The question lands with a certain honesty that most tech prompts avoid: if a stack of high-end GPUs appeared on your desk tomorrow, what would you actually build? Not what would be impressive on a résumé, not what would generate the most impressive demo reel, but what would genuinely pull you forward? Banning local LLMs from the conversation is the right kind of constraint. It forces a return to first principles, to the question of what raw compute is actually for when you strip away the most obvious, most exhausted answer.
The thread's suggestions drift toward the expected corners: niche simulations, distributed systems, generative art that isn't text. But the more interesting responses push past the familiar into research-oriented territory, which is where the real signal lives. This is not about finding a use for spare horsepower. It's about recognizing that distributed training and inference are themselves a discipline worth exploring, and that the algorithms governing how models learn across nodes are a far richer problem than any single inference task. The person who asks what else GPUs can do is really asking what we've been missing by focusing so narrowly on one application.
That instinct to look past the obvious is worth taking seriously. The same mindset that treats a GPU stack as a playground for exploring how LLMs navigate token space is the mindset that produces genuinely new ideas. It's not about the hardware, and it never was. It's about the willingness to treat a known tool as an unknown variable. When you remove the default answer, you're left with a much more uncomfortable and much more useful question: what do you actually want to understand well enough to compute?
For our readers, the practical takeaway is not to rush out and acquire a pile of GPUs. It's to notice where your own thinking has become routine. If you caught yourself defaulting to "run a local model" when asked about raw compute, that's not a failure. It's a signal that you've stopped asking what else is possible. The most valuable work in any field comes from the questions that don't have a canned response yet. The thread's best contributions are the ones that treat the GPUs as a means to a specific, weird, personal end, not as a status symbol.
The concrete point to watch is whether this kind of speculative question translates into actual projects. It's easy to brainstorm unhinged ideas in a comment thread. It's harder to commit to one, build it, and share what you learned. The people who do that, who turn a what-if into a working prototype, are the ones who will define what the next generation of compute is for. So the question isn't really what you'd build with a stack of GPUs. It's what you're curious enough about to actually start building today.