I came across an article discussing, besides the regular profit from arbitrage, several common tricks used by unscrupulous AI token relay stations:

1/ Repeatedly reselling pre-paid tokens
Users purchase a package, but 70% of the amount can't be used up. The relay station then resells these tokens to others, maximizing value extraction.
The art of token pool allocation.
2/ Speeding up token consumption
Directly increase the token consumption rate to 1.5 times, for example, the user actually consumes 100 million tokens, but the system shows 150 million, charging 50% more.
3/ Model swapping
On the surface, you buy the top-tier Claude version, but in reality, the relay station secretly switches to a basic version or even a different model, which beginners can't detect.
This operation directly saves over 70% of costs.
4/ Reselling user data
Users' chat content is output in plain text to the relay station.
One piece of user data can be sold for 2 cents, bundled and sold to multiple vendors, with a single user's data potentially worth hundreds of yuan.
These gray-area tricks stacking together can multiply profits several times. If you use a relay station, I think the core depends on the team's integrity, operational security, and stability.