As a outside observer, NAND and DRAM prices have skyrocket ed with the AI infrastructure boom just as the China-based fabs are coming online.
It is wise for these Chinese fabs to eventually use a very aggressive dumping strategy to price well below cost push out other players forever, especially in DRAM.
But right now it seems they can max out their supply capacity without selling below cost.
Appears to me like China's endless state led (often unproductive) investment in semiconductor manufacturing subsidies (for decades) is about to pay off with some industry dominance soon.
Dumping is when you sell things for below cost. It is not dumping when you charge a 500% markup instead of a 1000% markup, even if the market is currently selling at that markup.
> It is wise for these Chinese fabs to eventually use a very aggressive dumping strategy to price well below cost push out other players forever, especially in DRAM.
Crucial's departure from the consumer market left such a gaping hole, that CXMT doesn't even need to push other players out to gain a footing.
The downside in general is that other countries lose production capacity in steel, heavy industry, semiconductors, machine tools etc - industries that took decades to build and can't be easily replaced.
Also they gradually lose the ability to meaningfully innovate in those sectors because there's no grounding against production reality anymore.
This has geopolitical consequences further down the line.
First, they're not selling at a loss; the huge price increases have allowed them to push aggrssively in the legacy markets. They're making "slightly smaller" profits than other manufacturers (of which there are now very few).
Second, they can drive out all competition and then have a captive audience for whatever prices they want, as the barriers to entry in these markets are very high. This is essentially what's happened with all higher-end manufacturing in the west over the past 30+ years.
It's funny that you call this an "very aggressive dumping strategy" while AI vendors are doing the same but with even greater losses and on a much larger scale.
It's all simply a fight for market share.
The original sin is the existing DRAM vendors selling their entire (spare) capacity to the likes of OpenAI.
Can we please stop with this irritatingly persistent myth? AI companies, at least the big ones, do not sell inference at a loss - far from it. This has been debunked and explained many times and yet it keeps being repeated.
The numbers aren't public but most guesses I've heard are that Anthropic's markup is around 50% on average, and that if considered in isolation, most models are profitable overall. The constant losses are instead due to training the next models, which will also eventually recoup but later, and forward capex investment.
This idea that big AI companies are normally and systematically selling inference at a loss as some kind of market share strategy is just not supported by the facts.
> It is wise for these Chinese fabs to eventually use a very aggressive dumping strategy to price well below cost push out other players forever, especially in DRAM.
It's not dumping, it's the opposite.
Sam Altman's stunt has created massive amounts of fictitious demand (OpenAI isn't using those wafers it's ordering) and triggered massive panic-buying from everyone else.
Prices are arteficially high, this has turbocharged China's fab and R&D budgets as you observe.
> is about to pay off with some industry dominance soon.
They're not looking to dump the semiconductor markets. They're looking to invade Taiwan.
All this buildout in their semiconductor industry is to detach themselves from the western semiconductor industry that will either sanction them if they invade Taiwan, or in the case of TSMC, suffer major damage in the ensuing conflict.
That the collapse/destruction of the Taiwanese semiconductor and electronics industries will utterly ruin the western tech industry is somewhere between a happy coincidence and acceptable collateral damage to them. No dumping required.
I realize Intel has done some serious ball dropping over the past two decades but you do realize the US has on shore cutting edge fabs, right? It's only luxury consumer electronics and the highest end corporate gear that use cutting edge nodes to begin with.
Disruption of the cutting edge would certainly wreak havoc on the pricing and specs of high end luxury electronics but that would hardly be the end of the world. I still use a desktop with DDR3 on a daily basis (granted the GPU is much newer with GDDR6) and my laptop is from the early era of DDR4 ...
> CXMT is in the process of converting wafer capacity equivalent to about 20 percent of its total DRAM output — some 60,000 wafers per month — at its Shanghai plant to the fourth-generation HBM3 chip production
I've heard that Samsung's business practices can be quite predatory. Basically if you have cool tech and you try to sell it to Samsung, you'll often get a few meetings and then they will go silent and then what you were trying to sell them will be offered by them as a new product about a year later. At least this was the situation like a decade ago.
I think this is because they are a huge conglomerate and there are divisions and groups that specialize in everything and their (Samsung) culture is to do everything as much as possible in house.
Apple has planned to explore cooperation with Chinese memory chip manufacturers Yangtze Storage (YMTC) and Changxin Storage (CXMT) to strive for more favorable supply contracts [from the big three]
That explains the cheap DDR4 DIMMs on AliExpress. Can get 2x 16GB DDR4-3200 DIMMs for A$252 delivered to Australia. A local PC store has same spec “name brand” RAM for around $380-$400.
DDR4 going from $1.35 to $11.50 in a year shows this market was already distorted before CXMT showed up.
Legacy DRAM is still over half of Samsung and SK hynix's production capacity. That's where the volume pain actually lands while they're betting everything on HBM4.
A near 10 times price rise of legacy chips and surely a lot higher profit increase couldn't motivate SKamsunix to increase their capacity for that segment, even more, they proudly informed the market about their decision to stop making DDR4 at the end of this year... to focus on even higher margin products.
But they aren't going to stop whining about China, no matter how much pain the market experiences.
Everyone here wants to be able to buy RAM at a reasonable price again.
After reading articles about CXMT and repeatedly reviewing the comments here - my take is there's nothing in play that will lead to reasonably priced RAM anytime soon.
If I'm wrong please illuminate us. We could use some hope.
Could Intel's US fabs make DRAM? Much depends on whether Apple, Dell and HP lobbyists can convince US regulators to lift restrictions on working with Chinese memory manufacturers, at least until (2028?) Micron's new semiconductor fabs can ship US-made DRAM and SSDs to US OEMs, https://wccftech.com/cxmt-ymtc-removed-from-pentagon-list-op...
The Pentagon has withdrawn the document that suggested updates to Section 1260H
Each fab will be 600,000 square feet—the size of more than 10 football fields—making them some of the biggest “clean rooms” ever built in America. To prepare the site, engineers have already blasted through more than 7 million pounds of dynamite. An army of construction workers, building contractors and architects have set up a small city’s worth of trailers so they can work around the clock. Each Boise fab is expected to use 70,000 tons of steel (almost as much used to build the Golden Gate Bridge) and 300,000 cubic yards of concrete (enough for four Empire State Buildings).
This feels like a classic business blunder. Focus hard on a single business segment, leaving an opening in the market for your competitors. Not because it wasn't profitable, but because it wasn't profitable enough for you, right now. Only downside is that now you've created an opening for a new player in the market.
This feels like a short coming of western business/stock market thinking. Focusing on profit within the next few quarters, and not caring about the longer term consequences. For all it's flaws and shady business practises at least China can think beyond a single fiscal year.
> Focusing on profit within the next few quarters, and not caring about the longer term consequences
Anything new? From my non-American view, American companies has done similar things for a very long time now. It happened in the consumer electronics, it might happen again in the IT industry.
It's not the fault of the companies, they simply just wanted more certainty and the consumer market is not (when compare to cooperate contracts).
But from the stand point of a nation, if no one creates low-end products, then no one will be providing low-end/entry-level jobs. That's when you got structural problems.
Ok but this is how the market is supposed to work. If the incumbents aren't doing what their customers want, then competitors can rise and fill the gap and compete.
This isn't a shortcoming, it's a competitive market working as intended.
Lets not forget there is no competitors. There is one competitor- the chinese state, one huge company willing to subsidize any endeavor that will help it fmgain more marketshare with already captured markets.
CXMT sells the vast majority of their bits at the prevailing market rate, just like everyone else. They are adding capacity as quickly as they can, with a 5-10 year planning horizon, just like everyone else. It’s really not that deep!
It's not really a blunder though. Given that total capacity is tightly constrained, Samsung and SK Hynix are happy to focus on what they do at their best and with the highest margins. Why shouldn't they supply the HBM market?
It's putting all eggs in one basket. If/when the higher-margin category collapses, they'll have no fallback. Imagine Chevrolet had discontinued Impalas and other low-margin cars, and switched to Corvettes during the pandemic
There is really nothing about the stock market that means only thinking about the mext few quarters. See all the losses on the profit and loss statements of AI tech giants, or, say, game console companies? Why are their stocks still valued so highly during these periods? The answer: investors are thinking long term.
It is really impossible to have quality long term thinking without capitalization accounting and similar instruments that come out of the "wester" system of business that chinese free enterprise gladly and speedily copied when it was made free.
I think the sentiment here is about management's tie of bonuses to near-term stock performance. Maybe not about the market itself, I agree with your view on investors want long term gains over short term fluctuations mostly.
Because that’s just the price for the dies, still need packaging, integration and then retail distribution. Also raw BOM is like 1/5th of the retail price usually.
Back in the 1990’s everyone had to have a unix workstation for unclear reasons (why not run Linux for < 10% the cost?).
There were crazy bubble economics schemes that meant doomed startups got unix boxes for free.
When the bubble popped, the workstation vendors hit a triple whammy: Inferior $/perf, unlimited used inventory at low prices, and an economic downturn.
The same exact thing is happening now, except the hardware is being jammed into data center models.
Anyway, when the bubble pops, people making affordable consumer stuff will be fine (like this CXMT company).
People that went all-in on firing all non-hyperscaler customers (like micron/crucial) will find they’re building the wrong chips for end-user devices, there is no server market anymore (for a few years), and they have a total addressable market of maybe 1000 distressed companies, globally.
I predict the people making these decisions and destroying their companies to juice Q2 2026 financial outlook numbers will genuinely be surprised when the bankruptcies start.
Has DDR5 caught up to DDR4 latency yet? I remember it was worse at least in the beginning. There's more bandwidth per channel but a hw design can always add more channels for the desired BW. Not so for latency.
and unfortunately increase latency even more with registered DIMMs. Comparing bandwidth increase (50 GB/s) to the stagnated latency (~80..120 ns total, less than ~0.1 GB/s) over last decades, I'm wondering, whether one still can call today's RAM random memory (though sure it can be accessed randomly). Similar to hard disk drives. Up to 300 MB/s sequentially but only up to less than 1 MB/s 4KB random (read).
I was wondering when people would find out about CXMT. I wish them luck and hope the US doesn't sabotage them. We need diversity and competition right now.
It might be very effective marketing. The big non-Chinese OEMs trust and use Korean and Japanese DRAM, and they might have been unwilling to put DRAM from CXMT into their products. (CXMT is newish, does not have access to ASML gear, which ASML would like you to believe makes it harder to make high-quality DRAM, DRAM is historically not a very large fraction of the cost of most non-huge-memory machines, and a bad DIMM is an expensive mistake for a company like Dell or HPE that is on the hook for repairs.)
But now CXMT seems to have gotten at least Dell, HP (I wonder if the article meant HPE), Acer and Asus to buy and attempt to qualify samples. If CXMT lands some serious purchasing agreements while still selling well above cost, that’s a win for them.
Does the last part of your comment explain it? They need revenue to expand capacity and the market has opened up a window to become a bigger supplier while still being profitable.
Because market rate is a 400%+ markup right now and not everyone is a greedy American kleptocrat with a diagnosable addiction to extracting every possible cent of wealth on the planet within a single fiscal quarter.
Is there a reason GPU's don't use insane "blocks" of sdcard slots (for massively parallel io) so the model weights don't need to pass through a limited PCI bus?
Yes. Let's do the math. The fastest sd cards can read at around 300 MB/s (https://havecamerawilltravel.com/fastest-sd-cards/). Modern GPUs use 16 lanes of PCIe gen 5, which is 16x32Gb/s = 512Gb/s = 64 GB/s. Meaning you'd need over 200 of the fastest SD cards. So what you're asking is: is there a reason GPUs don't use 200 SD cards? And I can't think of any way that would work
SD is obviously the wrong interface for this but "High Bandwidth Flash" (stacked flash akin to HBM) is in development for exactly this kind of problem. AMD actually made a GPU with onboard flash maybe a decade ago but I think it was a bit early. Today I would love to have a pool of 50GB/s storage attached to the GPU.
One thing to note, those aren't the fastest SD cards, those are the fastest UHS-II SD cards. The future is SD Express and you can already get microSDs at 900 MB/s.
Some years ago I realized that if I had oodles of money to spend I would totally get someone to make a PCIe card with like several hundreds microSD cards on it.
You can buy vertical microSD connectors, so you can stack quite a lot of them on a PCIe card. Then a beefy FPGA to present it as a NVMe device to the host.
Goal total capacity, as you can put 1TB cards in there. And for teh lulz of course.
The next gen inference chips will use High Bandwidth Flash (HBF) to store model weights.
These are made similarly to HBM but are lower power and much higher capacity. They can also be used for caching to reduce costs when processing long chat sessions.
As far as I understand, the "entity list" you are referring to is part of the "Export Administration Regulations", so it restricts sales from the US to restricted entities, not the other way around.
It ain't the price raise that's the mistake (even if that's what's currently painful for those of us looking to buy RAM). It's the willingness to only raise prices, and not meaningfully expand production, that's the mistake.
OTOH, now I read small Taiwanese manufacturers who are left out of the Nvidia supply chain are reverting to DDR4 motherboards because of the DDR5 shortage. Strange times.
The pertinent question, then, is who's gonna eat China's lunch?
My guess would be south/southeast Asia (India and Vietnam seem especially promising), but if the US was smart it'd put its efforts fully toward as many infrastructure investments and trade agreements and immigration agreements as possible to create a pan-American economic union. We have the resources and technology to turn every country in North and South America into an industrial and technological powerhouse. We have the resources and technology to finally conquer the Darién Gap and connect North and South America with highways and high-speed rail. We have the resources and technology to go on the offense against drug cartels (while also eliminating the failed border controls and drug prohibitions that keep those cartels in business in the first place).
If we're gonna be imperialists, then by golly let's at least be productive about it.
As a outside observer, NAND and DRAM prices have skyrocket ed with the AI infrastructure boom just as the China-based fabs are coming online.
It is wise for these Chinese fabs to eventually use a very aggressive dumping strategy to price well below cost push out other players forever, especially in DRAM.
But right now it seems they can max out their supply capacity without selling below cost.
Appears to me like China's endless state led (often unproductive) investment in semiconductor manufacturing subsidies (for decades) is about to pay off with some industry dominance soon.
Like the electric vehicle sector.
It's amazing what can be achieved when you can plan 5 years in advance, instead of just making the line go up for the next quarter.
History is littered with the corpses of those slaughtered by the millions in the name of great leader’s 5 year plans.
https://en.wikipedia.org/wiki/Great_Leap_Forward
4 replies →
Not five, but likely twenty-five. Not specific plans but attention: situation changes all the time, but the interest remains.
Don’t forget all the times those plans fail because the world did not turn out to be what you thought it would.
Dumping is when you sell things for below cost. It is not dumping when you charge a 500% markup instead of a 1000% markup, even if the market is currently selling at that markup.
> It is wise for these Chinese fabs to eventually use a very aggressive dumping strategy to price well below cost push out other players forever, especially in DRAM.
Crucial's departure from the consumer market left such a gaping hole, that CXMT doesn't even need to push other players out to gain a footing.
It's more like everyone else abandoned the market, and CXMT realised it was free real estate.
How's it dumping below cost when hey can simply sell for 100% margins instead of western makers selling for 400%.
It's easy to misread, but they're not arguing that. Note the "eventually" and "but right now".
Because only western companies are allowed to make massive profits at the expense of entire nations, it's not greed when they do it apparently.
4 replies →
I personally fail to see the downside of any manufacturer selling forever at a loss, except for the manufacturer itself.
You become dependent on the supplier.
The downside in general is that other countries lose production capacity in steel, heavy industry, semiconductors, machine tools etc - industries that took decades to build and can't be easily replaced.
Also they gradually lose the ability to meaningfully innovate in those sectors because there's no grounding against production reality anymore.
This has geopolitical consequences further down the line.
23 replies →
First, they're not selling at a loss; the huge price increases have allowed them to push aggrssively in the legacy markets. They're making "slightly smaller" profits than other manufacturers (of which there are now very few).
Second, they can drive out all competition and then have a captive audience for whatever prices they want, as the barriers to entry in these markets are very high. This is essentially what's happened with all higher-end manufacturing in the west over the past 30+ years.
If you actually believe this, then what is your explanation for a manufacturer to do this?
Do you think they are just stupid?
Microsoft gave away Internet Explorer at a loss, and what happened to internet standards?
the currency eventually collapses
1 reply →
> It is wise … to price well below cost push out other players forever
I challenge you to name a single successful example of this that isn’t state enforced.
Uber
Amazon?
2 replies →
All VC funded companies that release free or underpriced products and services, capture market share, then raise prices or enshittify?
The entire business model of VC funded tech?
4 replies →
It's funny that you call this an "very aggressive dumping strategy" while AI vendors are doing the same but with even greater losses and on a much larger scale.
It's all simply a fight for market share.
The original sin is the existing DRAM vendors selling their entire (spare) capacity to the likes of OpenAI.
No one sold their capacity to OpenAI. The vast majority of DRAM is transacted in what is essentially a quarterly auction.
2 replies →
Can we please stop with this irritatingly persistent myth? AI companies, at least the big ones, do not sell inference at a loss - far from it. This has been debunked and explained many times and yet it keeps being repeated.
The numbers aren't public but most guesses I've heard are that Anthropic's markup is around 50% on average, and that if considered in isolation, most models are profitable overall. The constant losses are instead due to training the next models, which will also eventually recoup but later, and forward capex investment.
This idea that big AI companies are normally and systematically selling inference at a loss as some kind of market share strategy is just not supported by the facts.
1 reply →
> It is wise for these Chinese fabs to eventually use a very aggressive dumping strategy to price well below cost push out other players forever, especially in DRAM.
It's not dumping, it's the opposite.
Sam Altman's stunt has created massive amounts of fictitious demand (OpenAI isn't using those wafers it's ordering) and triggered massive panic-buying from everyone else.
Prices are arteficially high, this has turbocharged China's fab and R&D budgets as you observe.
> is about to pay off with some industry dominance soon.
They're not looking to dump the semiconductor markets. They're looking to invade Taiwan.
All this buildout in their semiconductor industry is to detach themselves from the western semiconductor industry that will either sanction them if they invade Taiwan, or in the case of TSMC, suffer major damage in the ensuing conflict.
That the collapse/destruction of the Taiwanese semiconductor and electronics industries will utterly ruin the western tech industry is somewhere between a happy coincidence and acceptable collateral damage to them. No dumping required.
> utterly ruin
I realize Intel has done some serious ball dropping over the past two decades but you do realize the US has on shore cutting edge fabs, right? It's only luxury consumer electronics and the highest end corporate gear that use cutting edge nodes to begin with.
Disruption of the cutting edge would certainly wreak havoc on the pricing and specs of high end luxury electronics but that would hardly be the end of the world. I still use a desktop with DDR3 on a daily basis (granted the GPU is much newer with GDDR6) and my laptop is from the early era of DDR4 ...
1 reply →
Chinese investment has not been unproductive. It gave them independence so that the US could not cut them of- see Cuba.
Also moving into HBM:
> CXMT is in the process of converting wafer capacity equivalent to about 20 percent of its total DRAM output — some 60,000 wafers per month — at its Shanghai plant to the fourth-generation HBM3 chip production
Whenever China is mentioned, it's always about government subsidies, as if Samsung didn't receive government subsidies when it was developing
remind me is Samsung's source subsidies also regularly attacking every country's telecoms and infrastructure
You mean the US? Yeah, they are.
1 reply →
I've heard that Samsung's business practices can be quite predatory. Basically if you have cool tech and you try to sell it to Samsung, you'll often get a few meetings and then they will go silent and then what you were trying to sell them will be offered by them as a new product about a year later. At least this was the situation like a decade ago.
I think this is because they are a huge conglomerate and there are divisions and groups that specialize in everything and their (Samsung) culture is to do everything as much as possible in house.
1 reply →
is CXMT attacking every contry?
2 replies →
https://wccftech.com/apple-eyeing-a-partnership-with-chinese...
That explains the cheap DDR4 DIMMs on AliExpress. Can get 2x 16GB DDR4-3200 DIMMs for A$252 delivered to Australia. A local PC store has same spec “name brand” RAM for around $380-$400.
https://a.aliexpress.com/_mssanU1
These prices are insane. I can’t believe that’s what constitutes a good deal right now.
For context, I spent $105 USD (all in) for 2x16gb DDR5 ram last April
3 replies →
DDR4 going from $1.35 to $11.50 in a year shows this market was already distorted before CXMT showed up.
Legacy DRAM is still over half of Samsung and SK hynix's production capacity. That's where the volume pain actually lands while they're betting everything on HBM4.
A near 10 times price rise of legacy chips and surely a lot higher profit increase couldn't motivate SKamsunix to increase their capacity for that segment, even more, they proudly informed the market about their decision to stop making DDR4 at the end of this year... to focus on even higher margin products.
But they aren't going to stop whining about China, no matter how much pain the market experiences.
Everyone here wants to be able to buy RAM at a reasonable price again.
After reading articles about CXMT and repeatedly reviewing the comments here - my take is there's nothing in play that will lead to reasonably priced RAM anytime soon.
If I'm wrong please illuminate us. We could use some hope.
Could Intel's US fabs make DRAM? Much depends on whether Apple, Dell and HP lobbyists can convince US regulators to lift restrictions on working with Chinese memory manufacturers, at least until (2028?) Micron's new semiconductor fabs can ship US-made DRAM and SSDs to US OEMs, https://wccftech.com/cxmt-ymtc-removed-from-pentagon-list-op...
https://www.wsj.com/tech/micron-is-spending-200-billion-to-b... | https://archive.is/XKSpC
This feels like a classic business blunder. Focus hard on a single business segment, leaving an opening in the market for your competitors. Not because it wasn't profitable, but because it wasn't profitable enough for you, right now. Only downside is that now you've created an opening for a new player in the market.
This feels like a short coming of western business/stock market thinking. Focusing on profit within the next few quarters, and not caring about the longer term consequences. For all it's flaws and shady business practises at least China can think beyond a single fiscal year.
> Focusing on profit within the next few quarters, and not caring about the longer term consequences
Anything new? From my non-American view, American companies has done similar things for a very long time now. It happened in the consumer electronics, it might happen again in the IT industry.
It's not the fault of the companies, they simply just wanted more certainty and the consumer market is not (when compare to cooperate contracts).
But from the stand point of a nation, if no one creates low-end products, then no one will be providing low-end/entry-level jobs. That's when you got structural problems.
Ok but this is how the market is supposed to work. If the incumbents aren't doing what their customers want, then competitors can rise and fill the gap and compete.
This isn't a shortcoming, it's a competitive market working as intended.
The market doing what it's supposed to do does not negate that the market segment has only been left open because of overly myopic businesses.
4 replies →
Lets not forget there is no competitors. There is one competitor- the chinese state, one huge company willing to subsidize any endeavor that will help it fmgain more marketshare with already captured markets.
Only if you don't want or need any geopolitical gradient at all.
Everyone gets mad when Chinese do capitalism...
6 replies →
CXMT sells the vast majority of their bits at the prevailing market rate, just like everyone else. They are adding capacity as quickly as they can, with a 5-10 year planning horizon, just like everyone else. It’s really not that deep!
Are you sure? In the past they explicitly said they are not going to increase production.
https://www.tomshardware.com/pc-components/dram/memory-maker...
3 replies →
It's not really a blunder though. Given that total capacity is tightly constrained, Samsung and SK Hynix are happy to focus on what they do at their best and with the highest margins. Why shouldn't they supply the HBM market?
It's putting all eggs in one basket. If/when the higher-margin category collapses, they'll have no fallback. Imagine Chevrolet had discontinued Impalas and other low-margin cars, and switched to Corvettes during the pandemic
2 replies →
There is really nothing about the stock market that means only thinking about the mext few quarters. See all the losses on the profit and loss statements of AI tech giants, or, say, game console companies? Why are their stocks still valued so highly during these periods? The answer: investors are thinking long term.
It is really impossible to have quality long term thinking without capitalization accounting and similar instruments that come out of the "wester" system of business that chinese free enterprise gladly and speedily copied when it was made free.
I think the sentiment here is about management's tie of bonuses to near-term stock performance. Maybe not about the market itself, I agree with your view on investors want long term gains over short term fluctuations mostly.
> the average fixed contract price of PC DRAM DDR4 8Gb stood at $11.50
If that's the case, then why are the cheapest options I can find online multiple times that much?
That's a small b, that's gigabit. And that matches e.g. a "crucial pro udimm 64GB kit ddr4-3200 cl22-22-22 2Rx8" being at about 100$ a year ago.
Currently they sell here in Germany for 409€ each, that's 6.25€ for half of each of the 16Gbit chips on that kit.
> That's a small b, that's gigabit.
That explains it, thanks!
Because that’s just the price for the dies, still need packaging, integration and then retail distribution. Also raw BOM is like 1/5th of the retail price usually.
Raw BOM is much higher than 1/5th of retail price for ram. Some industries have higher margins than others.
11.5*8 = 92. I'm seeing plenty of DDR4 sticks below that price.
Brent crude is $71 a barrel, but you can't buy gas for $1.69. Why's that?
Are you able to buy crude by the gallon and refine it yourself? That would bring the cost down a bit.
2 replies →
Back in the 1990’s everyone had to have a unix workstation for unclear reasons (why not run Linux for < 10% the cost?).
There were crazy bubble economics schemes that meant doomed startups got unix boxes for free.
When the bubble popped, the workstation vendors hit a triple whammy: Inferior $/perf, unlimited used inventory at low prices, and an economic downturn.
The same exact thing is happening now, except the hardware is being jammed into data center models.
Anyway, when the bubble pops, people making affordable consumer stuff will be fine (like this CXMT company).
People that went all-in on firing all non-hyperscaler customers (like micron/crucial) will find they’re building the wrong chips for end-user devices, there is no server market anymore (for a few years), and they have a total addressable market of maybe 1000 distressed companies, globally.
I predict the people making these decisions and destroying their companies to juice Q2 2026 financial outlook numbers will genuinely be surprised when the bankruptcies start.
Has DDR5 caught up to DDR4 latency yet? I remember it was worse at least in the beginning. There's more bandwidth per channel but a hw design can always add more channels for the desired BW. Not so for latency.
> add more channels
and unfortunately increase latency even more with registered DIMMs. Comparing bandwidth increase (50 GB/s) to the stagnated latency (~80..120 ns total, less than ~0.1 GB/s) over last decades, I'm wondering, whether one still can call today's RAM random memory (though sure it can be accessed randomly). Similar to hard disk drives. Up to 300 MB/s sequentially but only up to less than 1 MB/s 4KB random (read).
People have been wondering that for a while: https://news.ycombinator.com/item?id=19304281
Awesome. Hopefully storage is next.
I was wondering when people would find out about CXMT. I wish them luck and hope the US doesn't sabotage them. We need diversity and competition right now.
More competition is always good
This is just marketing. Why would you sell at 50% of market rate? Chinese production of NAND and DRAM is not significant, it's single digit %
It might be very effective marketing. The big non-Chinese OEMs trust and use Korean and Japanese DRAM, and they might have been unwilling to put DRAM from CXMT into their products. (CXMT is newish, does not have access to ASML gear, which ASML would like you to believe makes it harder to make high-quality DRAM, DRAM is historically not a very large fraction of the cost of most non-huge-memory machines, and a bad DIMM is an expensive mistake for a company like Dell or HPE that is on the hook for repairs.)
But now CXMT seems to have gotten at least Dell, HP (I wonder if the article meant HPE), Acer and Asus to buy and attempt to qualify samples. If CXMT lands some serious purchasing agreements while still selling well above cost, that’s a win for them.
Does the last part of your comment explain it? They need revenue to expand capacity and the market has opened up a window to become a bigger supplier while still being profitable.
> Why would you sell at 50% of market rate?
Because market rate is a 400%+ markup right now and not everyone is a greedy American kleptocrat with a diagnosable addiction to extracting every possible cent of wealth on the planet within a single fiscal quarter.
Is there a reason GPU's don't use insane "blocks" of sdcard slots (for massively parallel io) so the model weights don't need to pass through a limited PCI bus?
Yes. Let's do the math. The fastest sd cards can read at around 300 MB/s (https://havecamerawilltravel.com/fastest-sd-cards/). Modern GPUs use 16 lanes of PCIe gen 5, which is 16x32Gb/s = 512Gb/s = 64 GB/s. Meaning you'd need over 200 of the fastest SD cards. So what you're asking is: is there a reason GPUs don't use 200 SD cards? And I can't think of any way that would work
SD is obviously the wrong interface for this but "High Bandwidth Flash" (stacked flash akin to HBM) is in development for exactly this kind of problem. AMD actually made a GPU with onboard flash maybe a decade ago but I think it was a bit early. Today I would love to have a pool of 50GB/s storage attached to the GPU.
2 replies →
One thing to note, those aren't the fastest SD cards, those are the fastest UHS-II SD cards. The future is SD Express and you can already get microSDs at 900 MB/s.
Some years ago I realized that if I had oodles of money to spend I would totally get someone to make a PCIe card with like several hundreds microSD cards on it.
You can buy vertical microSD connectors, so you can stack quite a lot of them on a PCIe card. Then a beefy FPGA to present it as a NVMe device to the host.
Goal total capacity, as you can put 1TB cards in there. And for teh lulz of course.
2 replies →
The next gen inference chips will use High Bandwidth Flash (HBF) to store model weights.
These are made similarly to HBM but are lower power and much higher capacity. They can also be used for caching to reduce costs when processing long chat sessions.
Maybe latency. IIRC flash is a lot laggier than DRAMs and SRAMs.
The cure for high prices is high prices.
Still no confirmation if CXMT or YMTC actually removed from entity list, until then this is jus cheap domestic inputs.
As far as I understand, the "entity list" you are referring to is part of the "Export Administration Regulations", so it restricts sales from the US to restricted entities, not the other way around.
Great moment to break into the market if you're willing to forfeit profits
This is good news. The price you pay for jacking up your prices is losing market share.
Once established, the Chinese vendors will retain most the market share if the quality is ok. The SK/JP vendors are making a big mistake.
It's not clear that raising your prices to match the supply/demand curve is a mistake
They will compete on price if they are forced to, but they aren't forced to right now
It ain't the price raise that's the mistake (even if that's what's currently painful for those of us looking to buy RAM). It's the willingness to only raise prices, and not meaningfully expand production, that's the mistake.
1 reply →
Everyone is completely sold out and adding capacity as quickly as possible.
Are they really adding capacity?
6 replies →
Good news. Now we need Chinese manufacturers of DDR4 chipsets and motherboards.
> Good news. Now we need Chinese manufacturers of DDR4 chipsets and motherboards.
Search aliexpress for X99 dual socket motherboards.
Chipsets don't determine the RAM type and all motherboards have been made in China for a while.
Taiwan dominates the global motherboard industry. (ASUS, Gigabyte, MSI, ASRock, etc. and MediaTek etc)
The big Taiwanese manufacturers are chasing the AI dragon.
https://www.digitimes.com/news/a20251021PD219/ai-server-asro... (Oct 2025)
OTOH, now I read small Taiwanese manufacturers who are left out of the Nvidia supply chain are reverting to DDR4 motherboards because of the DDR5 shortage. Strange times.
This decade is going to end with Chinese dominance in everything. Trump and AI handed them everything they need on a platter.
Western European countries got dominant, then got arrogant, letting the USA eat their lunch.
USA got dominant, got arrogant, letting China eat their lunch.
China is indeed getting dominant. They will get arrogant one day. Meanwhile, Western Europe and the USA are still very good places to live.
Exactly, there are no losers here, everyone gets lunch.
The pertinent question, then, is who's gonna eat China's lunch?
My guess would be south/southeast Asia (India and Vietnam seem especially promising), but if the US was smart it'd put its efforts fully toward as many infrastructure investments and trade agreements and immigration agreements as possible to create a pan-American economic union. We have the resources and technology to turn every country in North and South America into an industrial and technological powerhouse. We have the resources and technology to finally conquer the Darién Gap and connect North and South America with highways and high-speed rail. We have the resources and technology to go on the offense against drug cartels (while also eliminating the failed border controls and drug prohibitions that keep those cartels in business in the first place).
If we're gonna be imperialists, then by golly let's at least be productive about it.
2 replies →
> China is indeed getting dominant. They will get arrogant one day
China was dominant. They got arrogant before 1840, then there's 100 years of humiliation.