Yea, AFAIK the problem is that manufacturers don't want to increase production. It's expensive, takes years to build new factories and last time they did it prices crashed and they got stuck holding the bag. So they'd rather ride the current high.
Our only "hope" is that the AI bubble pops. The fact that this is a real possibility is another reason why manufacturers don't want to invest in more factories.
Once it stops being subsidized, and prices go up 30-50%, there will definitely be a correction. It's also a race to the bottom - frontier models will continue to improve but "good enough" has already been hit, and at that point companies will need to compete on price, and they can't go too much lower in pricing without losing money.
That said they ARE building new fabs, that shit just takes years.
It’s also largely good enough now that you can do coding (which is a big driver of usage) on a local model. It’s slower, but so what. Give it a prompt and go make a cup of coffee, vs give it a prompt and instantly get your code. Still faster than doing it yourself.
This guy codes, and as soon as I can run a decent model with 8 gpu (the max a commercial motherboard can take), I am 100% getting some uncensored model and using it for my main LLM, MAYBE with the occasion claude request ($20 per month) for ultra-complex stuff). Bye bye chatgpt and gemini, I wont miss you.
I haven’t played with open router yet but apparently that’s the killer feature: use a local model most of the time, automatically switch to a paid one when needed.
The main issue is growing silicon crystal. It basically takes 10 years to be profitable. Chinese companies are trying to make it, and should get a 30-40% "yield" by... 2030.
They're trying to make it. Thus far they're failing. They have rather inferior seed material. That's one natural resource of the United States the quislings haven't sold yet.
Far as I know they're not expecting any usable yield from their current batch.
I think in 2030 they should be able to hit 40%. Their first real factory won't be there until 2029 at the earliest. By 2035 they will be on par with Micron/etc.
That and they're expecting a functional unification with Taiwan, and the acquisition thereby of TSMC.
Plus since TI was stupid enough to build a fab there a while ago and train up all Chinese personnel, they're not lacking in expertise worth at least one fab.
China has DRAM domestic production and their stuff is okay. Like other DRAM manufacturers their profits shot past the Moon since 2025.
CXMT is the biggest of them. As far as costumer products are concerned, their only limit is that they can't make 32GB RAM DDR5 sticks.
Their density caps at 20GB RAM per stick. Nobody cares if they can flood the foreing market at 2025 prices ( which they can't at the moment ).
The USA is, of course, constantly trying to keep CXMT from developping. The US government is banned from doing buisness with them or buying their products, and they were added to the list of "military-linked" China companies ( as if there was a single Chinese company that isn't indirectly funding China's military ).
Regardless, if they can supply the world minus the USA, this still increases overall supply everywhere.
Something of concern is that CXMT is opening a new high-bandwith-memory fab, so all that production goes straight to AI datacenters.
Not helped by valve timing their hardware reveals around black friday last year screwing up peoples plans and now this shit.
How many price hikes have we seen with hardware in the last decade or so.There was the HDD factory or ram factory hit by earthquakes, the bitcoin miner bullshit jacking up GPU prices,covid 19 and now A.I cartel price fixing.
eBay is going to be flooded with used enterprise hardware in the near future as companies offload last generation servers. If you're into homelabbing you know that you can pick up a beefy server with a half terabyte of memory for the cost of a mediocre consumer PC with 64GB of ram. Throw your 5090 in a Dell R740 and game on.
They're going to make it illegal to resell. Nvidia will have some kind of buyback or mandatory recycling program because the hardware is worth more to them destroyed. That's why they want to put data centers in orbit, so the hardware never comes back and is simply consumed.
It other news...games developers are looking into tightening their code, not relying on heavy asset management swaps and wondering how they're going to buy the GI Joe with the Kung Fu Grip for their kids this Christmas.
Unfortunately you cannot optimize much for ram. If you want 4k textures, you're looking at gigabytes, periods. And if you want the game to have loading time, it NEEDS to be stored in the RAM and you NEED fast ram. Nothing to be done here, unless you use compression algorithms, but it won't look as good...
The 15 years between Batman: Arkham Asylum and Suicide Squad: Kill the Justice League called and would like to discuss this further with you :)
The reliance on massive RAM/VRAM is often a band-aid for inefficient asset streaming and bloated code, not a requirement for visual fidelity. Virtual texturing and texture streaming have existed for years to solve this. Larian Studios uses these techniques to run massive worlds without dumping everything into RAM.
Furthermore, APIs like DirectStorage exist specifically to bypass these bottlenecks by streaming assets directly from SSD to GPU. The fact that many AAA devs still aren't using them effectively proves this is a choice of convenience, not a technical impossibility.
But this is solely graphics and detail on imaginary objects we're talking here. The entire idea of new and original has gone the way of the dodo. I'd rather play Tetris than another clone of whatever is popular today. If people want to develop games they should know what a game is first, stuffing an engine full of assets which have contemporary fashion value (And nothing else) will have all of these games in landfills next to E.T. before 2050.
The only walls I want to climb are these prison walls of banality which pass for acceptable by a community which hasn't seen how hard the technology can be pushed in its current form. Lazy devs, bean counters making decisions and nothing of the proper gaming eras which brought down the boundaries which tried to enclose it.
Have fun with the yellow paint, goes with the yellow journalism that praises it. I'm looking for something stiffer.
I find the older games to look better with modern 4K rendering anyway. They used to use lighting so much more creatively and now it's at best you can get an outdoor game where they use the sun. I've played a ton of such games lately. Fear, Half Life 2, SWAT 4, Dead Space. Some of those are a bit rigid looking, but the atmosphere is outstanding and they don't look in the realm of ~20 years old. Go a bit newer, MGS V looks absolutely outstanding at 4K with the highest texture setting. I'm playing all this stuff on a 65" OLED, so it's not just hiding detail.
Yeah maybe these new games all have "4K" textures, but their game sucks, and most aren't even nice to look at. Nintendo gets it, and yet they continue to sell games easily. None of the game urinalists bitch about those not being the highest quality 4K textures.
Larian Studios is not using AAA 4k textures all over the place. Black Flag Resynced takes 65GB, the original (which already had high-res textures) was 30GB. Where does the other 35GB come from? Did they really write less efficient code? Not really. Just higher def textures.
SSD has what, 5% of the speed of RAM? And if you'Re copying from the SSD to the GPU, you're not copying from the SSD to whatever else you need (including ram).
I know for a fact 125% of current RAM production is sold under 2031. That is, 125% of what is being produced right now is already sold. And we're not even talking about the new ram that is coming out/has started coming out this year.
That is, if Micron and Sy Henix are not able to increase their production by 25%, they are in deep trouble. Fortunately, they are more or less likely to double their production by 2028 at the latest.
While I do think a lot of this corrects pretty aggressively, I just think it will be years and it gets worse before it gets better. I'm considering locking up another new GPU as VRAM is becoming king and while I'm generally happy I don't want to wait five years either.
I do think there may be a time when used SSDs are a glut, because I suspect this whole AI datacenter thing to blow up a bit, a few places building them will go bankrupt and be liquidated, etc. But we will see. The problem with their GPU capacity is it's not things that can be used on the secondary market for graphics, but a ton of that becoming available secondhand would at least make local AI more feasible for the masses.
Yea, AFAIK the problem is that manufacturers don't want to increase production. It's expensive, takes years to build new factories and last time they did it prices crashed and they got stuck holding the bag. So they'd rather ride the current high.
Our only "hope" is that the AI bubble pops. The fact that this is a real possibility is another reason why manufacturers don't want to invest in more factories.
Once it stops being subsidized, and prices go up 30-50%, there will definitely be a correction. It's also a race to the bottom - frontier models will continue to improve but "good enough" has already been hit, and at that point companies will need to compete on price, and they can't go too much lower in pricing without losing money.
That said they ARE building new fabs, that shit just takes years.
I can tell you for a fact that the $200 ChatGPT plan can end up costing them $17,000............................
FUCK GAMING, STOCKPILE AMMO!
Are you lost? You don't belong here.
It’s also largely good enough now that you can do coding (which is a big driver of usage) on a local model. It’s slower, but so what. Give it a prompt and go make a cup of coffee, vs give it a prompt and instantly get your code. Still faster than doing it yourself.
This guy codes, and as soon as I can run a decent model with 8 gpu (the max a commercial motherboard can take), I am 100% getting some uncensored model and using it for my main LLM, MAYBE with the occasion claude request ($20 per month) for ultra-complex stuff). Bye bye chatgpt and gemini, I wont miss you.
I haven’t played with open router yet but apparently that’s the killer feature: use a local model most of the time, automatically switch to a paid one when needed.
ive not checked in in a while, what are the best local models right now?
The main issue is growing silicon crystal. It basically takes 10 years to be profitable. Chinese companies are trying to make it, and should get a 30-40% "yield" by... 2030.
By then the AI craze will/should be dead.
They're trying to make it. Thus far they're failing. They have rather inferior seed material. That's one natural resource of the United States the quislings haven't sold yet.
Far as I know they're not expecting any usable yield from their current batch.
I think in 2030 they should be able to hit 40%. Their first real factory won't be there until 2029 at the earliest. By 2035 they will be on par with Micron/etc.
That and they're expecting a functional unification with Taiwan, and the acquisition thereby of TSMC.
Plus since TI was stupid enough to build a fab there a while ago and train up all Chinese personnel, they're not lacking in expertise worth at least one fab.
Tsmc makes chips not memory, at least not the one ai needs
They don't make the final product RAM cards. They make the components thereof. At least a large amount of them.
Even high channel DDR5?
China has DRAM domestic production and their stuff is okay. Like other DRAM manufacturers their profits shot past the Moon since 2025.
CXMT is the biggest of them. As far as costumer products are concerned, their only limit is that they can't make 32GB RAM DDR5 sticks.
Their density caps at 20GB RAM per stick. Nobody cares if they can flood the foreing market at 2025 prices ( which they can't at the moment ).
The USA is, of course, constantly trying to keep CXMT from developping. The US government is banned from doing buisness with them or buying their products, and they were added to the list of "military-linked" China companies ( as if there was a single Chinese company that isn't indirectly funding China's military ).
Regardless, if they can supply the world minus the USA, this still increases overall supply everywhere.
Something of concern is that CXMT is opening a new high-bandwith-memory fab, so all that production goes straight to AI datacenters.
Not aicmemroy tho..
You are wrong. Every major company with a production fab is ramping up right now.
Welcome to the ammo market
The US will have to have a different presidency in order for the AI bubble to pop. It’s being propped up almost completely by the government.
Not helped by valve timing their hardware reveals around black friday last year screwing up peoples plans and now this shit.
How many price hikes have we seen with hardware in the last decade or so.There was the HDD factory or ram factory hit by earthquakes, the bitcoin miner bullshit jacking up GPU prices,covid 19 and now A.I cartel price fixing.
eBay is going to be flooded with used enterprise hardware in the near future as companies offload last generation servers. If you're into homelabbing you know that you can pick up a beefy server with a half terabyte of memory for the cost of a mediocre consumer PC with 64GB of ram. Throw your 5090 in a Dell R740 and game on.
They're going to make it illegal to resell. Nvidia will have some kind of buyback or mandatory recycling program because the hardware is worth more to them destroyed. That's why they want to put data centers in orbit, so the hardware never comes back and is simply consumed.
LOL How are they going to power 10 megawatt datacenters from space?
Indians in hamster wheels
16 square kilometers of solar panels: https://i.imgur.com/HM0ul3Q.jpeg
Look, I don't think this is going to happen but these are the ideas being tossed around startups and at nvidia. This is where the thinking is.
It other news...games developers are looking into tightening their code, not relying on heavy asset management swaps and wondering how they're going to buy the GI Joe with the Kung Fu Grip for their kids this Christmas.
Unfortunately you cannot optimize much for ram. If you want 4k textures, you're looking at gigabytes, periods. And if you want the game to have loading time, it NEEDS to be stored in the RAM and you NEED fast ram. Nothing to be done here, unless you use compression algorithms, but it won't look as good...
The 15 years between Batman: Arkham Asylum and Suicide Squad: Kill the Justice League called and would like to discuss this further with you :)
The reliance on massive RAM/VRAM is often a band-aid for inefficient asset streaming and bloated code, not a requirement for visual fidelity. Virtual texturing and texture streaming have existed for years to solve this. Larian Studios uses these techniques to run massive worlds without dumping everything into RAM.
Furthermore, APIs like DirectStorage exist specifically to bypass these bottlenecks by streaming assets directly from SSD to GPU. The fact that many AAA devs still aren't using them effectively proves this is a choice of convenience, not a technical impossibility.
But this is solely graphics and detail on imaginary objects we're talking here. The entire idea of new and original has gone the way of the dodo. I'd rather play Tetris than another clone of whatever is popular today. If people want to develop games they should know what a game is first, stuffing an engine full of assets which have contemporary fashion value (And nothing else) will have all of these games in landfills next to E.T. before 2050.
The only walls I want to climb are these prison walls of banality which pass for acceptable by a community which hasn't seen how hard the technology can be pushed in its current form. Lazy devs, bean counters making decisions and nothing of the proper gaming eras which brought down the boundaries which tried to enclose it.
Have fun with the yellow paint, goes with the yellow journalism that praises it. I'm looking for something stiffer.
I find the older games to look better with modern 4K rendering anyway. They used to use lighting so much more creatively and now it's at best you can get an outdoor game where they use the sun. I've played a ton of such games lately. Fear, Half Life 2, SWAT 4, Dead Space. Some of those are a bit rigid looking, but the atmosphere is outstanding and they don't look in the realm of ~20 years old. Go a bit newer, MGS V looks absolutely outstanding at 4K with the highest texture setting. I'm playing all this stuff on a 65" OLED, so it's not just hiding detail.
Yeah maybe these new games all have "4K" textures, but their game sucks, and most aren't even nice to look at. Nintendo gets it, and yet they continue to sell games easily. None of the game urinalists bitch about those not being the highest quality 4K textures.
Larian Studios is not using AAA 4k textures all over the place. Black Flag Resynced takes 65GB, the original (which already had high-res textures) was 30GB. Where does the other 35GB come from? Did they really write less efficient code? Not really. Just higher def textures.
SSD has what, 5% of the speed of RAM? And if you'Re copying from the SSD to the GPU, you're not copying from the SSD to whatever else you need (including ram).
Exactly what I would say too if I was cornering the market.
Reminder that the PS3 had 256MB of ram + 256MB of vram and had essentially modern games on it. We'll have to go back.
I know for a fact 125% of current RAM production is sold under 2031. That is, 125% of what is being produced right now is already sold. And we're not even talking about the new ram that is coming out/has started coming out this year.
That is, if Micron and Sy Henix are not able to increase their production by 25%, they are in deep trouble. Fortunately, they are more or less likely to double their production by 2028 at the latest.
Just to give you an idea...
On the other hand - latest model providers are now turning away customers:
https://x.com/kimi_moonshot/status/2078855608565207130
OpenAI going bankrupt would solve the problem. ( They huff 40% of the entire world's DRAM production. )
We gotta throw Sam Altman into a fucking volcano bros
his software doesn’t even work when you need it most:
https://x.com/davidsacks/status/2078991100057141620
While I do think a lot of this corrects pretty aggressively, I just think it will be years and it gets worse before it gets better. I'm considering locking up another new GPU as VRAM is becoming king and while I'm generally happy I don't want to wait five years either.
I do think there may be a time when used SSDs are a glut, because I suspect this whole AI datacenter thing to blow up a bit, a few places building them will go bankrupt and be liquidated, etc. But we will see. The problem with their GPU capacity is it's not things that can be used on the secondary market for graphics, but a ton of that becoming available secondhand would at least make local AI more feasible for the masses.