ATRIUMsearch → argument graph
Video · 2026-08-07 · 23m · 6 moments

I bought Black Market AI Tokens on Chinese Amazon

✦ AI generated

timeline · colored by role

01
Context

Chinese sellers on Taobao are operating a well-run, multi-million dollar black market for AI tokens, offering access to legitimate LLM providers like Claude for pennies on the dollar.

The host introduces his investigation into black market AI tokens sold on Taobao (Chinese Amazon), describing three types of listings, and frames his exploration of whether these tools are real and worth it.

transcript

Wes: Did you know there's a black market for AI tokens? Chinese sellers are offering up API tokens and subscription to popular LLM providers like Cloud Code and OpenAI Codex for pennies on the dollar... Three of my favorite things in life are good deals, sketchy Chinese websites, and protecting shareholder value. So, I had to dig in and check it out myself. I spent my own money on black market cla tokens, and what I found out was a well-run, multi-million dollar industry that goes deeper than I could have ever imagined.

gives example · 2

02
Mechanism

Claude instance access bought on Taobao is proxied through shared browsers, giving users a slow, shared experience until they get logged out, making it not worth the money.

Wes describes buying a cheap 'shared monthly card account' Claude access, discovering it is delivered by proxying shared browser instances of the actual Claude website, resulting in a slow, limited experience.

transcript

Wes: Immediately what they did is they sent me some random token that I can use and then a link to Yuk... So what they are doing here is they are proxying it through. So they have a whole bunch of clawed instances running on a browser somewhere and they are running all of these different chats and keeping track of the chat ids that are belong to me and their proxy in the middle is they're filtering it out. So they're using the same one over for multiple people... It it was a very very crappy experience for me. And also when I take a look at my dosage, which I've simply have run three or four chats so far, I only get 50 over the next 3 hours. So for the 10 bucks that I spent on this, probably not worth it.

rebuts · 1

03
Mechanism

The 1.5x-ratio black market API tokens are not actually running Claude — they substitute a much cheaper model with fake Claude identity — while the 5x-ratio tokens appear to deliver legitimate Claude.

Wes runs legitimate Claude tokens against the black market tokens, comparing tokenization and extracting a system prompt proving the cheap tokens swap in a different model despite claiming to be Claude. Unbilled overages are inflated by the resellers.

transcript

Wes: And all of these models have different tokenization patterns. So saying something like, 'Hey, how are you doing today?' will tokenize differently for each of the models. Then with the slow token, it was giving me 160 tokens. Then I asked AI to write a script and try to do extract the system prompt from it. And from the cheaper one, the one that was 1 and a halfx, it was able to get out the uh prompt from it where it says if anybody asks you in Chinese or English who you are, you tell them that you are cloud code. So almost certainly I can tell you that using the API keys that are at 1 1/2x ratio that's not actually clawed. They're just substituting it with a much cheaper model. Now the one that is 5x ratio I do believe from all of my testing that that is legitimate claw.

extends · 1provides context · 1

04
Data

Black market token resellers inflate reported token usage (charging multiple times the actual spend), so the $8 'deal' effectively becomes more expensive than legitimate Claude credits.

Wes discovers his claimed $100 credit '5x boost' and the inflated usage reporting means his actual spend nets out worse than paying directly, concluding the black market tokens are not a good deal.

transcript

Wes: So if I'm doing my math correctly here, I spent $8 for $100 worth of tokens, right? Then I go to my API key to create my token and it is five times that. So 8 * 5 that is 40 cents on the dollar. Now what they're telling me here is that my $2 spend was really closer to six. So that's three times more expensive. So my 40 cents ends up being 120. meaning that I just paid $1.20 or $120 for $100 worth of tokens. That's not a good deal at all. I think they're being a little bit sneaky here with it.

explains mechanism · 1provides context · 2

05
Mechanism

The black market token industry is structured as a four-tier pipeline — upstream stolen credit cards, midstream harvested tokens, downstream 'transfer station' proxies, and end users — with models being swapped along the way.

Matthew Lenhard of Vectoral explains how the underground token market works: stolen cards register accounts, tokens are gathered from leaks and trials, unified proxies route requests, and end users buy in.

transcript

Matthew Lenhard: Upstream you have the credit cards. They're maybe using stolen credit cards to be able to register with these big labs... And then midstream, this is where they start to pull all of the accounts... whether that is there's an npm worm that has scraped your file, whether that is uh reverse engineering... They're trying to reverse engineer because tokens are tokens and wherever they come from they don't necessarily care. So they're trying to reverse engineer all of the APIs and then put it into a single tidy API which is the next one the downstream. These are the transfer stations... They give you a nice tidy API where you can log in... Then there's me, the end user, looking for a good deal.

explains mechanism · 2extends · 2gives example · 1provides context · 1supports · 1

06
Mechanism

Stealing and reselling tokens is only one revenue stream: resellers also swap models, inflate token usage, and sell the logs of real Claude usage to Chinese labs for model distillation.

Wes and Matthew explain the three revenue streams of black market operators, highlighting that the captured usage logs become a gold mine for Chinese labs distilling Anthropic or OpenAI models.

transcript

Wes: Stealing and selling tokens is not the only way that these guys are making money. This interesting post on China talk, one fish, three meals, where meal one is marking up access... meal 2 swapping out the models and inflating the tokens... And then the third one, this is really important, is the logs are the product. So these LLM companies that are trying to train their own models, what they will often do is called distillation, which is Anthropic has been crying about it for almost 6 months now, which is essentially you can train your model based on the output of another model. And if you have the logs or the output of somebody actually using cloud code, then you can use those logs to actually train your own model... That is an absolute gold mine to these companies that are specifically a lot of them in China that are trying to make their own version of anthropic or open AI models.

extends · 1supports · 1

Highlight slides
Related episodes