest pattern in tech and politics: the have-nots embrace openness to compete; the haves discover the virtues of control.
Why is China reportedly restricting its open-source AI models?
According to a Wall Street Journal report, Beijing is weighing export restrictions on the country’s most advanced AI models. The shift comes as Chinese labs including DeepSeek, Qwen, Zhipu, and Kimi have closed the capability gap with U.S. frontier models. Once a strategic asset reaches parity, the incentive to give it away collapses.
What is the “have-nots become the haves” framework?
The framework holds that challengers embrace openness because it helps them compete, while incumbents discover the virtues of control. It’s the same dynamic that drove Linux against proprietary Unix, Kubernetes against VMware, and Git against every centralized version control system before it. Openness was the strategy — never the destination.
Has this happened before in open source?
Yes. HashiCorp relicensed Terraform, prompting the OpenTofu fork. Elastic relicensed Elasticsearch. MongoDB relicensed after AWS commercialized its work. Redis Labs did the same. Each was a “have-not” that became a “have” — and quietly closed the door behind them.
Can you actually fork a frontier model?
Not in the way you can fork software. A model checkpoint is a frozen artifact — billions of weights, the residue of a training process the recipient never sees. Without the training corpus, the pipelines, the RL infrastructure, the evaluation frameworks, and the institutional knowledge, the open community inherits artifacts, not capabilities. A museum, not a movement.
What does this mean for U.S. companies using Chinese open-source models?
Companies including DoorDash and Harvey have reportedly routed production traffic to Chinese models. If Beijing restricts access, dependency risk rises overnight. Enterprises building on any open-weight frontier model — Chinese or otherwise — should assume the license, availability, and update cadence they see today are subject to strategic reversal.
The bigger question for China open-source AI
The bigger question isn’t whether China reconsiders open source. It’s whether any nation that develops strategically important AI ultimately makes a different choice. America ran this playbook first, from cryptography export controls to today’s semiconductor rules. Founder’s remorse, it turns out, scales to nations too.
Key takeaways
- China’s reported AI export restrictions aren’t ideological — they’re structural.
- Open source has always been a challenger’s strategy, not a permanent commitment.
- Frontier model weights are artifacts, not forkable ecosystems.
- Every enterprise AI stack now carries geopolitical dependency risk.
- Expect more sovereign AI, more selective open weights, and fewer truly open frontier releases.
Alan Shimel is the founder and CEO of Techstrong Group. Watch the video edition of Shimmy Says on YouTube and Techstrong.tv.
Transcript
Shimmy says, Shimmy says, Shimmy, Shimmy, Shimmy says, ask me almost anything! " I want to talk today about China. Well, actually, I don't want to just talk about China.
I want to talk about something even bigger than China. I want to talk about power. And if we're going to talk about power, I got to tell you a quick story.
I was a political science major, history major, double major in college. And I remember my senior year, I had a political science seminar where we just... It was a grad-level course almost you took in your senior year of undergrad.
My professor was Dr. Henry Paolucci. I'll never forget that man.
You probably have never heard of Dr. Henry Paolucci. Dr.
Henry Paolucci had the honor of being the 1964 candidate for senator from the great state of New York for the Conservative Party. Not the Republican Party, the Conservative Party. I don't know if the Republican Party even ran someone for senator in 1964 because Dr.
Paolucci had the unlucky chore of running against a guy named Robert Kennedy. Not the Robert Kennedy that you may know running around who nonsense about vaccines and snorts cocaine off toilet bowls. I'm talking about his father, who was a real man, who was a real icon of his time, Bobby Kennedy.
And Kennedy steamrolled Paolucci, and Paolucci never got over it. But I learned something from Dr. Paolucci that I never forgot, and that is when it comes to power, especially in political world and geopolitics, very often it's about the haves and the have-nots.
Right? And this, if you look at the world through the lens of have and have-nots, a lot of things really become very clear to you. Because at the end of the day, it's the same old story.
The have-nots wants what the have-haves. And the haves, once they have it, want to keep it for themselves and not share it with the have-nots. It's human nature, I guess.
When Dr. Paolucci taught us this, I thought it was the ramblings of a frustrated politician. Thought that was interesting.
But here we are all these years almost, I don't know, 45 years later, and I understand now that's truly how the world works. It's about the haves and the have-nots. Now I know what you're thinking.
"Shimmy, why are you telling us this? " It has a lot to do with AI today because that's what I want to talk to you about. " There was a couple of stories, one in the "Journal" and another one I think I saw in "The Washington Post," that China, the "Wall Street Journal" story anyway, is China is actually thinking about restricting access to some of its most advanced AI models.
Now, I know what you're saying. "Shimmy, those Chinese open source models are open source. They're open as in open source.
How could China restrict that? " But you got to look at this versus the haves and have-nots, and this is just another chapter in this AI Cold War that's playing out between the US and China. On the US side, we were all about restricting.
We had export controls. We stopped Anthropic from releasing Fable and Mythos. We delayed OpenAI.
6. At the same time, China was a have-not. The US was the haves here.
China's the have-not, and they didn't have a horse in the AI race, but they desperately wanted one, and we weren't letting them have ours. We were hoarding what we had, what we have. So China took the classic have-not story, right?
They didn't have an OpenAI in China. They didn't have Anthropic to go. They didn't have Google.
They didn't even have SpaceX or have a big copilot. So China had to play catch-up, and they went to one of the oldest scripts in the book when you're a have-not for playing catch-up. They said, "We're going to release our models open source.
We're going to make them inexpensive, free as in beer, and we're going to make them open that we can't pull them back. " I spoke about this on my July 4th thing, right? Free as in beer, free as in freedom.
What could be bad? Developers all over the world started using these Chinese models. Even right here in the US, a whole bunch of companies in Silicon Valley.
As a matter of fact, the last numbers I saw said something like 30% or 35% of the AI model market is now with Chinese open source models. This wasn't an accident. It worked just the way the Chinese wanted it to work.
It worked just the way open source always works. Because you know why? Open source is the great infiltrator.
Think about Linux. Linux didn't beat Unix because CEOs loved Linux, or CFOs loved Linux. It was the developers who installed it, who loved it, who brought it in, system administrators who deployed it.
And then the companies woke up one day and realized that Linux was already everywhere. What were they fighting? Kubernetes, another example.
Same story, right? There were a lot of container orchestrators. Kubernetes infiltrated as open source.
The rest is history. Git, whether you're talking about GitHub, GitLab, or any of the Gits, same old story. Python, PostgreSQL, the list goes on.
Open source rarely attacks from the front door or through the top. It comes in the side door from underneath. Developers adopt it.
The ecosystem forms around it. Business follows. And it gets better because of that community.
And China understood this playbook, and as a have-not, they didn't have a lot of other choices, so they played the best card they had, this open source card. Whether they believed in open source philosophically really didn't matter. But for some of us, I got to tell you, like me, for instance, the irony of the Chinese Communist Party taking the road that they're the open source alternative and that the US are the big, bad closed source guys who are blocking everyone's access to this indispensable technology.
The irony should not be lost on any of us because that's not who they are, and quite frankly, it's not who we are either to be closing stuff up. But anyway, open source was the perfect strategy for the Chinese here in trying to play catch-up and move from have-not to have. So what's happened?
Take a look at this chart that I was able to pull up. It's not my chart. The attribute's right on the chart, and you can see it for yourself.
But if you look here, I think it has the top 10 or 12 open source models. And look, the US ones, as we expect, are right up there front and center. But you already see it's called ZAI is the popular one, but the official name is there.
But it's not that the Chinese models are not in their top five, but when you look at this top 10 or 12, the gap between them and the ones on the top are closing, and they're closing rapidly. There's really not as big a difference as there used to be between the top-performing models and the other ones. 5's right there.
Anthropic's at the top. But look at that GLM. It's roughly on par.
Qwen, another Chinese one, DeepSeek, Kimi, they're all right there. These guys used to be 20, 50 points behind. Now they're single digits.
They're within striking distance. And that's incredibly important because you know what? For a lot of tasks, you don't necessarily need the biggest, most expensive model out there.
These models are good enough for you to use, and that's what companies all over the world are discovering, and that's why they're grabbing that market share. So when you're good enough, you're no longer a have-not, and that's where we are with the Chinese. They're not a have-not anymore.
They're a have. The Wall Street Journal pointed it out, and I think it's fascinating. Companies like DoorDash are using the Chinese models.
Harvey's using these Chinese models. They've become the model of choice for the startups in the Valley. They're using Chinese models, think about that, instead of our own.
Why? Because they're good enough, and the incentives change now. China no longer has to play the have-not game.
They're playing the have game. They don't have to convince the world that they could build good AI. They have.
The world already knows it. So now where does this go? The Chinese Communist Party shows their real stripes, right?
They don't need to champion open source because open source, after all, is built on transparency, collaboration, sharing, freedom, not just freedom to be, freedom to modify. Not exactly the words we associate with the centralized government over in China. Unfortunately, not exactly the words we associate with this government, but we'll come back to that.
So the question is: Did China ever believe in open source? Who knows? Do you care?
No. Because the question is: Did open source help China become a have in the AI race? Yes, absolutely.
So now here's the important part. It's not just that it's China playing this dance that we're playing. They played the open one; we played the closed one.
We're doing, as I said before, the same thing here in America. We're restricting our advanced chips. We're restricting semiconductor equipment.
We're debating who gets access to the frontier models. Anthropic had to restrict access to some of its most advanced capabilities. Governments everywhere are reaching the same conclusion.
This isn't just about ordinary software. This isn't Linux, Unix. This is now we're talking about strategic infrastructure, and that, my friends, is a very different world.
So this is what really fascinates me. I don't think the biggest question is: Will China stop releasing models? I think the bigger question is: Can you actually fork a frontier AI model?
So if China stops releasing their models and it's open source and we can fork it, just like OpenTofu forked Terraform. Linux can be forked. PostgreSQL was forked.
Can we fork GPT? Can DeepSeek or Qwen, can we fork these? I know some people are going to say yes, but I'm not so sure because a model checkpoint is not enough.
You need training data. You need synthetic data. You need reinforcement learning.
You need evaluation systems. You need the infrastructure You need all the institutional knowledge that created that model in the first place. It's not like software code.
Maybe we don't know how to actually fork intelligence, and that's a question I haven't seen enough answers to or enough people talking about. And it fits a bigger story yet still. If you've watched me talk, you've heard me on Shimmy Says, Techstrong Gang, et cetera.
It's about something I call the intelligence grid, AI factories, infrastructure, power generation, networking, semiconductors. I've been saying it for months. We're not building software anymore.
We're building infrastructure. We're building an intelligence grid. Because if intelligence becomes infrastructure, why would governments treat it differently than electricity or nuclear technology, advanced chips?
I don't think they will. This is exactly what we should have expected. It gets even more, the Chinese announced today that they found a backdoor in Anthropic that was specifically aimed at finding Chinese users of Claude.
Anthropic admitted there was a backdoor for a short time only back in March or May, and it was just an experimental thing that was taken out, but maybe it wasn't taken out. Why was it in there to begin with? Did Anthropic want it in there?
Did our government want it in there? Did Anthropic put it in there because our government said, "If you do that, maybe we'll let you release it"? Who knows?
But this is the intrigue. This is the game we're playing now with AI as a strategic lesson. Because, my friends, let me bring it all back for you.
It's about the haves and the have-nots. We are the haves. We've been the haves.
The Chinese were the have-nots. Now they're the haves, too. Now we're both haves.
And what are we doing? Because we're two halves, we're both hoarding what we have, not to let the rest of the world have total access to it so we can control it. Because control starts looking pretty attractive when you're a have.
It always has. That's true for companies, it's true for countries, and frankly, it's true for people. It makes me wonder whether we're looking back at the last couple of years as a unique moment in history, or a moment when the world's most advanced AI models were treated like software when they're not.
Because if intelligence is really going to become a strategic infrastructure, the intelligence grid is real. This may be the last generation of frontier models that we see right now that anyone can simply upload for the world to download, like these Chinese models. And that's something worth thinking about.
It's something worth watching. Can we fork and continue to develop them? Hey, that's it for this week.
If you enjoyed this episode, please like it, subscribe, share it with someone who you may think that this story is important to, because it isn't just China, as I said. But until next time, guys, I'm Shimmy, and this is what Shimmy said. Shimmy says, Shimmy says, Shimmy, Shimmy, Shimmy says.
Ask me almost anything
