If you check the actual tokens consumption and compare it to the API you will find a factor 10.

Training models cost tens of millions, their revenues from sub + api are well above hundreds of millions.

If you look at minimax IPO data, you can see that they spent 3x their revenue on "cloud bills".

So yes, it's probable that they do subsidize inference through subscriptions in order to capture market.

No inferrence provider is profitable, and most run on VC money to serve customers.