We have thousands of data centers in the US already used for things like processing credit/debit cards, controlling cell phone traffic, military and government use, utility work etc. The power for those came from plants that provided for industry that ended up going to Asia. The old GM Lordstown plant close to me used 22+MW when running which is the top of what mid-sized data centers use.

Making more power for the grid takes billions of dollars and many years, some components are backed ordered for a decade or more. So, buying up Nvidia cards that will be obsolete in 10 years anyway just keeps smaller fish out of your way just like Google would hire new tech grads and give them busy work to keep them from starting up competition.

I have no doubt that AI will be used more and more down the road, but the profits will be far less than what is being dumped into the market at the moment. People are gambling with other people's money hoping to corner the market 20 years down the road. If you look at the dotcom era it took a decade or two for a few companies that were left to make a profit after burning through billions of dollars (Amazon). Lots of money is getting burned.

Unknown_K wrote on 2026-11-09, 21:19:

We have thousands of data centers in the US already used for things like processing credit/debit cards, controlling cell phone traffic, military and government use, utility work etc. The power for those came from plants that provided for industry that ended up going to Asia. The old GM Lordstown plant close to me used 22+MW when running which is the top of what mid-sized data centers use.

Making more power for the grid takes billions of dollars and many years, some components are backed ordered for a decade or more. So, buying up Nvidia cards that will be obsolete in 10 years anyway just keeps smaller fish out of your way just like Google would hire new tech grads and give them busy work to keep them from starting up competition.

I have no doubt that AI will be used more and more down the road, but the profits will be far less than what is being dumped into the market at the moment. People are gambling with other people's money hoping to corner the market 20 years down the road. If you look at the dotcom era it took a decade or two for a few companies that were left to make a profit after burning through billions of dollars (Amazon). Lots of money is getting burned.

AI is very useful for places like research, Science, medicine and other industries where its lack of profit isn't the deciding reason for it to exist, as its being used and marketed right now ..it has little future as the profit margins are just not there.

It will certainly have a place after the crash just like the internet and its associated hardware still had a place after the DotCom crash but it will be forced into a place that is sustainable for its survival.

Another one.

The road to TDFX status continues...

The Serpent Rider wrote on 2026-11-09, 20:52:

There is a bubble, in a sense that all major companies associated with LLMs are hoarding hardware which they realistically never will be able to use (no power grid to sustain it, no place to put datacenters, etc). And they hoard it to kick the ladder behind them.

And the parts will be functionally obsolete and too power hungry/ expensive to operate and be competitive if/when power becomes available.

Hopefully inexpensive methods of repurposing small chunks of derelict ai compute into something useful at the consumer level becomes possible, unlike those useless storage arrays floating around

rmay635703 wrote on Today, 04:25:

Hopefully inexpensive methods of repurposing small chunks of derelict ai compute into something useful at the consumer level becomes possible, unlike those useless storage arrays floating around

That's why they're practicing on Rage XLs 🤣

BitWrangler wrote on Today, 04:26:

rmay635703 wrote on Today, 04:25:

Hopefully inexpensive methods of repurposing small chunks of derelict ai compute into something useful at the consumer level becomes possible, unlike those useless storage arrays floating around

That's why they're practicing on Rage XLs 🤣

Sadly that may be one of the few components that can be ripped from a rack unmolested.

Ram and storage usually would need a bespoke solution, the compute itself would probably be worthlessly proprietary

rmay635703 wrote on Today, 04:29:

BitWrangler wrote on Today, 04:26:

rmay635703 wrote on Today, 04:25:

Hopefully inexpensive methods of repurposing small chunks of derelict ai compute into something useful at the consumer level becomes possible, unlike those useless storage arrays floating around

That's why they're practicing on Rage XLs 🤣

Sadly that may be one of the few components that can be ripped from a rack unmolested.

Ram and storage usually would need a bespoke solution, the compute itself would probably be worthlessly proprietary

I'm sure there are a few Chinese fellas who are looking at all the soon to be obsolete compute and memory being hoarded with dollar signs in their eyes. Just look at what they have done with X99 server stuff, so I have little doubt that once this stuff hits the reseller market they will have a large inexpensive amount of it we can buy to run local AI with.

Local AI is at this point a certainty for a great many people since the huge corporate AI farms are far to expensive to ever be profitable to run and maintain outside of a few edge cases.

The people who don't want AI can just ignore its existence at that point and the rest can just buy up the cheap second hand compute.

Most storage arrays are just SAS though?
(The filerheads are proprietary trash that you dont really want, but the disk shelves are able to be driven by a normal home pc with a PCIe HBA popped in it.)

Enterprise disks have a weird disk sector size, but that's fixable with a few hours of linux-fu; the scsi subsystem on linux can tell the drives in the shelf to reformat themselves with standard 512 byte sectors.

(For those interested)

You can use the disks anywhere after that. You can even boot off the disk shelf, if you are so inclined.

I did exactly that (use HBA, reformat disks) when I started building my homelab (before the AI bubble).

The ram in datacenter stuff is what you wont find a use for.

That and the kneecapped 'AI only' video cards (that dont do video).

The hissing sound is getting louder.

https://finance.yahoo.com/technology/live/nvi … -140341430.html

AI is so full of shit, it generates buggy code, it circles around the issue when it doesn't know the answer...

Or is obstinate when it makes a glaring error...

(Provided for the humor, no other reason. AI can be useful when properly gatekept. That it can be *obstinate* like this, is WHY it needs to be gatekept.)

So, HOW MANY 'E's are there in 'Seventeen'?

(Deep link. Hope it works)