~ Companies are now ceasing their mandate to workers to use AI to do things that used to be done without AI. They’re now saying don’t use AI just to use AI. They’re trying to reduce their token purchases.
~ The estimation now for AI use cases is to reconsider if tokens cost ten times their current price. If the use case is still worth it, then that use case will likely survive the bubble. If it’s not worth it, it’s time to hire back employees.
~ The hyperscale AI industry will have to make $6 trillion annually to break even. Amazon makes $2 trillion by selling people material stuff. Walmart is similar. There may not be a market for $6 trillion in tokens every year, even from government projects. Also, it’s a bad sign if government projects are propping up the whole stock market.
~ China is mostly turning to a software AI model which does home and small business AI tasks fairly well without buying compute from a hyperscaler. You get a gaming machine, get open source AI software and a dozen terabytes of training data, and you should be able to create slop, vibe code or fix the grammar and style of your report. Also, hyperscale models are opinionated and don’t like certain topics. Home-grown AI doesn’t have those objections.
I think for small to medium sized businesses, those who are using cloud providers already, they will continue to purchase ai compute just like they do any other cloud service. For larger enterprises that have the manpower to maintain physical infra, it will be cheaper to self host inference.
For regular users, I suspect most people will use a provider, but are not nearly as willing to pay for it (at least not that much more than a typical streaming service). Power users (especially younger ones with more time in their hands to spare) will use local inference.
How much of Amazons $2T in earnings is AWS earnings? If that already includes AWS spend, then $6T isn’t gonna happen any time soon.
So I saw a few relevant newspieces:
~ Companies are now ceasing their mandate to workers to use AI to do things that used to be done without AI. They’re now saying don’t use AI just to use AI. They’re trying to reduce their token purchases.
~ The estimation now for AI use cases is to reconsider if tokens cost ten times their current price. If the use case is still worth it, then that use case will likely survive the bubble. If it’s not worth it, it’s time to hire back employees.
~ The hyperscale AI industry will have to make $6 trillion annually to break even. Amazon makes $2 trillion by selling people material stuff. Walmart is similar. There may not be a market for $6 trillion in tokens every year, even from government projects. Also, it’s a bad sign if government projects are propping up the whole stock market.
~ China is mostly turning to a software AI model which does home and small business AI tasks fairly well without buying compute from a hyperscaler. You get a gaming machine, get open source AI software and a dozen terabytes of training data, and you should be able to create slop, vibe code or fix the grammar and style of your report. Also, hyperscale models are opinionated and don’t like certain topics. Home-grown AI doesn’t have those objections.
I think for small to medium sized businesses, those who are using cloud providers already, they will continue to purchase ai compute just like they do any other cloud service. For larger enterprises that have the manpower to maintain physical infra, it will be cheaper to self host inference.
For regular users, I suspect most people will use a provider, but are not nearly as willing to pay for it (at least not that much more than a typical streaming service). Power users (especially younger ones with more time in their hands to spare) will use local inference.
How much of Amazons $2T in earnings is AWS earnings? If that already includes AWS spend, then $6T isn’t gonna happen any time soon.
Some companies already run Chairman Xi’s big models on-prem and do not limit token usage on those.