Are stagnating models really a problem? Because if so, then how is "if we stop training we are profitable" an acceptable answer to the profitability question?
Groq don't train their own models, they serve open ones. They're a few versions behind the frontier open weights models. The top model they serve is GPT OSS 120B, released last summer.
They used to serve Kimi K2 (I think they said experimentally), which I think for many purposes is a better model, but discontinued it with the suggestion to use GPT OSS instead.