I don't use Lichess. If I did, I would have donated. But after knowing that they throw money down the sink like this, I would be hesitant.
Maybe you could rewrite it in assembler over the weekend and run it on a Raspberry Pi, right?
I worked on Java codebases that absolutely squeezed every bit of performance from a machine and that would be very hard to do in other languages.
While I know many people are saying don't look at the cloud costs, when I look at those numbers it makes me think of CAPEX vs OPEX trades, and the OPEX here is non-trivial, but potentially a variable offset by some labor costs. I think some of this is being driven by a dated tech stack and technical debt.
Now, I am not one to trivialize the work that may need to go into getting this efficient, but I think you are right, we should consider investing into modernization of the stack.