Video games are the key example: all video games are competing against other games for what comes down to more complicated physics calculations (often light-based physics: "more realistic" shadows from Raytracing, or "cooler" shaders like Dragonball FighterZ / Guilty Gear (very "unrealistic", but clearly requires modern GPU-shaders to calculate).
Though Google tried to get video games "into the cloud", it still required data-centers full of high-end GPUs... and even then had latency issues.
Other performance critical applications include Stockfish Chess, LeelaZero Go (and KataGo), Blender 3d modeling, etc. etc. Having the user "spend the money on more compute" is economically a more feasible move, than centralizing compute costs to a server somewhere (especially costly GPU-hosting).
--------
Good news! There's plenty of performance-critical applications waiting to be explored.
But if you just need to deploy a simple application with low compute costs, centralizing the compute into a single server (or even mild decentralization through the Javascript interpreter) is good enough.