Seagate Sparks Memory Sell-Off As CEO Says It Would 'Take Too Long' To Build New Factories (cnbc.com) 70
Seagate CEO Dave Mosley said Monday that building new memory chip factories or adding capacity would "take too long" to keep up with AI-driven storage demand. "If we took the teams off and started building new factories or bringing up new machines, that would just take too long. You would end up with more capacity, but then you'd slow the rate of growth on that technology," Mosely said. CNBC reports: Memory chip stocks have soared in recent months as a flood of AI investing has sent demand soaring, with the chips a key part of the AI buildout in data centers. Chip production cycles stretch over many quarters for a single unit, and investors are increasingly wary of how long the leading memory makers can capture demand. CME Group is launching a new futures market for semiconductors, enabling more traders to lock in prices and hedge against the rising prices of computing power.
At Monday's conference, Mosely also addressed the "very long lead times" and maintaining predictability with its clients. "We know what's coming out a year from now," he said. "And we've basically gone to the customers and said, 'Look, if you want to plan this really well, which it should be for your data centers, we know what's coming out. You can buy this stuff up to a certain period.' And so we want to keep that four or five quarters of visibility very, very solid for what's being built. But the demand is significantly higher than that."
At Monday's conference, Mosely also addressed the "very long lead times" and maintaining predictability with its clients. "We know what's coming out a year from now," he said. "And we've basically gone to the customers and said, 'Look, if you want to plan this really well, which it should be for your data centers, we know what's coming out. You can buy this stuff up to a certain period.' And so we want to keep that four or five quarters of visibility very, very solid for what's being built. But the demand is significantly higher than that."
Technobabble translation... (Score:5, Insightful)
He thinks its a fad and won't invest in new factories... but he can't SAY that because he wants that sweet, sweet AI bubble money.
Re:Technobabble translation... (Score:5, Interesting)
No, I suspect it's got more to do with short-term profits and his overall compensation, given he probably wouldn't still be the CEO by the time any new factories were brought online.
Corporate boards typically only reward short-term thinking.
Re:Technobabble translation... (Score:5, Interesting)
No, I suspect it's got more to do with short-term profits and his overall compensation, given he probably wouldn't still be the CEO by the time any new factories were brought online.
Corporate boards typically only reward short-term thinking.
No not at all. Firstly he's been the CEO of Seagate for a decade and there's no indicating that he'll step down in 3 years. On the flip side we have already seen major cracks in the AI industry.
1. OpenAI cancelled a huge order with SK Hynix when they aborted the Stargate datacentre. - A good sign that the industry is cracking under it's overpromises.
2. Wall Street has actively created indices to allow hedge funds explicitly to invest in companies *NOT* in the AI space. - A good sign that the finance industry thinks AI will crash.
3. Even with current prices, fabs takes years to build and many many years more to make an ROI. This is not a decision you make in the wake of a 1-2 year bubble. It's the kind of decision you make when you're sure of demand 10 years from now.
Anyone here thinking long term will not be building anything either. Building something in a bubble that won't be completed until after the bubble pops is the kind of reckless investment that would actually get a CEO to no longer be in their position a few years from now.
Re: Technobabble translation... (Score:3)
Well, you could say that xai can make more money leasing out their ai datacentre than they can by using their ai datacentre fot ai. Doesnt that tell you that ai isnt very profitable at xai?
Re: (Score:2)
I don't understand what the endgame is supposed to be for all of this. The numbers being thrown around don't make any sense at all--trillions of dollars in spending chasing billions of dollars in revenue looks insane and even worse than the dotcom era.
There's a meme out there that I can't help but agree with: "The reason RAM prices went up 4x is that a massive amount of not-yet-manufactured memory was bought with money that doesn't really exist to be put into GPUs that haven't been made yet, to be installe
Re: (Score:2)
Deals fall apart all the time, you clown.
Of course they do. That's kind of how bubbles start popping, when deals fall apart. By the way what is a "deal"? The OpenAI Stargate project wasn't "a deal", it was a landmark project for the industry, a $500bn datacentre which would have consumed 40% SK Hynix's manufacturing capacity.
Investments this year will surpass $2.5 trillion.
And yet a single project from one company just knocked 20% off that figure. That's how bubbles work, when suddenly a double digit percentage of an industry changes on a whim you know you're in a period of instability.
Are you retarded?
You not h
Re: (Score:2)
I'm also going to add that the number of "GW" of AI promised to be installed exceeds the power generation manufacturing capacity of the world. Building more of the latter to match the former exceeds the heavy industry capacity of the world.
There's no physical way to get from where we are now to the growth claims that are buoying AI stocks!
Re: (Score:2)
There's also been some developments in AI that means that the demand may level off a bit. Google said yesterday that they can now train AI with servers distributed over the globe. If one region has cheap renewable electricity right now, they can move more of it there. So the need to build more capacity and more storage is reduced as well.
There are some similar developments coming from Chinese companies too, particularly around making AI more efficient.
Hopefully by the end of the year we will be seeing reduc
Re: (Score:2)
That Nvidia is plowing it's hardware profits back into AI/Datacenter companies, to keep the ball rolling, it's somewhat worrisome.
Seems like there's not a lot of profits in AI field overall, except for what the shovel sellers are making.
Nvidia Pours Billions Into the Companies Supporting AI Growth [nationalcioreview.com]
Re: (Score:2)
I think you're on to it.
Like many things, AI is subsidized to gain traction. Once the true costs are put on to consumers to run the models and what they're doing, demand will drop dramatically as measuring the change to using AI products, cost, ROI will have to be weighed much more carefully. When that happens, investors will slow on AI investments.Supply will start to catch up with demand etc.
Yeah I'm just basically saying the same thing you are, it's a bubble and people are expecting this one to pop.
Re:Technobabble translation... (Score:5, Insightful)
Sandisk has been devloping an HBM-killer memory called High-Bacndwidth-Flash [techpowerup.com]... but they aren't sharing much information. I'd say this announcement either means High-Bacndwidth-Flash isn't panning out or that they won't need new factories to produce it.
As far as what others are doing that makes investing in DDR5 dumb:
1. Cerebras doesn't use memory chips, just a crazy crazy amount of on-CPU cache [heygotrade.com].
2. Intel is re-entering the memory market with a sorta radical new ZAM memory that is considered an HBM-killer [tomshardware.com].
3. Turboquant is fairly useless at actually speeding things up currently, but it proves that we can shrink the need for memory if compute can handle this type of quantization in-line [research.google]. In the near-term this will mean accelerators coming out in the next few years will have these schemes baked into the hardware and need far less RAM. This is roughly comparable to when we started getting hardware accelerated texture compression with 3D graphics cards. It looks like the Turboquant idea was plagiarized and overhyped, but the idea is likely amazing if applied at the CPU level [biggo.com].
Re: (Score:3)
Re:Technobabble translation... (Score:5, Informative)
Building out a new fab takes years. I have done it.
You have to order the hardware (which has to be made to fill your order -it is not sitting waiting on a shelf), build the facility (some of the equipment is delicate and large enough that you pour a slab, crane in the equipment, and then assemble the building around it). You also need trained workers, and an entire material supply chain. Reliable power and water supplies. Once all that is in place you can begin the process of tuning the equipment -it does not just turn on and start producing chips.
Re:Technobabble translation... (Score:5, Interesting)
Well, to be fair, the memory companies have suffered the past 3-4 times that memory prices skyrocketed, they brought more factories online, and then prices plummet just before the factory comes online so they're selling increased capacity into a surplus market.
It's why a company like Kingston exists - Kingston exists solely to absorb surplus memory. If memory makers make too much RAM, they sell it to Kingston and Kingston makes a bunch of memory modules from it. Likewise if they make too much NAND flash, it goes into USB sticks and SSDs. It's why Kingston RAM and flash products are so variable - they just take the surplus parts and put them into products. You might get Samsung products one day, the next day it's SK Hynix because that's what's coming in the factory.
None of the memory makers are bringing up their timelines of increased RAM production from the late 2020s/2030 because they don't want o bring it online into a market th's flooded.
Re: (Score:1)
Apple seems to feel the same, and thus doesn't want to tie up too much money in AI. After the poppage*, Apple can snap up AI-ware at a nice discount.
* "bursting"?
Also he's not afraid of competition (Score:2)
So basically I suspect the memory producers have all agreed not to increase production and there's no way a new player could enter the market because with the complete lack of antitrust law enforcement the memory
Re: (Score:2)
Indeed. He expects this insane and really mindless hype to be over well before any new factories could start producing things. That means likely less than 5 years from now. I think he is perfectly correct on that one.
Cartel (Score:5, Interesting)
Re:Cartel (Score:5, Informative)
Yep... back as far as 2006, several execs were indicted for price-fixing RAM:
https://www.justice.gov/archiv... [justice.gov]
There was a class action lawsuit in 2018 over the same issue.
I assume Apple, at least, feels they've taken steps to control RAM availability with their transition to the M series ARM processors, because they integrate the memory and the video memory into the CPU itself?
Re: (Score:2)
Apple and RAM as part of CPU (Score:2)
Ok... you have me questioning the details now, but the AI overview I just checked says:
"Yes, Apple's M-series chips integrate the RAM directly onto the processor package. Because the memory is built-in as "Unified Memory," it cannot be upgraded or replaced after your Mac or iPad is purchased."
Re: (Score:2)
Do not just trust AI answers, what that is referring to is the RAM soldered on the motherboard and if you question the AI's answer it will apologise for misleading you and provide a better answer.
Re: (Score:2)
Apple do NOT integrate RAM in the CPU (Score:3)
Apple include at most at most a few hundred MB of RAM(L1+L2+system level cache) in the CPU/GPU.
CPU RAM is SRAM which generally uses 6 transistors per bit. DRAM uses 1 transistor and 1 capacitor and as such is cheaper to produce but is significantly slower.
Also note that there is high demand for CPU/GPU/AIPU manufacturing as well as RAM which can easily be verified by noting the increase in share price of e.g. TSMC.
There will continue to be high prices for RAM, GPUs etc until the AI bubble bursts or produces
Re: (Score:2)
This is more complicated than I had realised.
With the M4 processor TSMC integrate DRAM(from another manufacturer as TSMC do make DRAM) into the SoC using InFo-oS(Integrated Fan-Out on Substrate see https://3dfabric.tsmc.com/engl... [tsmc.com] and click the InFo-oS tab). So the SoC package includes a fixed amount of DRAM. So while I was technically correct that the DRAM is not part of the CPU it is now part of the M4 SoC package.
Re: (Score:2)
sorry meant to write TSMC do NOT make DRAM
Re: (Score:2)
The chinese are getting into the game I belive. They're behind for now but it might not be long until they are taking a big chunk of the market. That's why Seagate should build new factories and sell the memory at a competitive price, but going for temporary gains over long term ones is the american way.
Re: (Score:2)
US has proposed laws to not only ban ASML from selling China DUV scanners but even servicing what's already there. Would be the end of China's memory and flash fabs.
Though that would likely push China into a full on trade war, catapult flash prices into the stratosphere, and hell might even get EU to get into a tradewar with the US. EU is cowardly, but even they will have a breaking point. It comes on top of the tariffs, alternative forms of protectionism and the ICC shenanigans.
Re: Cartel (Score:2)
Re: (Score:2)
Re: (Score:2)
don't expect them to add capacity because if they all agree not to they make more money
While they were indeed colluding, that's not the issue here. The issue is that largely many people think that AI expansion now is a bubble. RAM prices skyrocketed due to demand that may ultimately not be fully realised. Just look at SK Hynix - they just had a massive order cancelled because OpenAI won't proceed with Stargate.
It takes many years to build a new fab. It's not the kind of activity you do during the bubble. It's the kind of activity you hope you did before the unpredictable bubble arrives. There
Re: (Score:2)
I refuse to believe the claim that "this would require billions of dollars and at least five years to get a factory operational."
There is clearly enormous amounts of money circulating in the industry right now. If a company like Nvidia genuinely wanted to manufacture its own memory, it absolutely could. Even with initially poor yields, the economics could still work. A 50% yield rate is far less concerning when RAM prices have increased by 200%, especially for a company purchasing memory in massive volumes
tl;dr (Score:2)
$orry I can't hear you.
Re: (Score:3)
$orry I can't hear you.
... through this huge pile of dollar bills.
Covid broke everything (Score:5, Insightful)
Before Covid, we had a semblance of free-market supply and demand. Demand would ultimately influence supply to increase, driving down prices. But in Covid companies learned that with just-in-time supply chains and huge barriers to entry, they could pretty much be cartels with no legal consequence and make a ton of money. So when demand goes up they no longer bother with trying to increase supply to meet that demand. In fact they often now reduce supply to drive up prices even farther. In hindsight it's an obvious flaw in the free market system as we practice it, but it was never exploited nearly to the degree it is now.
Re: (Score:2)
Re: (Score:2)
I'm pretty sure at this point everyone exploited Covid to do this. Destroy the supply chains....weaken supply....raise prices....squeeze the public dry.
This time next year you won't be talking about the computer you'll buy...but the $1200 terminal you have to buy to use the $200/month cloud PC service. They won't be selling to anyone but datacenters.
That's called a smartphone (Score:2)
It costs $1200 and acts as a terminal to cloud services.
Re: (Score:2)
Sorry but that's a load of crap for many reasons.
a) JIT supply chains have been a focus for nearly a century now. COVID changed absolutely nothing and prior to COVID we have also had massive cost demand swings in various industries including tech due to supply and demand disruption.
b) There's very little JIT going on here. JIT is something that affects you in months. The fact that we have had computers delivered at reasonable prices many months after part costs shot up, and pre-builds only really started be
Diminishing returns (Score:2)
Re: Diminishing returns (Score:1)
DDR6 isn't any faster than DDR4. The only real difference is that DDR6 is less reliable.
Re: (Score:2)
Depends on the application. I can point to quite a few places where memory speed actually shows it's face. For example....I can run 8b LLMs on my AMD laptop....but the minute the context window gets anything in it....boy does the speed tank. The internal graphics cores render everything just fine...but the fact the memory is 85GB/s vs 250GB/s makes a HUGE difference.
Re: (Score:1)
When storing big things, moving quickly is unsurprisingly also important. In the context of what's driving the market with LLMs, it basically just boils down to how fast the model can respond. Add a lot of context, a lot of parameters, some "reasoning", and some safety layers on top then it takes a lot of memory, a lot of processing, and a lot of moving information around.
People get pretty frustrated when services are slow to respond or when the answers aren't very good, so that's why there's a lot of compe
That kind of thinking brings in new players.. (Score:5, Insightful)
Either the companies think this AI hoarding will be short enough to not risk expansion (every economist and corporate type loves constant expansion) or they think their stranglehold on the market is too strong for anyone to offer new competition.
For us mere retail plebs - the corporate greed situation just sucks and is getting worse. This is the situation anti-trust laws were supposed to prevent.
Re: (Score:3)
Ideally there's a 3D memory breakthrough soon. DRAM made like flash would reduce production cost to a fraction (ie. one exposure for 100's of storage layers, it's not the 3D part which is important but the fact it takes only one exposure).
Re: (Score:2)
Re: (Score:2)
NAND 3D flash has only one high resolution exposure, for the entire stack of storage layers. The per layer litho for the staircase contacts is low resolution. Cells which can be formed in similar manner have been proposed for DRAM.
https://ieeexplore.ieee.org/ab... [ieee.org]
https://www.niar.org.tw/en/xmd... [niar.org.tw]
https://neosemic.com/neo-semic... [neosemic.com]
I have my doubts it can actually work reliably, would be nice though.
Re: (Score:2)
Another one from IMEC.
https://www.imec-int.com/en/pr... [imec-int.com]
The destructive streaming reads makes it fundamentally different from normal DRAM, but works fine for large matmuls (ie. AI).
Re: (Score:2)
Re: (Score:2)
They all say they use NAND like processing. In 3D NAND the full stack of alternating layers is created first, it's literally impossible to use lithography on the internal layers beyond what comes from on top. As far as high resolution goes, that's just the holes.
https://semiengineering.com/3d... [semiengineering.com]
There's an old paper which title catches the essence of 3D NAND well ... "3D memory: etch is the new litho".
PS. I'm ignoring the logic, but that's not repeated with the cell layers.
Re: (Score:2)
Re: (Score:2)
Except the barrier to entry is too high. It would be fun if someone could manage to disrupt the memory consortium.
The Chinese are massively rising in the ranks right now. Major PC vendors are currently qualifying CXMT's RAM offerings. But the problem is you're thinking on the wrong time scale. A bubble like this doesn't bring in new players, and indeed the Chinese new players who are entering the market right now actually simply got really lucky as they started building their fabs years ago, before AI was shoved down our throats.
Re: (Score:3)
Except the barrier to entry is too high.
Yeah, not just for manufacturing, but also for licenses and patents. When there's this few companies in this market worldwide, it's too legally complicated and risky to try to disrupt the existing consortium.
Just a taste of what is coming! (Score:1)
What they have, are Massive Automation Machines which they are marketing and selling as AI. Which require huge data centers and the supporting vendors.
If/when AGI does happen, I doubt it will require hundreds of huge centralized data centers.
What? is each robot going to be bound to communicating with a data center for every little fork in the logic.
Elon Musk and his out-of-the-box thinking ... (Score:2)
... might just give the IC industry the kick in the butt it needs, just like with spacecraft and electric vehicles. He recently stated that the whole IC-Fab thing is done wrong these days and that he might just end up eating a hamburger and smoking a cigar right next to a microfab with higher cleanroom efficiency to prove his point once the first Terafab is up and running.
I'm no engineer, but the "copy-exactly" and "clearroom design" of the late 70ies sure has become long in the tooth and my intuition says
Re: (Score:3)
I'm no engineer,
Yes. Musk is not one as well. He has no clue how semiconductor production works. Here is a hint: A current ASML lithography system is 2 stories high, 160 metric tons in weight and requires 250 experts 6 months to assemble. And this idiot is hallucinating about "microfabs".
Memory chip stocks (Score:2)
By that he means the NYSE or NASDAQ stock exchange price of memory chip companies, not the quantity of memory chips in the warehouse.
chips that don't exist for data centers that don't (Score:2)
The market is a wet dream of manufacturers, of course. Already knowing that for all the forseable future however much you can produce will be sold at very good prices - amazing.
Until the house of cards comes down. Most of the stuff ordered is, as someone put well into a meme, money that doesn't yet exist buying chips that don't yet exist for data centers that have not yet been built.
Seagate Sparks Memory Sell-Off (Score:2)
Seagate Sparks Memory Sell-Off
What exactly does this mean?
Re: (Score:1)
I, too, was confused by the article summary and the original article on CNBC itself.
I found this article that does a better job of explaining what happened to the stocks of interest that day, and why:
https://www.investing.com/news... [investing.com]
It's interesting to note that Seagate's stock has recovered its loss from Monday already (and probably the other ones, too).