The model-swapping section is the most interesting part. The CISPA finding — 37% vs 83.82% on benchmarks — is stark. But what's striking is that the market corrects around it: the proxies that actually route to Claude are more expensive, and there's a segment of Chinese developers who specifically seek those out. Not just cost arbitrage; they need the capability gap.
The deeper irony is that tighter access control may not eliminate demand. It may simply professionalize the gray market around that demand.
If Chinese developers, researchers, and companies still want Claude, but the official channel is blocked or unstable, the market will not disappear. It will reroute. The result is not necessarily less access, but less visible access: more intermediaries, weaker accountability, more data leakage, and a larger incentive to monetize logs.
This is why the problem is bigger than China or Claude. AI access is becoming a political economy of channels. Whoever controls the model matters. But whoever controls the routing layer, payment layer, identity layer, and user logs may matter almost as much.
Blocked by their own government through the Great Firewall (GWF), unable to freely access the international internet;
Restricted by American AI companies and related policies, unable to normally use frontier models like Claude and ChatGPT;
Burdened by their own national enterprises — companies like DeepSeek, GLM, MiniMax, and Alibaba are industrially distilling American large models, which in turn objectively intensifies the level of control and restrictions;
Fooled and held hostage by their own country’s large models — on one hand, all domestic models have built-in political censorship mechanisms, and some frontier models are even restricted for purchase within China (for example, GLM5.2 is deliberately made scarce on domestic platforms while being readily available on international ones); on the other hand, some Chinese AI companies, in pursuit of international reputation, divert resources belonging to their own people (with a per capita GDP of around $13,000) — resources that enjoy government tax breaks, infrastructure subsidies, resource allocation preferences, and market favoritism — to subsidize overseas consumers (with a per capita GDP of around $80,000), all just to gain “false international praise”;
Betrayed by their own compatriots — Zilan Qian wrote this article possibly only for money, reputation, or even out of a sense of superiority or morbid curiosity! Yet the article shows no reflection or concern whatsoever for ordinary Chinese people’s right to use humanity’s frontier AI.
In summary, a puzzling question arises:
As members of humanity, should 1.4 billion Chinese people have the right to use frontier AI products?
Another unavoidable question follows:Are all Chinese people (one-seventh of humanity) equivalent to the Chinese Communist Party?
Your primary reporting is terrific and essential to thoughtful conversation on these issues. Thank you. However I strongly disagree with you that knowledge exchange and open knowledge makes a threatening world. The facts are the opposite, see Thomas Shelling Nobel prize for “the logic of conflict in nuclear scenarios”: withholding and secrecy leads to knowledge mostly being used to maintain dominance, and the result is multiple actors competing for dominance. This is an unstable game.
Moreover your reporting shows that the tactical steps Anthropic and the US industry are taking are not in fact working— they are not slowing diffusion. I would suggest that the solution for achieving a better world is not doing more of the same.
Instead, try to focus on what can be done and by whom. And accepting what cannot be done because a single point of dominance of global society is not plausible in a world of multiple highly capable societies.
1. In model safety, this is the responsibility of the model makers. Use of model for bad end: this is an Anthropic problem no matter where the user comes from or what they pay. A full price full blooded US user can have bad intentions.
2. Reselling and price arbitrage is inherent in the deepest human idea of markets: customers “buy” something because they believe they will “own” it. And have freedom of action to use what they buy. Meaning they are free to imagine a use for it that will benefit them, and that you as the seller may not have thought of. Owning means that you as an owner are free to create new uses and fully enjoy the product.
A key part of this is the right to resell. Resale creates opportunities for buying and selling—arbitrage—what hedge funds facilitate—and liquidity in markets-that is the ability to efficiently buy and sell. Most of the US restrictions on technology are based on limiting ownership rights and weakening the definition of ownership and controlling or preventing resale—while at the same time promoting markets.
The US approach, most strongly associated with the Biden administration, is perhaps not exactly wrong headed, but it denies centuries of experience with people and markets. Check out behavioral economics for more.
3. The world we live in is at its foundation one of open knowledge exchange. This is the great advantage of US society- it is the most open to knowledge and initiative- it was open source, anti-empire from the beginning. The attempt to lock knowledge up and keep it from non-US people is failing on all sides because among other things it is counter to the values of most of the tech community. Ai success in San Francisco and Toronto and China are connected.
Preventing the diffusion of knowledge beyond the US borders will not succeed beyond certain unusual situations such as lithography. This strategy becomes more and more difficult when global education levels surpass those of the US and large portions of primary knowledge development accomplished by global communities. Hard as this may sound, if we want global safety and advancement we need to find common purpose with others.
The V2EX angle here adds texture. V2EX (China's primary developer forum, ~1M monthly actives) has daily threads on Claude Code workflows, API proxy recommendations, and transfer station quality reviews. The participant profile matches exactly what you describe: SaaS builders, CS students, indie devs. Not researchers trying to catch up. The meta-commentary is visible in real-time, users tracking which proxies have reliable uptime, which ones silently rate-limit, which break when Anthropic tightens verification.
The Singapore routing is in the data too. SG consistently recommended on tech forums as the lowest-latency, most stable path. The 'Singapore leads per capita Claude use' stat is not accidental. There's a whole infrastructure of Alibaba Cloud SG instances behind that number.
Russia has similar marketplaces advertising access to LLMs that residents of that country wouldn't be able to access due to sanctions (see: plati.market). Do you think that this configuration could be how Russian sellers are providing access to buyers?
Another day another Zilan banger
Pretty telling traveling in China how they’re all Claude pilled
Impressive to write an article that starts off disturbing and then gets continually worse as you keep reading.
wonderful piece - thank you zilan; i learned a lot
This is so interesting. I learn so much from you
Fascinating.
The model-swapping section is the most interesting part. The CISPA finding — 37% vs 83.82% on benchmarks — is stark. But what's striking is that the market corrects around it: the proxies that actually route to Claude are more expensive, and there's a segment of Chinese developers who specifically seek those out. Not just cost arbitrage; they need the capability gap.
The deeper irony is that tighter access control may not eliminate demand. It may simply professionalize the gray market around that demand.
If Chinese developers, researchers, and companies still want Claude, but the official channel is blocked or unstable, the market will not disappear. It will reroute. The result is not necessarily less access, but less visible access: more intermediaries, weaker accountability, more data leakage, and a larger incentive to monetize logs.
This is why the problem is bigger than China or Claude. AI access is becoming a political economy of channels. Whoever controls the model matters. But whoever controls the routing layer, payment layer, identity layer, and user logs may matter almost as much.
AI era — the daily life of an Chinese person:
Blocked by their own government through the Great Firewall (GWF), unable to freely access the international internet;
Restricted by American AI companies and related policies, unable to normally use frontier models like Claude and ChatGPT;
Burdened by their own national enterprises — companies like DeepSeek, GLM, MiniMax, and Alibaba are industrially distilling American large models, which in turn objectively intensifies the level of control and restrictions;
Fooled and held hostage by their own country’s large models — on one hand, all domestic models have built-in political censorship mechanisms, and some frontier models are even restricted for purchase within China (for example, GLM5.2 is deliberately made scarce on domestic platforms while being readily available on international ones); on the other hand, some Chinese AI companies, in pursuit of international reputation, divert resources belonging to their own people (with a per capita GDP of around $13,000) — resources that enjoy government tax breaks, infrastructure subsidies, resource allocation preferences, and market favoritism — to subsidize overseas consumers (with a per capita GDP of around $80,000), all just to gain “false international praise”;
Betrayed by their own compatriots — Zilan Qian wrote this article possibly only for money, reputation, or even out of a sense of superiority or morbid curiosity! Yet the article shows no reflection or concern whatsoever for ordinary Chinese people’s right to use humanity’s frontier AI.
In summary, a puzzling question arises:
As members of humanity, should 1.4 billion Chinese people have the right to use frontier AI products?
Another unavoidable question follows:Are all Chinese people (one-seventh of humanity) equivalent to the Chinese Communist Party?
Anthropic was not first to install KYC. OpenAI did this with ImageGPT 1.
Qian please don't call them distillation attacks T_T do me a favor please
Wait hold on. That's fucking crazy that the white house is on this too. Isn't the whole anthropic claim about deepseek distilling just bullshit?
https://www.youtube.com/watch?v=_k22WAEAfpE
Anthropic staff will be so happy to see this post.
Your primary reporting is terrific and essential to thoughtful conversation on these issues. Thank you. However I strongly disagree with you that knowledge exchange and open knowledge makes a threatening world. The facts are the opposite, see Thomas Shelling Nobel prize for “the logic of conflict in nuclear scenarios”: withholding and secrecy leads to knowledge mostly being used to maintain dominance, and the result is multiple actors competing for dominance. This is an unstable game.
Moreover your reporting shows that the tactical steps Anthropic and the US industry are taking are not in fact working— they are not slowing diffusion. I would suggest that the solution for achieving a better world is not doing more of the same.
Instead, try to focus on what can be done and by whom. And accepting what cannot be done because a single point of dominance of global society is not plausible in a world of multiple highly capable societies.
1. In model safety, this is the responsibility of the model makers. Use of model for bad end: this is an Anthropic problem no matter where the user comes from or what they pay. A full price full blooded US user can have bad intentions.
2. Reselling and price arbitrage is inherent in the deepest human idea of markets: customers “buy” something because they believe they will “own” it. And have freedom of action to use what they buy. Meaning they are free to imagine a use for it that will benefit them, and that you as the seller may not have thought of. Owning means that you as an owner are free to create new uses and fully enjoy the product.
A key part of this is the right to resell. Resale creates opportunities for buying and selling—arbitrage—what hedge funds facilitate—and liquidity in markets-that is the ability to efficiently buy and sell. Most of the US restrictions on technology are based on limiting ownership rights and weakening the definition of ownership and controlling or preventing resale—while at the same time promoting markets.
The US approach, most strongly associated with the Biden administration, is perhaps not exactly wrong headed, but it denies centuries of experience with people and markets. Check out behavioral economics for more.
3. The world we live in is at its foundation one of open knowledge exchange. This is the great advantage of US society- it is the most open to knowledge and initiative- it was open source, anti-empire from the beginning. The attempt to lock knowledge up and keep it from non-US people is failing on all sides because among other things it is counter to the values of most of the tech community. Ai success in San Francisco and Toronto and China are connected.
Preventing the diffusion of knowledge beyond the US borders will not succeed beyond certain unusual situations such as lithography. This strategy becomes more and more difficult when global education levels surpass those of the US and large portions of primary knowledge development accomplished by global communities. Hard as this may sound, if we want global safety and advancement we need to find common purpose with others.
The V2EX angle here adds texture. V2EX (China's primary developer forum, ~1M monthly actives) has daily threads on Claude Code workflows, API proxy recommendations, and transfer station quality reviews. The participant profile matches exactly what you describe: SaaS builders, CS students, indie devs. Not researchers trying to catch up. The meta-commentary is visible in real-time, users tracking which proxies have reliable uptime, which ones silently rate-limit, which break when Anthropic tightens verification.
The Singapore routing is in the data too. SG consistently recommended on tech forums as the lowest-latency, most stable path. The 'Singapore leads per capita Claude use' stat is not accidental. There's a whole infrastructure of Alibaba Cloud SG instances behind that number.
Russia has similar marketplaces advertising access to LLMs that residents of that country wouldn't be able to access due to sanctions (see: plati.market). Do you think that this configuration could be how Russian sellers are providing access to buyers?