… is the ram used in data centers even compatible with desktop PC. I hear a lot “cheap ram when bubble bursts”
You must log in or # to comment.
The resources they’re exhausting aren’t the physical components, it’s the production capacity.
The shit being fed into this ‘AI’ grift will have very limited reuse potential when their universe implodes.
What we’re seeing is a confluence of unconstrained parasitic consumption, and artificially constrained cartel production.
The parallels I see are with OPEC in 1973. JEDEC have found themselves in a position of power at the choke point of a Crucial industry, and are leveraging it to maximise profit and influence.
Unions of individuals are good for society, cartels of corporations are carcinogenic.
If/when the bubble bursts we’ll have bigger problems than the price of RAM.
If it bursts the world wide economy collapses, because most “wealth” is loans against stock, which are then invested in stocks driving the price up.
I think we’re past a trillion sunk into the ponzi scheme just in AI stock, but if it goes down banks call in their loans triggering automatic sales of whatever collateral they used.
Billions and billions being sold automatically regardless of price would cause cascading crashes…
But if it works…
Corps can fire the majority of their employees and starving desperate people turn to Mad Max after a few consecutive missed meals.
If it bursts the
world widewestern economy collapses
*when
Haven’t seen this mentioned yet: the RAM they are making for these AI chips are built to burn up after 6 years. The RAM is built into the processing units to reduce latency but the whole unit is designed with short planned obsolescence so burning out after 5-6 years is expected. The assumption is that by the time they burn up they you will replace them with a more energy efficient unit.
How do you know it’s going to burn out in 6 years?
It’s an accounting thing. They can deduct the value of the hardware over its expected usable lifespan. So they have to replace it after 4-6 years even if it’s still working fine. Sometimes they sell it used, and you can buy it pretty cheap, but dollar for dollar, consumer grade stuff that’s current will usually consume less power and game better than a 6 year old server. Now if you need a home NAS or VM host…
Trump also changed the depreciation law as part of the BBB, so I think companies pay less tax by buying this year. https://www.bdo.com/insights/tax/one-big-beautiful-bill-act-expands-100-depreciation-expensing-opportunities
It could be different but that is what the companies are stating as their operational life span when they report their financials. They used to say they would only last 4 years…
No, in addition to the reasons brought up by the other commenters, I’m starting to think that “computers as a service” will start to be a thing.
Google’s Stadia by all accounts wasn’t horrible, but it was pricey and the selection was subpar.
But what if Amazon, Azure, and Google start up some post-AI burst equivalent that provides a use case for all that processing power? Sure, the GPUs used by commercial AI aren’t designed for gaming, but Nvidia could see the writing on the wall and start partnering with hyper scalers to create massive racks of gaming GPUs. And it would be one step closer to the ultimate goal of removing personal ownership of things! Pay a subscription for a cloud gaming PC or try your luck on building your own.
People will subscribe to computers too… :) Haha this world is beyond silly now.
NZXT has this now. And why people went ballistic on them last year. You pretty much “rent” a PC but the price comes out to be nearly triple what you’d pay for a premium PC over the course of a few years.
Thats really weird. I guess you can sell anything to people who are afraid of owning something, by just telling them service is included or something.
That scam works well for AWS. My last company was a very small software dev (120 people) and spent $6M a year on AWS. Can you imagine the infra you could build with that much money!? And still have plenty left over to pay a small team to be responsible for it.
Yeah I know. I also worked for a small company and they spent a fortune on aws, mainly just for virtual machines and network traffic between regions.
None of my colleagues ever run anything on aws themselves for their own homelab stuff, since its extreamly overpriced.
Yeah the flak was justified though. They had all these rules like it couldn’t be"used" in any way, so people weren’t able to upgrade as often as they promised. I believe Gamer Nexus did a video about it.
The RAM that has been sold will not be viable for desktop systems, but especially with manufacturing capacity build up, you’d have memory vendors a bit more desperate to find a target market for new product. Datacenter clients will still exist but they could actually subsist on the hypothetical leftovers of a failed buildout, so consumer space may be their best bet.
I’m assuming RAM factories are investing in new production lines to meet the increased demand. This means there should be ample capacity to spare when demand decreases. That’s the ideal time to purchase those 500 GB of RAM for all your Chrome tabs.
The plan is that when interest rates rise again, investors won’t have access to cheap loans and AI companies won’t be able to build more data centres. This should lead to a demand collapse and manufacturers being left with surplus capacity. I doubt they can reduce production quickly enough, so they’ll likely push cheap products onto the market for a while.
It is not compatible with desktop PC. It is a different manufactured kind of ram. The increase price is from companies flooding in high amounts of large orders. So consumer ram for PCs is getting backlogged so to say. Or just slowed down
Specifically, desktop RAM is slabs of silicon, placed into little packages, soldered onto circuit boards in DIMM form or similar, to be plugged into a motherboard slot for RAM.
The AI demand is for the silicon itself, using advanced packaging techniques to put them on the same package as the complex GPUs with very high bandwidth. So these same pieces of silicon are not even being put into DIMMs, so that if they fall out of use they’ll be pretty much intertwined with chips in form factors that a consumer can’t easily make use of.
There’s not really an easy way to bring that memory back into the consumer market, even after the AI bubble bursts.
No. Its ecc ram and most PC motherboards aren’t compatible with it. If it crashes we might have cheap GPUs and cheap cloud gaming.
Those of us running home labs will have a field day upgrading our own servers.
Unfortunately not even then. Nowadays the GPUs are a pretty alien form factor, usually not pcie cards. SXM and now HGX.
Datacenter gear has resembled consumer systems less and less after a period of getting closer in the 90s and 2000s.
Not only ecc, but registered. Most AMD systems would be compatible with ecc u-dimm, but not with the r-dimm found in servers
The datacentre GPUs are also useless for consumers. They don’t have video output.
That’s not much of an issue these days, you can offload rendering pretty much wherever in Linux. Level1Techs has a whole section on their forums on SR-IOV and working with offloaded compute in homelabs for awhile now.
Probably not someone that a consumer would want to do unless they’re interested in the challenge.
I forgot about the absence of PCIe connectors until jj4211 mentioned it in another comment. That would also be problematic for consumer use. I know it’s possible to do the conversion as I’ve looked into it in the past, but it doesn’t come cheap.
I’ll make my own video output! With blackjack, and hookers!
Why would companies lower their prices after consumers have been forced to get used to them?
In order to maximize revenue. Selling 1,000 units at $50 profit apiece makes you more money than selling 400 units at $100 profit each.
But if you fire half your staff while only making 400 units…
Now you’re thinking like a capitalist!
Sure, but what if you guranteed your customers through loosely organized collusion and endless consolidation/buyouts that there is no other practical place to buy ram?
That works during a bubble. What happens though after the bubble when demand presumably plummets?
I wouldn’t count on it anytime soon. The data center build out is a years long process, and even after they get built, their owners will want to use the initially installed hardware for as long as possible before upgrading.
Expect demand for new RAM to remain higher than it was before the AI datacenter build out began. Maybe not as high as it is right now, but some degree of increased demand will be permanent due to the increased count of datacenters that will eventually want to replace their obsolete, worn out hardware.
There might eventually be a larger supply of used RAM on the market, but a lot of it might not be compatible with mid-shelf consumer grade motherboards.
If it bursts we’ll get a lot of data centers looking for something to do. And RAM prices will come down but … What’s causing the bottle neck right now is that your average RAM factory needs a couple of years from designed to built to working. So the supply is limited right now while the demand is high. While us end users can’t use the data center gear, in a pinch they could use ours. So the bottle neck gets tighter. So if the bubble collapses, supply will increase and that will bring prices down. If it bursts about two years from now, all the hastily built RAM factories will churn out cheap RAM. But none of this is guaranteed, not the busting and not the dead cheap prices. Because the demand for RAM will not drop off a cliff, it will most likely decrease slowly. All this processing power in post-burst idle data centers will find a way to be used - with what I do not know. There will still be a higher demand for RAM compared to pre-ChatGPT times. So RAM will not flood the market, we will just return to a relative equilibrium of the market.
The only reason why memory might get repurposed is because the demand for AI has collapsed. The bubble that’s most likely to burst is a financial bubble, and is unlikely to reduce the demand for AI from users… So I wouldn’t count on it.
