There is currently a lot of turmoil throughout the proxy and relay station industry regarding Claude's new KYC regulations. I have worked in this business for about two months, and last week, when this news broke, everything changed.

Initially, I thought the impact wouldn't be significant because Claude's KYC is still triggered randomly. But the very next day, everything shifted. Now, the total cost of the account pool has increased, and channel stability is also affected. Some people are buying new numbers, some are looking for KYC providers, but the real problem is that this could spread to all users in the future.

80% of my clients are Indian, the rest are from Europe, Latin America, and South Asia. Interestingly, no one is from the US. All these people have high-end requirements for Claude's performance—mostly developers, research teams, and B2B companies. In the beginning, some casual users also came when OpenClaude became popular, but they gradually disappeared because the costs were too high.

The entry barrier in this proxy relay station business is very low—just a couple of customers are enough to start. But competition is extremely fierce. Some people directly advertise to their customer groups, some attack websites. I saw the strangest case where the original price was 1.8, with a discount it dropped to 0.8, and the seller claimed it was genuine. When the customer noticed issues in the evening, a dispute ensued, and the seller disappeared.

At the beginning of April, when Claude adjusted its limits, the entire supply became very compressed. At that time, I had to push my technical team to create various channels, and also add some reliable external channels. This experience taught me a counterintuitive lesson: if the model controls risk across the entire network and your pool dries up, customers tend to behave sensibly. But if your pool isn't dry and you're committing fraud, you'll get caught. So if you've been heavily penalized and have no credibility, maybe you're genuine.

The biggest challenge in this business is finding trustworthy channels. The best quality is Claude services provided by AWS, followed by officially reverse-engineered Claude services, and then industry-reputed channels like certain platforms. If no one can tell where it originated from, put a question mark.

Different channels affect the output. Output from AWS and official channels is very clean. It's also important to pay attention to cache hit rates. If you talk to the same model three times and it can recall previous information, that’s a high hit rate that saves you money. You can also test official features like native web search and multimodal image understanding.

My biggest concern is for third-world customers—especially from countries like India and Iran. For them, Claude's price is very heavy relative to their monthly salaries. The price difference between Sonnet and Opus can be 2-3x, so they naturally turn to more cost-efficient channels. There was an Iranian customer who was blocked due to war, but he desperately needed smart AI. Talking to him made me realize that war is not so far from me.

We thought we would focus on relatively developed countries, but in reality, many third-world customers are also facing these difficulties and want to use AI. So they have to turn to relay stations. No one knows what Anthropic's true intention is—whether they want to stop middlemen, or distillation attack labs, or limit users in certain regions. Maybe all of the above. But I have mixed feelings about this. On one hand, it closes many paths for general users, and those who truly need Claude only turn to intermediaries. On the other hand, finding reliable channels is very difficult for general users, and the market is chaotic due to intermediaries selling fake goods.

I think this is not a good situation. Those who can sustain a Claude subscription require high stability. Still, they need to figure out if there's any fraud involved, which can lead to disillusionment with intermediaries.

Where will this relay station business go? I believe it will continue as long as domestic or foreign customers need cheap APIs. If American models remain strong, relay stations will persist because they will always be expensive. The dollar exchange rate remains the same—training the same model costs $1000 in the US, while in China it might be just 1000 yuan. That price difference will always exist.

But if Chinese models truly emerge and become strong enough to beat American models, there will be no game for relay stations because acceptance of Chinese models will be higher. This means that, in the long run, American models must stay at the top for this business to survive.

According to the fundamentals of this business, entry barriers are low—anyone can enter—but the upper limit is very high. It’s like a "mutual success" platform— the more the platform restricts, the less intermediaries disappear.

But overall, I am disappointed with AI. I thought that when the AI era arrives, tired physical work, basic intellectual tasks, and dangerous jobs would be solved by machine intelligence, leaving people free to focus on the spiritual world and humanity. But I realized that even if many services are regionally priced—like a 60% discount on Claude in Africa—this discount is still too expensive for local people.

In these regions, those who use AI are like flying, and those who don’t are like walking. This speed gap will only widen the divide between poverty and prosperity, concentrating more wealth among the richer. Open-source models are advancing, but only a few can be deployed on personal devices.

What I want to do long-term is to take some open-source models or relatively affordable Chinese models abroad—directly to Latin America, less developed European countries, and then South Asia and Southeast Asia. In AI, logically, it should be available to everyone. This is a very serious issue. As a business, I will continue my relay station work, but outside of this, I will lean toward actual open-source projects. That’s what I see and can do.