Yep, it seems that the only ways out of this crisis is either China makes significant progress in domestic lithography, or AI bubble in USA collapses, but the second one is much more problematic outcome.
The problem with AI bubble bursting is that it willl result in a major crisis in the industry, and it won't fill the market with cheap hardware, as factories right now contracted to do server side hardware, not suitable for consumers PC's. So we're not only gonna get a lot of bankruptcies and unemployment, but no hardware also for years. The only positive is potential drop in prices for hosting in USA.
Yes. USA just wants to maintain its dominance and sabotage China. ML also solves a lot of military problems and create advantages. So they ride AI hype train to divert attention from the real reasons to sabotage computing globally.
The US does not dominate the memory industry -- it's dominated by two South Korean companies. China's DRAM industry is largely driven by gov't directives and *incentives*.
Sure, the US doesn't "dominate" or even control the lithography industry.
The US gov't however granted more or less exclusive EUV license to ASML eons ago, in exchange for significant domestic manufacturing and R&D in the US. Today, ASML maintains its 2nd largest R&D center in CT; Cymer (light-source/laser), a US subsidiary ASML acquired over a decade, in San Diego, and software R&D in Silicon Valley. Slightly less than 1/3 of ASML's business is actually in the US and that's why the US export control affects ASML's business in China. But this is not the case for Nikon or Canon, ASML's lesser competitors.
Most DRAM are made on 12+nm -- doesn't require EUVs and China has plenty of DUVs.
Sure, the US doesn't "dominate" or even control the lithography industry.
That's not true. ASML was directly prohibited by USA to sell modern lithography to China, and I've seen some reports that spare parts for the older tech is also banned.
ASML was directly prohibited to sell any lithography to Russia since USSR times.
So, on papers it's an independent company in a foreign country, in reality it can't sell its products to anyone USA doesn't like.
The same situation with Japan, that on paper is an independent country, but in reality it's still occupied by USA and have no right to make decisions in multiple sensitive areas.
not true. The US gov't has control over 1/3 of ASML supply-chain in US, which includes key components such as light-source, software, etc. The US has no such control over the Japanese lithography makers -- they don't depend on US suppliers. The US gov't instead has to negotiate the Japanese to issue their own export control. The Japanese competitors were never invited to license EUV from the US gov't in the early 1990's.
ASML was directly prohibited to sell any lithography to Russia since USSR times.
ASML was founded in 1984 and their first commercial product PAS 5500 was released in 1991 -- that's when the Soviet empire collapsed.
Further, in Q4 2025, China was ASML's largest customer with 35% of all systems sales going to China.
As it stands, there's no way AI is becoming cheap enough to run at scale for millions of users, across billions of prompts, without the prices going up at least 10x of what they are currently.
The models themselves are becoming more and more memory hungry as they continue iterating. The approx $20 a consumer pays per month towards a subscription like ChatGPT puts OpenAI in the red with that particular customer before they've rattled off 10 prompts.
Right now, everything is VC subsidized. It's the same deal for companies too. They will not be seeing their hiring costs reduce once these investor subsidies go away. If they're paying $1000 a month for a subscription, they'll end up paying $10,000. If they're paying $100k, it'll become a million.
Unless AI becomes drastically more lightweight (unlikely) or hardware becomes exponentially more powerful and efficient (impossible), the cost of AI is about to go up very soon.
I agree, but there's also an aspect of the military usage. AI is a "double purpose" tech, and significant investments by USA government and military corps are to get an edge in this area.
I'm not under the impression of "AI bad, AI expensive, therefore AI go bust".
I see positives with AI, I see negatives. FWIW, AI is here to stay in some capacity or the other, and what that will look like is only after we get past the current bubble. I definitely agree with you that military and armed forces around the world are rapidly arming themselves with this to get an edge on tomorrow's battlefield.
But I think the strength in military applications lies partly with local AI, wherein the models don't have to phone back to a data center to make real-time tactical decisions on whether to take out a target, move, or do recon and mark the target for some heavier firepower, phone home about new information at a location, etc.
My only prediction is that consumers and companies alike are about to get a nasty awakening once the VC money dries up and companies like OAI and Anthropic raise their prices to show us what it really costs to use these models. Same for the thousands of AI companies that have popped up being just a wrapper around existing models.
But I think the strength in military applications lies partly with local AI, wherein the models don't have to phone back to a data center to make real-time tactical decisions on whether to take out a target, move, or do recon and mark the target for some heavier firepower, phone home about new information at a location, etc.
Look at what they're doing with Starlink. Under the disguise of a civilian usage, it's used all over the globe to communicate with drones, and not only for reconnaissance.
But so far the military usage is in global control over logistics, troop movements etc, that was previously very hard to track as it required too many people to look at satellites data. The more datacenters they have, the better their capacity to track anything around the globe with minimal delays.
My only prediction is that consumers and companies alike are about to get a nasty awakening once the VC money dries up and companies like OAI and Anthropic raise their prices to show us what it really costs to use these models. Same for the thousands of AI companies that have popped up being just a wrapper around existing models.
I agree, unless we have some massive breakthrough in tech performance and/or price in a couple of years, "AI revolution" will shrink to a few companies with very pricey products in limited niches. In most cases humans will still be a much cheaper solution.
RAM isn't just used in techbros AI machines, but across the economy in pretty much every industry. From industrial PLCs, microwave control panels, TVs, mobile phones, microwaves, cars, planes, etc.
The price increases are massively affecting all companies producing any kind of physical product and that won't be fixed with HBM RAM.
RAM suppliers have been bitten by this in the past - spend millions over 3 years to ramp up production, only to find at the end of 3 years that the bubble popped and now the demand is too little to break even.
The smart people in the industry have seen this before, so they aren't going to simply ramp up production, because at the end of all that investment they might just go out of business if the bubble had popped by then.
It’s not like they’ve been found to be price fixing in the past anyways
Price fixing works best when there are few parties; makes collusion easier. When suppliers ramp up demand that doesn't exist in 3 years, they go out of business, reducing the number of suppliers.
Counter-intuitively, we want measured and controlled response to demand, just so that many players are left still alive when the dust settles.
Counter-intuitively, we want measured and controlled response to demand, just so that many players are left still alive when the dust settles.
What everyone wants ironically including the manufacturers themselves, at least in the long term is for RAM manufacturers to tell hyper scalers to go fuck themselves so that when the bubble pops, there will still be anyone left for them to sell to.
This isn't a measured response, this is an understanding that AI will pay anything no matter how insane so why not let them bid up the price to grab stock. It'll fuck the entire economy, but they'll make a lot of money right now.
I am not expecting them to ramp up production to the 10x factor. Since they are probably right that once the factories are finished the demand is gone.
However I would expect them not to lower production once demand rises...
Great, maybe instead of Electron crap, new generations learn proper data structures and algorithms to fit into memory constrained devices, like in the old home computer days, or game consoles.
I wish. But this is lost ground. Vibe coding will just make things worse. Instead of not knowing what their careless use of stuff causes under the hood, new programmers won't even know what they do at all, I believe. Only few will make it to the low level. I know only a few people where I'd think, they'd use bit arithmetic and see an IP(v4) address as an integer...
Are the memory shortage, AI, and infrastructure investments all part of a planned strategy?
Of course, I do think Korean companies are moving very conservatively because of the unfair pricing pressure they faced from big tech in the past. At the same time, however, if China, as a massive manufacturing power, begins producing low-cost memory in earnest, the impact will be severe.
That could become a problem in both the consumer market and the enterprise market.
If they are truly thinking strategically, then by the time China catches up in the memory industry, they may expand capacity and use the price competitiveness they have built up in the meantime to launch another zero-sum game.
Filias9@reddit
For cheaper memory, we are waiting for China. To start their production en mass and break cartel. This or big economical collapse.
PraetorRU@reddit
Yep, it seems that the only ways out of this crisis is either China makes significant progress in domestic lithography, or AI bubble in USA collapses, but the second one is much more problematic outcome.
lazyhustlermusic@reddit
I’m down with the bubble regressing mainstream AI for the next five years or so.
PraetorRU@reddit
The problem with AI bubble bursting is that it willl result in a major crisis in the industry, and it won't fill the market with cheap hardware, as factories right now contracted to do server side hardware, not suitable for consumers PC's. So we're not only gonna get a lot of bankruptcies and unemployment, but no hardware also for years. The only positive is potential drop in prices for hosting in USA.
lazyhustlermusic@reddit
That’s a human generated problem blamed on AI
PraetorRU@reddit
Yes. USA just wants to maintain its dominance and sabotage China. ML also solves a lot of military problems and create advantages. So they ride AI hype train to divert attention from the real reasons to sabotage computing globally.
tooltalk01@reddit
The US does not dominate the memory industry -- it's dominated by two South Korean companies. China's DRAM industry is largely driven by gov't directives and *incentives*.
PraetorRU@reddit
It's pretty much the same as claiming that USA doesn't dominate lithography market as ASML is in Netherlands.
tooltalk01@reddit
Sure, the US doesn't "dominate" or even control the lithography industry.
The US gov't however granted more or less exclusive EUV license to ASML eons ago, in exchange for significant domestic manufacturing and R&D in the US. Today, ASML maintains its 2nd largest R&D center in CT; Cymer (light-source/laser), a US subsidiary ASML acquired over a decade, in San Diego, and software R&D in Silicon Valley. Slightly less than 1/3 of ASML's business is actually in the US and that's why the US export control affects ASML's business in China. But this is not the case for Nikon or Canon, ASML's lesser competitors.
Most DRAM are made on 12+nm -- doesn't require EUVs and China has plenty of DUVs.
PraetorRU@reddit
That's not true. ASML was directly prohibited by USA to sell modern lithography to China, and I've seen some reports that spare parts for the older tech is also banned. ASML was directly prohibited to sell any lithography to Russia since USSR times. So, on papers it's an independent company in a foreign country, in reality it can't sell its products to anyone USA doesn't like.
The same situation with Japan, that on paper is an independent country, but in reality it's still occupied by USA and have no right to make decisions in multiple sensitive areas.
tooltalk01@reddit
not true. The US gov't has control over 1/3 of ASML supply-chain in US, which includes key components such as light-source, software, etc. The US has no such control over the Japanese lithography makers -- they don't depend on US suppliers. The US gov't instead has to negotiate the Japanese to issue their own export control. The Japanese competitors were never invited to license EUV from the US gov't in the early 1990's.
ASML was founded in 1984 and their first commercial product PAS 5500 was released in 1991 -- that's when the Soviet empire collapsed.
Further, in Q4 2025, China was ASML's largest customer with 35% of all systems sales going to China.
lazyhustlermusic@reddit
What’s the best tin foil to make a hat from?
KinTharEl@reddit
As it stands, there's no way AI is becoming cheap enough to run at scale for millions of users, across billions of prompts, without the prices going up at least 10x of what they are currently.
The models themselves are becoming more and more memory hungry as they continue iterating. The approx $20 a consumer pays per month towards a subscription like ChatGPT puts OpenAI in the red with that particular customer before they've rattled off 10 prompts.
Right now, everything is VC subsidized. It's the same deal for companies too. They will not be seeing their hiring costs reduce once these investor subsidies go away. If they're paying $1000 a month for a subscription, they'll end up paying $10,000. If they're paying $100k, it'll become a million.
Unless AI becomes drastically more lightweight (unlikely) or hardware becomes exponentially more powerful and efficient (impossible), the cost of AI is about to go up very soon.
PraetorRU@reddit
I agree, but there's also an aspect of the military usage. AI is a "double purpose" tech, and significant investments by USA government and military corps are to get an edge in this area.
KinTharEl@reddit
I'm not under the impression of "AI bad, AI expensive, therefore AI go bust".
I see positives with AI, I see negatives. FWIW, AI is here to stay in some capacity or the other, and what that will look like is only after we get past the current bubble. I definitely agree with you that military and armed forces around the world are rapidly arming themselves with this to get an edge on tomorrow's battlefield.
But I think the strength in military applications lies partly with local AI, wherein the models don't have to phone back to a data center to make real-time tactical decisions on whether to take out a target, move, or do recon and mark the target for some heavier firepower, phone home about new information at a location, etc.
My only prediction is that consumers and companies alike are about to get a nasty awakening once the VC money dries up and companies like OAI and Anthropic raise their prices to show us what it really costs to use these models. Same for the thousands of AI companies that have popped up being just a wrapper around existing models.
PraetorRU@reddit
Look at what they're doing with Starlink. Under the disguise of a civilian usage, it's used all over the globe to communicate with drones, and not only for reconnaissance.
But so far the military usage is in global control over logistics, troop movements etc, that was previously very hard to track as it required too many people to look at satellites data. The more datacenters they have, the better their capacity to track anything around the globe with minimal delays.
I agree, unless we have some massive breakthrough in tech performance and/or price in a couple of years, "AI revolution" will shrink to a few companies with very pricey products in limited niches. In most cases humans will still be a much cheaper solution.
lazyhustlermusic@reddit
Don't mind them, they just have been drinking Putin's kool aid.
lelanthran@reddit
We don't need that; last-gen die/tech is sufficient, even if slightly slower, because we want more HBM machines, not SOTA RAM.
A high bandwidth 256GB DDR4 machine is going to be better (for LLM usage, but would translate to almost all usage) than current DDR5 machines.
The Chinese don't need to get to 7nm fabrications to meet the demand.
Izacus@reddit
RAM isn't just used in techbros AI machines, but across the economy in pretty much every industry. From industrial PLCs, microwave control panels, TVs, mobile phones, microwaves, cars, planes, etc.
The price increases are massively affecting all companies producing any kind of physical product and that won't be fixed with HBM RAM.
tooltalk01@reddit
If China's "success" in the rare earth metal is any guidance, China would drive out foreign competitors and their start own cartel.
Not quite sure there are still folks who believe China Co. is charity, or that it's a supply issue.
BalleaBlanc@reddit
And prices... decades.
programming-ModTeam@reddit
Article is behind a paywall.
lelanthran@reddit
RAM suppliers have been bitten by this in the past - spend millions over 3 years to ramp up production, only to find at the end of 3 years that the bubble popped and now the demand is too little to break even.
The smart people in the industry have seen this before, so they aren't going to simply ramp up production, because at the end of all that investment they might just go out of business if the bubble had popped by then.
Seik64@reddit
It’s not like they’ve been found to be price fixing in the past anyways
lelanthran@reddit
Price fixing works best when there are few parties; makes collusion easier. When suppliers ramp up demand that doesn't exist in 3 years, they go out of business, reducing the number of suppliers.
Counter-intuitively, we want measured and controlled response to demand, just so that many players are left still alive when the dust settles.
recycled_ideas@reddit
What everyone wants ironically including the manufacturers themselves, at least in the long term is for RAM manufacturers to tell hyper scalers to go fuck themselves so that when the bubble pops, there will still be anyone left for them to sell to.
This isn't a measured response, this is an understanding that AI will pay anything no matter how insane so why not let them bid up the price to grab stock. It'll fuck the entire economy, but they'll make a lot of money right now.
Johanno1@reddit
I am not expecting them to ramp up production to the 10x factor. Since they are probably right that once the factories are finished the demand is gone.
However I would expect them not to lower production once demand rises...
And they are lowering production right now!
Orsim27@reddit
They shift capacities from DDR ram to HBM. Not lowering, just given the capacities to the players with more money
lelanthran@reddit
Are they? Those greedy fucks!!!
pjmlp@reddit
Great, maybe instead of Electron crap, new generations learn proper data structures and algorithms to fit into memory constrained devices, like in the old home computer days, or game consoles.
Ofen@reddit
I wish. But this is lost ground. Vibe coding will just make things worse. Instead of not knowing what their careless use of stuff causes under the hood, new programmers won't even know what they do at all, I believe. Only few will make it to the low level. I know only a few people where I'd think, they'd use bit arithmetic and see an IP(v4) address as an integer...
redimkira@reddit
Necessity is the mother of invention. From that perspective, part of me wishes memory continues to be scarce.
bawng@reddit
I thought it was laziness!
redimkira@reddit
That too haha
lonelyroom-eklaghor@reddit
Now that's a great scenario
Complete_Instance_18@reddit
Yeah, this is rough news, especially for anyone trying to
Internal-Factor-980@reddit
Are the memory shortage, AI, and infrastructure investments all part of a planned strategy?
Of course, I do think Korean companies are moving very conservatively because of the unfair pricing pressure they faced from big tech in the past. At the same time, however, if China, as a massive manufacturing power, begins producing low-cost memory in earnest, the impact will be severe.
That could become a problem in both the consumer market and the enterprise market.
If they are truly thinking strategically, then by the time China catches up in the memory industry, they may expand capacity and use the price competitiveness they have built up in the meantime to launch another zero-sum game.
Wouldn’t that be the most likely scenario?