Trade bans on AI models are the fastest way to build your own competitors. When you wall off a massive chunk of the global market, you aren’t protecting intellectual property or maintaining a strategic edge; you are simply creating a vacuum that someone else will happily fill with a clone. We have seen this play out in a dozen different industries over the last century, and the result is always the same: the banned party stops trying to buy the product and starts trying to build it. It is a basic law of economics that the more you restrict a highly desired resource, the more you incentivize the world to figure out how to make it without you.
The recent surge of “Mythos-like” models coming out of Asian startups is the inevitable result of Anthropic’s ongoing export restrictions. As detailed in this TechCrunch report, these companies aren’t necessarily trying to invent a new way of thinking—they are reverse-engineering the behavior and capabilities of a model they can no longer legally access. It is a bit like the automotive industry in the mid-century, where brands spent decades essentially copying the lines of whoever was winning the luxury segment because they couldn’t get the original blueprints. They aren’t innovating; they are distilling. They are using synthetic data generated by the original models to train their own smaller, leaner versions that mimic the original’s reasoning patterns. The goal isn’t to be original; the goal is to be “close enough” to the gold standard to be useful for the average developer.
But let’s be real about the technical reality here. The hurdle isn’t the architecture—anyone with a decent team can scrape a paper or prompt-engineer a distillation set to mimic the output style of a top-tier model (probably because the lawyers were bored and thought a ban would actually work). The real friction is the hardware. You can copy the weights or the training recipe all you want, but you still need the H100s or B200s to make it breathe. These startups are likely fighting a brutal war over GPU clusters and power grids just to get these clones off the ground. Who actually thinks a regulatory fence stops a well-funded team from routing traffic through three different shell companies in neutral territories just to get a few thousand chips? The cost of these GPUs is astronomical, and the latency of routing requests through proxies is a nightmare, but it is a price they are willing to pay.
Or maybe the clones are just slightly worse versions of the original—see below. It is possible that these models are just “skin-deep” imitations that fail the moment you push them into complex, multi-step reasoning tasks (though I suspect the gap is narrower than Anthropic wants to admit). But in a market where the alternative is having nothing at all, a “B-” grade model that is locally available beats an “A+” model that is blocked by a firewall. The psychological shift is what matters here. Once a region moves from being a customer to being a developer, they never go back to being just a customer. They stop asking “how do I get access to Mythos?” and start asking “how do I make this better than Mythos?”
Anthropic is playing a dangerous game of geopolitical chess where they think they’ve trapped the king, but they’ve actually just given their opponents a reason to stop paying for the subscription. By restricting access, they’ve effectively provided a roadmap and a guaranteed customer base for every startup in the region. They’ve traded a recurring revenue stream for a temporary sense of security, all while teaching their competitors exactly what the market wants. It is a strategic failure of the highest order. They are essentially funding the R&D of their future rivals by making the cost of failure for those rivals almost zero. If you can’t buy the best, you build the best, and the market demand ensures that the build will eventually succeed.
The momentum is already shifting. By Q4 2026, we will see at least two of these “clone” models outperform the restricted versions of Mythos on regional benchmarks. Once the local talent stops trying to “mimic” and starts optimizing for their own languages and cultural contexts, the gap will close faster than any export ban can keep up with. We are moving from the era of the “global model” to the era of “sovereign AI,” and the irony is that the US-based labs are the ones who pushed the world in that direction.
The ban didn’t stop the progress; it just changed who gets the check.
