Try building a high wall around software and see what happens. It doesn't work.
For two years, Washington and Beijing have tried to slice the global artificial intelligence market down the middle. American policymakers passed sweeping export controls to keep Nvidia silicon out of Chinese data centers. Beijing responded by urging domestic tech giants like Alibaba and ByteDance to dump American tech and subsidize domestic alternatives.
The narrative is clean, dramatic, and compelling: a new Silicon Curtain dividing humanity’s most important technology into two distinct, isolated halves.
It’s also wrong.
The idea that you can build digital fortresses around AI misjudges how software, open-weight models, and global hardware supply chains actually function. Despite billions spent on bans, tariffs, and subsidies, the wall isn't holding. It's leaking like a sieve.
The Myth of the Hard Divide
If export bans actually created a total wall, Chinese AI labs would be stuck in the stone age. They aren't.
Recent capability tests by global security bodies show Chinese open-weight models narrowing the performance gap with top-tier American frontier systems to just a few months. Models like Z.ai’s GLM series or Moonshot’s Kimi line match or exceed American benchmarks on complex coding and reasoning tasks while operating at a fraction of the cost.
How? Because software isn't raw material. You can't blockade an algorithm the way you blockade oil or steel.
When American companies publish research papers, Chinese researchers read them. When open-weight parameters hit the web, developers worldwide download them, tweak them, and distill them into leaner, faster models. Attempting to isolate AI technology ignores the fundamental nature of open-source software development.
Smuggling and Cloud Loop-Holes
Then there's the hardware problem.
The U.S. Bureau of Industry and Security set performance limits on GPUs to block high-end chip shipments to China. But money finds a way.
- Third-party intermediaries: Shell companies in Southeast Asian hubs bought tens of thousands of restricted chips before regulatory loopholes closed.
- Cloud access: Chinese developers simply rent compute time on servers located in non-restricted regions.
- Domestic acceleration: Banning top-tier American chips didn't starve Chinese tech—it forced Beijing to pour over $150 billion into domestic hardware giants like Huawei and SMIC.
China's domestic chip self-sufficiency jumped from 16% to nearly 28% in a single year. By trying to lock China out, American policy accidentally accelerated the birth of a direct hardware rival.
+-------------------------------------------------------+
| THE CHIP PARADOX |
| |
| US Export Bans ---> Forces China to Invest |
| | |
| v |
| US Loses Market <--- SMIC & Huawei Gains |
| Share Instantly Domestic Market Share |
+-------------------------------------------------------+
Beijing Is Building Its Own Fence
Ironically, the biggest threat to open AI access in China isn't coming from Washington. It's coming from Beijing.
After watching open-weight Chinese models spread globally, Chinese regulators began floating restrictions to limit overseas downloads of domestic model weights and training datasets. Officials worry that letting Western companies freely use Chinese models gives American businesses cheap, powerful tools without giving China anything in return.
This creates a bizarre stalemate. Washington tries to keep hardware in. Beijing tries to keep model weights and data locked down.
Yet both efforts run into the exact same problem: engineers don't care about geopolitics. They care about performance per dollar. When a model works, developers find a way to run it, mirror it, and ship it.
Why Complete Separation Will Fail
You can't partition the AI ecosystem without destroying the feedback loops that make the technology useful in the first place.
Middle-power nations across Europe, Asia, and Latin America refuse to pick a side. They want cheap Chinese inference models running on efficient hardware, but they also want access to top-tier American foundation models. If the U.S. or China demands absolute allegiance, enterprises in these regions simply route around the restriction.
When you look past the political grandstanding, three realities emerge:
- Inference is becoming a commodity. Running models is getting cheaper everywhere, rendering hardware choke points less effective over time.
- Distillation works. Smaller, cheaper models can learn from massive, expensive models, making proprietary advantages short-lived.
- Open source wins on distribution. Developers prefer models they can inspect, modify, and host locally without getting locked into a vendor's ecosystem.
The attempt to build national walls around artificial intelligence makes great headlines for politicians looking tough on trade. But in practice, it just creates a black market for compute and pushes innovation into the shadows.
If you want to stay ahead in AI, focus on building faster, cheaper, and more reliable systems. Relying on government regulations to keep the competition out is a losing strategy. Software always breaks free.
How to Navigate the Fragmented AI Market
If you're building software or managing infrastructure today, relying on a single national tech stack is a massive risk. Here is how to protect your operations:
- Decouple your model layer: Never hardcode your applications to a single AI provider or proprietary API. Use abstraction frameworks that let you swap between U.S. closed models and open-weight alternatives in minutes.
- Prioritize local inference: Where data privacy and security matter, run open-weight models on your own cloud or on-premise hardware.
- Audit your hardware supply chains: If you rely on specialized hardware, diversify your cloud providers to ensure vendor changes or export policy updates don't take your products offline overnight.
