I think you can see this clearly in financial markets, but also in academic fields like AI reaearch. A researcher at an average institution just can't compete with the kinds of compute their equivalents at Facebook/Google/OpenAI can call upon.
How long until this imbalance of compute resources becomes a major cause of resentment?
For instance if you are hungry or homeless is Sheryl Sandberg going to say "Let them eat Bitcoins?"
People who are "learning A.I." are going to see the cost of training as significant, but they're usually working with data from Kaggle or other public data such that (1) they don't see the cost of labeling and curating data, and (2) they are making a product w/o commercial value (the data set isn't meant to do that)
Commercially relevant A.I. by definition has to do inference many more times than it does training and at a scale much larger than the labeling (otherwise you could just get the labeler to do the task more economically than inference.)
It's backwards to how outsiders think, but if A.I. is going to change the world the cost of inference has to drop, not the cost of training; that's what A.I. needs to deliver value at a lower cost than training.
---
One of two personal projects I have now now is about delivering the maximum value from inkjet printing; some of that is communicating that my print with a material cost of 6 cents (+4 cents for the removable 3M double-sided tape to stick it to the wall) is a public affordance and I want you to "steal" it or otherwise take it off the wall so you can enjoy both sides of it, otherwise it's a failure.
People love to talk endlessly about how $5 M NFT(s) are going to change the world but with a $25 materials budget and relentless attention to quality and cost control I can transform your little piece piece of it.
I might sound like a subartistic nerd but I see attention to the little things and continuous progress as unlocking technology, not some alternation between worshiping and fearing "Big Tech".
https://en.wikipedia.org/wiki/Computer_Lib/Dream_Machines
In a very real way, supercomputing is more accessible price-wise than at any time in history (~$10K NVidia A100). And next gen satellite internet should plug in the next 4B users.
The bridge is compressed education at scale. It still requires 10k+ hours to become an AI researcher. The next revolution looks like the interactive learning environments prophesied by the OG hypermedia pioneers. Even if it means minting a new breed of Github Copilot "prompt engineers" who never delve low-level!