You probably mean they are operating with 80% margins discounting training expenses, which will continue to be pretty high for the foreseeable future.

Seeing how cheaply Xiaomi was able to train Mimo 2.6 I am starting to wonder if they are greatly over-exaggerating the real training costs to increase their valuations and investments.

that was rl, not training. its a fraction of the cost

Don't let facts and knowledge get in the way of a good conspiracy theory!

Source?

Not a source but a comparison with a weaker non-SOTA model:

Nvidia's top AI chip Rubin sells in 72-GPU racks for about $3.5–7.8M. A rack running Xiaomi's MiMo V2.6 Pro generates roughly 150–300B tokens a day, worth about $130–260k at Xiaomi's API price. That's a payback of the infrastructure in a few weeks in theory. After a few weeks or a month, the only cost is electricity, and whatever they make after that is pure profit.

OpenAI and Anthropic are practically scamming people with the token prices.