In June 2026, the United States put a lock on its own AI models. It ordered Anthropic to block access to Mythos 5 and Fable 5 for anyone who isn't American, and days later it asked OpenAI to limit the rollout of GPT 5.6 to a handful of government-approved partners. Its best models, the newest ones, were placed out of reach for the rest of the world.
That adds to something older: years of restrictions that already left Nvidia almost out of China's chip market. Two brakes, one new and one accumulated, with the same effect. And the effect isn't the one Washington was after: demand, money and talent ran toward Chinese AI.
I live this firsthand. I'm a founder in Chile, I can't use the newest from Anthropic or OpenAI, and my stack leans more and more on Chinese models like GLM and DeepSeek. This isn't ideology. It's what's left when they close the door on you.
What exactly did the United States block?
Two different brakes that ended up coinciding.
The first, the models, and it's from this month. In mid-June 2026, the US government ordered Anthropic to block access by non-Americans to Mythos 5 and Fable 5, its most powerful models. The official reason was national security, after suspicion that a China-linked group had accessed one of them. Anthropic suspended them worldwide to comply, and the Commerce Department lifted the restriction only for a short list of American companies.
Days later it was OpenAI's turn. The administration asked it to limit the rollout of GPT 5.6 to a small number of partners approved by the government itself, over its cybersecurity capabilities, which it considers equivalent to Mythos's. Sam Altman called it a "strange moment," with no clear rules, and accepted the condition to be able to launch the model. So it wasn't an isolated case: the two US leaders, Anthropic and OpenAI, ended up with their best technology under government control. When the owners of the frontier have to ask permission to hand out what they build, the rest of the world looks elsewhere.
The second brake, the hardware, isn't from June: it's a policy that's been tightening for years. The US has restricted the sale of Nvidia's AI chips to China since 2022, turn after turn. The state of things in 2026 is the most extreme, with an ironic twist: early in the year the government loosened up to let the H200 in, but by then China had already told its own companies not to buy the H20 or the H200, on security grounds. Nvidia ended up practically out of that market, squeezed from both shores. An analysis by Brookings puts it bluntly: the game's over, the US is out of the AI chip market in China.
The world runs to Chinese AI
When you take the best Western model away from the world, the world doesn't sit and wait. It moves to the next one. And the next one, today, is Chinese and open.
The clearest signal: shares of Z.ai, the company behind GLM, jumped more than 30% after it released its new open source model. Global demand shifted toward the Chinese alternatives, cheaper and with no access restrictions. It's no accident that right now half the internet talks about GLM 5.2 as if it were free: the block opened the door for it.
And it's not just that they're available. They're much cheaper. DeepSeek had already shown you can build a top-tier model for a fraction of what American labs spend, and that gap carries over to the price per use. For a founder counting every dollar, an open, accessible and cheap model stopped being plan B. It's the plan A that Washington made obvious.
I'm part of that statistic. As a founder outside the United States, I'm not on the list that can use the newest from Anthropic. So I test, I measure, and I use what I do have available, and a lot of it comes from China. In my AI model benchmark, DeepSeek and the Chinese models fight for the top spots. I don't pick them by flag. By access and by result.
And China stopped needing Nvidia
Here's the part that should worry Washington more than any other.
Blocking the chip didn't stop China. It pushed it to make its own. Analysts estimate that the export cap alone could increase China's domestic compute capacity by 250%. Huawei plans to ship 600,000 of its Ascend 910C chips in 2026, no matter what the US decides.
And it's no longer theory. Huawei leads the AI chip market inside China, where it beat Nvidia, and it's projected to control 60% of that market by the end of 2026, with revenue over 12 billion dollars. Its Ascend 950 series is comparable to the H200, one of Nvidia's most powerful chips. And DeepSeek's V4 model has already been adapted to run on Huawei's chips: a full stack, a Chinese model on Chinese hardware, with no stop in the United States.
The honest caveat: the individual chip is still behind. The Ascend 910C performs at around 60% of an H100, and China still depends on manufacturing equipment it doesn't fully produce at home. But the direction is clear, and so is the speed. When someone with resources has no other choice, they build. And they improve fast.
Why this is an own goal
The logic of the block is intuitive: take the best technology from your rival and you leave them behind. It works when the rival can't build the substitute. China can.
Every restriction did the opposite of what it promised. The model block sent the world to try the Chinese ones. The chip block handed Huawei a captive market of hundreds of thousands of units and a national reason to accelerate. The United States didn't stop China. It funded its independence.
It's an old lesson, and it's not geopolitics, it's business: when you close the door on someone with resources and pride, you don't stop them. You force them to build their own house. And sometimes they build it better than yours.
What this means for you, founder
Let's bring this down to earth, because what matters to you is what you build with.
First: your stack can't depend on a single provider. If you depend on one model and tomorrow a political decision cuts your access, you're left without operations. Mine is split across several models on purpose, and that stopped being a technical preference to become insurance.
Second: Chinese AI is part of the map, like it or not. GLM and DeepSeek are among the most capable and cheapest models out there today. Ignoring them by flag is competing with one hand tied.
Third, and that's why it's not blind faith: Chinese models come with their own fine print. It matters where your data is processed, what's stored and what's censored. I use them with judgment, for the right tasks, not for everything. Just like with any provider, the question isn't what country it's from, but what it does with what you give it.
The day the best model and the best chip come from the same place
The United States wanted to buy time and ended up giving away ground. For the founder on the outside, the takeaway is simple and a little uncomfortable: the future of your stack is no longer decided by Silicon Valley alone.
Test the Chinese models with the same coldness you'd test any other: by capability, by cost, by where your data goes. And don't take your eyes off the hardware. Today the Chinese chip is behind; in a year or two, that sentence may stop being true. The day the best model and the best chip come from the same place, and that place isn't the United States, the block will have achieved exactly the opposite of what it set out to do.
I write about AI and test models at cristiantala.com. Every month I publish a benchmark of 89 models.
Top comments (0)