This is why I (actually/literally) just purchased a Commodore 64; I miss the days of crazy optimization and hardware hacking to fit so much into so little :).
I was joking. Running an LLM on the C64. Loading it from the tape drive! I never had to and didn't feel like I'd missed anything by not participating in the 20 minute load by tape.
As a 6 year old it taught both patience and troubleshooting, which is what want to help instill in my 7 year old son - plus the notion that you can truly own, understand, and hack your device, which sadly is gone from vast majority of tablets and phones and tvs and computers and even washing machines.
But agreed that as an adult with kids and job and family etc the wait is not ideal. Luckily, even in the 80s much c64 software was available on cartridges, and today there are even faster FPGA solutions :).
One thing is certain: in that world, we would not have to brace for a future where pretty much any form of human labor is at risk of getting replaced by just having the machines oxidize even more carbon and telling those humans to just eat cake if they fail to sell any their skills for bread.
While I definitely agree that there will be a slow down in purchase decisions for both consumers and businesses I struggle with this idea that there will be “no consumer products” in the market. There will be.
The future is starting to resemble smartphones that use predominantly RAM for local compute and inference while all persistent storage is managed by the cloud.
Anthropic is expected to spend 500B of money it doesnt have. Based on these commitments, data-centers have ordered equipment which includes memory. So, a lot of demand for memory is build on a speculative foundation.
> Anthropic is expected to spend 500B of money it doesnt have.
... yet.
That is why the big AI corps all went public this year or are trying to do so. They are valued large enough that they will be included in the popular indices (NASDAQ, S&P 500, ...) from the start, which means an awful lot of money will flow into them each month automatically from all the retirement savings.
I absolutely loathe Nvidia at this point. They've completely abandoned the core consumer that built Jensen's empire. Why not offer pre-orders of 5090s and MSRP? Why has every single drop at Best Buy gone primarily to botters and scalpers? Why not give end users half a fighting chance like EVGA used to by letting end users getinline before Nvidia crushed them?
When the AI bubble bursts, Nvidia is going to be Enron.
Because it would immediately go to scalpers anyway. Like, say the 5090 - I'm seeing prices of almost 7000€ for that thing [1]. Assuming without AI it would be priced at 1500€ - I could make about 3 months of rent just by immediately selling the GPU to the next reseller who then peddles it to some AI training farm. And in this economy, oh hell yes I would.
> When the AI bubble bursts, Nvidia is going to be Enron.
Nah. NVIDIA was a money printer way before AI was a thing. Post-collapse they will still be a money printer, just at lower speed.
The ones who are going to get fucked over hard when the AI bubble pops is pensioners, the generation who is just entering the workforce (2007ff had an aftereffect in wage depression lasting for up to a decade), and everyone who took up investments to service the AI demand (utilities, construction companies and the likes).
>> Because it would immediately go to scalpers anyway.
No they wouldn't, tens of thousands of unique users would be in line with dozens or hundreds of scalpers, if they're all allowed to buy one every 180 days per payment method, the end users actually get a fighting chance. Today they have no chance.
And it's not just a few people. The sophisticated scalping operations have hundreds of straw buyers in their network. For items that generate far less profit than a 5090 currently does.
I've watched vans dropping of an endless stream of people in the parking lot of my local Costco to buy out all the cases of Pokemon cards within 20 minutes of the "random" drop they put on the floor. They go in, buy the max quantity of the item with cash provided by their keeper, bring it back to the van with a receipt and change, and get the $20-30 per item pay they were promised. There were 2 or 3 'legit' buyers out of hundreds of obvious straw buyers.
If the profit per item is $4-5k, expect this to go into turbo mode and be done in a marketplace/automated fashion and go viral. Just sign up to your favorite scalper's platform, submit your information, get a quote and a prepaid label and drop the box you got from nVidia off at USPS the next day. That "tens of thousands" of individuals will turn into "millions" really quick once it gets known there is rent money involved by clicking a few buttons and it goes viral on all the social media platforms.
The only way out from that is manufacturer enough cards to meet demand, and at that point the retail channel doesn't really matter.
Yeah, I will never buy an Nvidia GPU again. I assume many of their former customers feel the same way. When the bubble pops, I think Nvidia is going to feel the pain as they will have pushed many of their former customers into the arms of AMD.
Everyone, and I mean everyone, is gated by ASML machines. They can't make nearly enough EUV machines to satisfy the demand. ASML has also not scaled their production nearly as much because they're quite far away from the AGI-pilled companies in the supply chain.
Well, CXMT is not gated by ASML machines because they can't buy them. They can raise their capacity because they have access to the domestically produced steppers. They are not as good as the ASML ones, and if the memory prices were not bonkers, they'd never make a profit using them, but as it is the low yields and high die sizes don't matter.
They are. They are using ASML DUV machines for memory. Samsung, SK, and Micron have moved onto EUV machines for memory because they are better. So CXMT not only has to rely on ASML, they have to rely on a worse machine to make modern memory chips.
The problem is that you’re giving them an in into the market, and as time goes on, they will get better at making memory. That’s like saying you’re going to exclude the Chinese the space station. Look what happened. They now have their own space station?
The greed and shortsightedness from both the memory makers and their big customers trying to lock out competitors from memory purchases is palpable. The first act of a Shakespearean or Greek Tragedy is where the characters take the actions that lock in the circumstances that precede their downfall. And this sets up everything for the next act where CXMT has everything it needs to establish itself on the market with buyers that would perhaps would not have considered it in other circumstances.
So if capitalism resulted in higher prices that is causing consternation among small-scale buyers, what -ism is responsible for the investment in factories that increased supply of components that resulted in lower priced electronics for the past few decades?
The samples from recent history seem to say that well regulated competitve markets are beneficial but ironically that socialism provides better competitive markets then capitalism.
I would not cheer the AI tulip craze effect on memory as a benefit of capitalism as it has several obvious unchecked violations of well regulated markets.
CXMT has entered the market, but for now they cannot increase their production fast enough to influence significantly the global prices, due to the US "sanctions", which prevent them to buy modern fabrication equipment.
So they must use older equipment, which has low productivity for state-of-the-art memories. Because of the "sanctions" there are now Chinese companies that have progressed significantly in semiconductor manufacturing equipment, but the effect of this will also take some more time to manifest.
Nonetheless, if the 3 big memory producers will not be able to increase the availability of memory until 2028, then until then CXMT will grab with certainty a major part of the market (they have already grown quickly until 10% of the market, from a negligible percent previously).
The rest of the world won’t care. They will use whatever memory they can get, and if the Chinese are willing to supply it over the next 5 to 10 years, they end up on top…
I thought your price for 32g was you getting ripped off. Holy crap. The price is super high now. Now I am lamenting missing out on getting 128GB ram when it was $700, thinking it was too high. It is outrageously expensive now.
Man whose compensation package is strongly tied to the economic performance of his company says company's economic performance will be great in 2027 and 2028
Micron reported full year revenue of $133bn, up 256%, $100bn of which was DRAM, with non-GAAP earnings last quarter of their financial year of $37.7bn on $54.23bn of revenue. Seems pretty healthy to me. Any reason to think they'll go out of business?
Only last year Microsoft admitted to having GPUs in inventory that they couldn't power, due to a lack of electricity. So much of this production capacity could very well go into making memory chips for GPUs sitting in a warehouse.
The amount of money spend on the AI hype train is crazy. Meanwhile we're wasting fab time making chips that might never be powered on. It's absolutely insane that there are more money to be made on hardware for AI that may never be used, rather than producing a product that consumers and businesses need right now.
They are probably too big to fail already (and too friendly with the current US gov). And even if they fail, open-weight models will take over as the usefulness of the tech behind has been proven.
The catchup models are all basically distillations of the sota models. Without the next training cycle things are going to stagnate. And as all those companies you mentioned are all linked to OpenAI and Anthropic and relying on hosting deals and things they are all going to be in the ringer when things to go south.
So on the one hand you have all the sota model makers doing investor expectation management in saying wet need a slowdown for safety (may or may not be true, but also means they don’t spend on the next training cycle before IPO? Could be making their books look better too?) and on the other we have a sense that the models aren’t yet at the stopping place where we can just not train another cycle and use distillations of the current generation?
The people in that interview stating this are working at the forefront of open source models and research and includes a cofounder of OpenAI. They don’t say it like it is conjecture.
I expect them to sell everything if they go bust to recoup what little they can for investors. That would double whack hardware manufacturers are their final consumers are now both leaving the markets as buyers and entering as competitors.
Given that even small GPU repair companies in China can put GPU chips on new boards with double the original memory, I would not be surprised to see the same people selling HBM from scrapped AI data centers acting as RAM on adapter boards. ;-)
I doubt it will change much, inference (not training) is very profitable and demand for inference is quite high. See: Mistral serving GLM on their own servers.
Pure inference would have to compete with other service providers and even local hardware that becomes capable of running models for a price if hardware demand by openAI and Anthropic dwindles.
That is a very commodify market and as such I expect power cost per query ends up setting the pricing.
> if hardware demand by openAI and Anthropic dwindles
Would it ever? Demand for inference (and I suppose the same applies for training) _increases_ when models get cheaper.
I suppose real risk is that one player aggressively cuts margins (say, AWS) and others follow suite. However the release of increasingly more capable, exclusive closed-weight models prevents this to an extent.
Oh sure, they'll take profits while they can, but the supply increase will start pulling the market down. It's how price signals work.
Only way to keep prices this high would be a multinational cartel of RAM makers agreeing never to compete on price. One defector kills it. Not likely you could keep a cartel like this given that the players are all on different sides of geopolitical tensions. China's not going to want to help keep Taiwan, Japan, Germany, or the USA in the money.
Not gonna happen with China in the world, they will sell to every place else on the face of the planet this memory bump is temporary, also keeping it high means other companies will reconfigure what they do. if you want to continue to sell millions of devices at a reasonable price, you’re not going to sit on the sidelines and wait for the three-headed cartel.
You are going to get busy. and design and engineer around the cartel. Is there one company in the world that might do that? A company that has the engineering, design and the money with the capacity to execute over time?
That's today. You have to think about what the industry is going to look like in a decade.
Building a DRAM company from scratch is hard, expensive, and risky. You're burning a fortune to get started, and the result is a commodity product. But the AI bubble allows CXMT to have terrible yields, sell a legacy product, and still make a healthy profit! This is literally a once-in-a-lifetime opportunity to catch up.
And yes, right now that means selling DRAM for 10% below the crazy AI-inflated prices the incumbents charge. But sooner or later either demand will drop or supply will catch up. CXMT, like many other Chinese companies, is undoubtedly willing to operate at razor-thin margins to increase market share. Either the other manufacturers move down as well, or they'll be made irrelevant by CXMT stealing all their customers.
Absolutely. far too many people are thinking about the next 2 to 4 years. How many times must they see the Chinese play the long game before they figure it out?
That’s not the point. You’re speeding up the progress of them becoming a competitor, a relentless competitor worldwide.
And you are now encouraging other large pocket tech companies inside and outside China to give thought to taking memory in-house if that means making and selling devices at a reasonable price in a timely manner.
Because the LLMs were trained on average software which doesn't give a shit about performance and they'll revert to the mean for sufficiently complicated projects. While they can help with searching for performance improvements, they also are liable to cheat if you aren't reviewing them for correctness of any algorithmic changes. Also you need to be responsible for teaching them the right benchmarks to use, since, again, performance optimization has been niche and not the standard.
As someone who programmed in Rust since before 1.0 (I started by reading the github issues daily in college when that was still feasible, volume-wise) and worked professionally using it, it can’t write better Rust than I can yet.
Save for one area, which is mostly because it was annoying to write and check with miri, which is unsafe code. It can loop through testing edge cases at my direction quickly.
But it can write code better than the average programmer without doubt. It’s only at the edge you can tell the places where it suffers. It has improved a lot, though. That rate of progress doesn’t guarantee it will capture all of that remaining edge but it’s still impressive.
I like to tell people that "AI" is the "last bubble." It was VR, the metaverse, crypto, NFTs, AR... AI is the Last Great Hype. There's really nothing beyond it. When this bubble pops it'll be remembered for decades to come because nothing will replace it.
The LLM bubble may fizzle rather than pop in any dramatic fashion, but it's certainly not the last word in AI - much more the exact opposite.
We've had AI "bubbles" / hype-fests before (expert systems, Japanese 5th generation systems, and I don't expect this will be the last.
There will also be technology bubbles and hype-fests in the future outside of AI - it's human nature. Hard to predict what they will be of course (the future is like that).
Until governments stop printing money and sending it directly to people that have nothing to do with it but investing, people will keep inventing random useless stuff to invest into.
One factor constraining supply for RAM in the short-to-medium term is that memory manufacturers is that DDR5 (and assumedly DDR6) capacity is being moved to the much more profitable HBM.
I think supply will be tight until at least the AI market crashes, which could take years or happen within months. Nobody can really predict when this will happen but it's fairly clear (IMHO) that the revenue projections needed to justify the investments by OpenAI, Anthropic and SpaceX simply don't exist.
Without the AI boom, the economy is in recession and has been for awhile. We're in an affordability crisis and the energy crunch is coming.
I'm not a business guy, but I think I would be very worried if I created any opening (even on the very long term) for competition to justify itself being built up to meet demand (or national security requirements of other countries) on an industry I mostly control, because I prioritized extremely high prices for a few years that might have the effect of changing the arithmetic for potential competitors that wouldn't have done it before.
It makes me consider whether there would be retaliation in later years against the US intellectual property system in countries other than China. While RAM is a tangible product, the areas it serves aren't really.
I think neither CEOs nor boards of the public companies care. They want to maximize their paychecks and bonuses in the short term before they leave to other companies.
It's worse, sometimes even governments fail due to overconfidence.
Brazil and the UK (well, technically Japan) caused 2 natural rubber price spikes that first led to the UK smuggling rubber tree seeds to be imported to Malaysia, where the UK took over global natural rubber production, marginalizing Brazil.
Then Japan conquered the Malaysia during WW2, at which point everyone started developing artificial rubber and natural rubber was reduced to a much smaller market.
Except that in this case it's no longer 'collect money, move on to the next one'. It's 'collect money, become dynastically-rich, retire in 3 quarters from now'. After that, who cares.
This seems like a distinction without a difference. In the context of CEOs sacrificing the greater good for their bonus, [basically] nobody has done that for a $250k bonus. We're often talking 8 figures, tens of millions of dollars.
Is this really a categorically different thing if we're talking about $85 million vs. $12 million?
I've met dozens of ~$20M net worth people, they are just normal folks - in rich countries thats not even upper class money, thats "a nice home but thats it" kind of cash. $100M+ is "i can retire at will", while $500M are like a different species, surrounded by yes-men and usually quite damaged mentally from all the ass-kissing. So yeah, i think there is a meaningful distinction.
Even if this absolute nonsense is accurate (it's not), you're agreeing with me. I said there's not a category difference between $12M and $85M in this context. And you're saying $100M is where it starts being different.
I don't think the semiconductor business is so much like that - it's become more ingrained in them that you need to manage the business for long-term sustainability.
Whether they did this or not, I think the exponential demand for memory would still be there. So there would still be conditions that encourage new competing manufacturers.
For now, there are only 3 that matter, but the reason is purely because the US government has sabotaged their competition as soon as it became apparent that the incumbents would lose market share.
Without what the US government did and still does, the prices of memory would have never increased so much and all the world would not have been forced to pay so much for memory.
I assume that some Americans have downvoted what I have written, but this is a fact that is true beyond any reasonable doubt.
A few years ago, the US government has added the Chinese vendors of DRAM and of flash memory on their sanctioned entity list exactly at the time when Apple and other major computer manufacturers were evaluating the replacement of products from Micron and the like with the Chinese alternatives. This was prevented by the "sanctions".
There is absolutely no doubt that the so-called sanctions did not have any other purpose than removing the competition for Micron.
The official reason is absolutely ridiculous, i.e. that they are among the companies that provide products to the Chinese military.
If that would have been the reason, then 100% of all existing Chinese companies from any domain of activity would have to be put on the "sanctioned entity" list, because any of them would sell to the military of their country, if given an opportunity, like all companies from USA will provide products and services to the US military, given the opportunity.
While the Chinese memory makers will obviously also sell to the military, almost all their production goes into consumer products, like smartphones and laptops, which are the main products affected by the US "sanctions" (which are no true sanctions, because true sanctions against China would have been accompanied and conditioned by some political requests, which is not the case for any of the US "sanctions").
There is also no doubt that without the US "sanctions" the Chinese memory makers would have bought the latest semiconductor manufacturing equipment and they would have had the production capacity to flood the market with their products when Micron, Samsung and Hynix stopped providing enough memory for anyone who is not among their top 5 customers.
Thus their is no doubt that the US government has stolen a lot of money from my own pocket when I was forced to update a couple of old computers this year, by artificially restricting the amount of memory that is produced and preventing competition in this market.
> I assume that some Americans have downvoted what I have written, but this is a fact that is true beyond any reasonable doubt.
I would guess that the reason for any downvotes is that even if what you wrote is completely true it isn't really relevant, because the current problem is not lack of competition. The current problem is a huge increase in demand over a short time.
Even if there were no restrictions on China and they had latest generation fabs, they would not have had latest generation fabs just sitting around idle. They would be unable to meet the current increased demand for exactly the same reason everybody else can't.
« Without what the US government did and still does, the prices of memory would have never increased so much and all the world would not have been forced to pay so much »
If that is true (and given the reasons you provide below, and the general quality of your commenting, I assume it is) then the situation is similar (if less militarized) to that of the oil sector, where the cost increase is likely higher by perhaps three orders of magnitude.
It had mostly come down to 3 (Samsung, SK Hynix, Micron), but the Chinese manufacture's market share is growing rapidly, especially as they ramp up due to sanctions. CXMT is currently 10% of global DRAM, having seen previous YOY growth of 1%->4%->8%. YMTC is now global #3 in NAND.
It wouldn't be surprising to also see Japan make a comeback in memory - they are investing heavily (both privately & government) in semiconductor manufacturing.
The argument I've seen about why someone else doesn't start selling RAM is it takes a year or two to ramp production. But here's a CEO saying he's going to make sure prices stay high enough, long enough that you can spin up a competitor
Micron is a trillion dollar company. The CEO would never say something like that directly. He's talking in the "business dialect" of English and you have to read between the lines a bit.
The CEO is saying nothing that everyone who analyzes the situation and knows at least something about semiconductor production would say. Demand is way up, supply cannot keep up, it will take some time to increase supply.
Micron is working on increasing that supply. The started building a massive fab for DRAM in 2022 in Idaho. It should start production in 2027. They started a second massive fab for DRAM next to it in 2023 with production scheduled for 2028.
This is the info that people should be looking at. Building these facilities takes time. I'm guessing they are thinking that demand is not going to decline anytime soon to justify this buildout.
> But here's a CEO saying he's going to make sure prices stay high enough, long enough that you can spin up a competitor
2 years is not remotely enough time to spin up a competitor. It's not even enough time for a current leading competitor to spin up another fab.
If you are a newcomer you are talking a decade or more before mass production. If you are HK Hynix perhaps 5 years or so.
The majors all have large fabs under construction, but the majority of new production from those facilities won't be coming fully on-line until 2028 at the earliest. Even from the date of first wafer shipped most of these facilities take 6-12mo to reach mass production level output.
Chinese DRAM is interesting only because they started building capacity (expertise) a decade ago. Only now are the expected to have any material impact on the total volume of the market. And if they want to gain even more market share, they will be under the same ~5 years to spin up a brand new fabrication plant. Perhaps a bit less because they are China, but still multiple years.
And it's important to point out how long these fabs have been under construction.
Micron started building it's new fab in 2022. They are expecting their first chips to be out by 2027 (maybe). The projected cost, $50B.
This is a fundamental problem which the free market cannot handle. It takes multiple years and billions for competitors to spin up. No sane bank or business person would try and get into cutting edge memory fabrication. That means the players we have are the players we'll likely always have (With MAYBE china entering the arena).
The free market only works when the required upfront capital for a new player to enter a business is relatively low. In industries where the players are static or shrinking, you have a free market failure.
> 2 years is not remotely enough time to spin up a competitor. It's not even enough time for a current leading competitor to spin up another fab.
A thing I heard from a friend who works in at one of the largest providers of production machines in this space: they have the machines and for some slices of the market they could basically ship them to a new fab next week. The problem is more in the installation (since the installs themselves are complicated), where they're basically limited by not having enough people to go around and _do the installs_.
It's really funny to think about how the bottleneck is so deep in the chain.
How much of this issue is self-inflicted? Micron is building a massive fab in Syracuse. They've spent like 4 years seeking approval, lobbying politicians, doing BS "environmental impact studies" (which are mostly just rent seeking and political handouts). They finally started building, but if they had just been able to put shovel in the ground we would already have this fab be producing memory.
If only we didn't have 100+ years of factories shitting where we eat we might not have thousands of governments wanting assurances that companies don't harm people in their area.
Silicon Valley is littered with SuperFund sites of bankrupted companies who conveniently couldn't afford to clean up the toxic mess they left behind when they became the vessel full of debt after the corporate shell game moved on.
It took just over 3 years to get approved, and the reason it took so long is that it is massive project (1400 acres) and has enormous infrastructure requirements that needed comprehensive study by reviewing agencies to make sure the area could handle them.
The first two fabs there will require 17 million gallons perday of water and when fully operations in the early 2040s the facility will need 48 million gallons per day. That is more than existing infrastructure can handle so the project also includes a new pipeline extension to Lake Ontario.
That water requires extensive treatment before it can be discharged. It will be pretreated on site, then go to a new $1 billion industrial wastewater treatment plant being for this project, and then finally discharged into the Oneida river.
The energy requirements when fully operation are 1850 MW drawn continuously 24/7 every day of the year. That's 16 TWh a year.
For comparison that is a little above the annual electricity use of Ft. Worth, TX or Charlotte, NC, and little below Indianapolis, IN. It is about 14% under San Francisco, CA. A little over half of Chicago, IL, about 1/3 of Los Angeles, CA, and about 30% of NYC.
That's way beyond the existing grid capacity, so lots of infrastructure upgrades there are taking place.
> because I prioritized extremely high prices for a few years that might have the effect of changing the arithmetic for potential competitors that wouldn't have done it before.
Or maybe they believe the demand is not sustainable, and they don't want to fund $10B ghost fabs.
There will be new competitors. And there will be one or two companies that will take it (memory) in house and eliminate the memory companies out of the picture.
The three greedy memory companies are changing the equations in tech industry and it won’t be in their favor in the long term.
The counterpoint is that that new, and existing, mortgage originators did not fare so well after the mortgage market collapsed in the GFC. During large parts of the housing boom deep in the money mortgages could not be refinanced as quickly as borrowers would like due to capacity constraints at the originator level. Obv, not including subprime originators, where there was arguably 10X the needed capacity.
They are not prioritizing extremely high prices. TSMC has a similar waiting list, so does ever other tool manufacturer. Memory capacity is being increased approximately as fast as it can be.
You can find public statements in the last 12 months from Samsung and Hynix about their aversion to scaling out production or increasing supply because they are concerned about profits.
Hard to know who to believe, on one side there are the absurdly high prices and statements from companies and on the other hand there are strangers posting on HN that it's actually all fine.
The memory business has always been extremely cyclical, and I doubt this time will be any different. The manufacturers have learnt this, and therefore to avoid boom followed by bust they try build for what they perceive to be longer-term more sustainable demand, not just knee-jerk increase demand to meet an immediate shortage.
The risk to any of the manufacturers, especially those with most market share, would seem to be if they are not all acting in sync. Micron's consumer brand, Crucial, had a great reputation, but now they have exited that business in favor or devoting all their capacity to currently higher margin products... will this come back to bite them later if they want to re-enter the consumer business?
What's different this time is that LLMs require orders of magnitude more memory than we could have ever used before. And I think it's getting obvious by now that inference will move to local compute, So everyone's laptops and phones would need a lot of memory.
The question the memory manufacturers have is, sure, we have all these orders out to the next three years... but will they actually be there to pay when the memory is delivered?
The memory manufacturers have been through bubbles before.
I don't deny that AI is "real", in that there is certainly something there, but I would hate to be the guy betting billions or even trillions of dollars that we're not in a demand bubble well in excess of what is justified by that tech at the moment and expanding capacity is a great idea. Of course, I would also hate to be the guy saying "no we shouldn't expand capacity" when it turns out that, yes, we should have, but the staggering profits being made in the meantime would cushion that blow pretty well.
> And I think it's getting obvious by now that inference will move to local compute
From the bottom of my heart, I hope that's the future.
I'm not convinced that will be the case. The AI companies don't want you to have local control. They want you to subscribe to a service that they can change at any time.
I grit my teeth when I type this, but (god help us) I think Apple is maybe the best (least bad?) hope here. They are the only big player with the hardware chops and without a current vested interest in getting you addicted to monthly AI subscriptions. I'm not saying this is highly likely... I'm just saying that out of the current major players, they're the ones with the ability and motivation to move in this direction in the near future.
The next best hope is probably just nVidia or AMD catering to the consumer market once again, after the AI bubble bursts or the datacenter market reaches saturation.
The outside hope is that some startup makes a business out of burning open weight LLMs to silicon like ChatJimmy... although ChatJimmy was bought up by AMD AFAIK.
It doesn't really matter what the AI companies want, if they misjudge the market somebody will just start a competitor and take it.
The relevant questions are: 1) Where are the economies of scale in the technology stack? 2) What's "good enough" to consumers, and how does that stack up with the relevant computing power needed? 3) What are the transaction costs along various system boundaries?
I think that the biggest force keeping inference in large centralized services is simply that provides a better product for the average user who doesn't care about local control (and the average user doesn't care about local control; indeed, for most people it's a misfeature, as then they have to administer their own hardware). HN is full of nerds that want to own their own stack; they're willing to put up with a little loss of capability to run Qwen 3.8 locally. But from the folks I know that have tried it vs. Claude vs. Codex vs. Antigravity, the local models are still pretty weak compared to what you can get by paying for a service. Most people will just pay for the service until the performance becomes indistinguishable and the price becomes less.
>It doesn't really matter what the AI companies want, if they misjudge the market somebody will just start a competitor and take it.
The vast majority of people are uninterested in local vs non-local models (not just LLMs but ANNs in general), only the services the models are providing. Even if performant local models can actually be served on device quickly, if the frontier labs don't want to go that way, they simply can and there's little threat the market can be taken from them that way.
If you can serve local models, the 'advantage' is that they can be run on the users device without that incurring any cost to you thus allowing flexibilty in pricing, but if such models can be made then whatever cost would be incurred to labs running on their own servers would be a rounding error and also inconsequential to them. That is to say, they can also compete with you on pricing. So unless someone comes along with something the frontier labs can't figure out or replicate, then it doesn't matter. Open AI might be forced to drum down prices but only in situations where costs have also been drummed down. In such scenarios, margins may even become better for them.
It doesn't really matter what the AI companies
want, if they misjudge the market somebody will
just start a competitor and take it.
This kind of "free markets solve everything" magical thinking is cool when we're talking about lemonade stands, but unfortunately it just isn't reality when we're talking about markets with a really big moat such AI model training.
Even if you made the US the freest market ever, you have the small problem of your free-market hero startups needing to compete with state-sponsored models from China.
(Also, free markets depend on savvy consumers to function properly. We assume that consumers will not choose lemonade that poisons them, for instance. This works well for lemonade stands -- one assumes consumers will steer clear of the lemonade stand ringed by corpses -- but works less well when the product is more complex and people don't understand it. You know, like AI.)
> LLMs require orders of magnitude more memory than we could have ever used before
That's true. But will anyone have any money to pay for the LLMs? The AI market is currently being heavily subsidised by, ultimately, everyone else. But the economy in general is looking extremely dire at the moment, and that seems unlikely to last.
Basically everyone I know who doesn't have a tech job is struggling for money / feels like the world is unaffordable for them atm. Those without jobs are struggling to get them. And those with jobs are scared to change even if they don't like them.
And from what I read in the media, a lot of the economic indicators are backing this up: discretionary spending is down, private debt levels are rising, etc. All signs that the slack in the system is disappearing.
That's what people thought about a year ago, but with the amazing strides that smaller models have taken over the past year, people are now starting to question that.
That is not true. Semiconductor industry saw a similar supply crunch during covid and spent heavily for capacity. And when the crunch tapered off, they were left holding the bag.
Micron would probably not even exist, or be a sticker brand for some Chinese conglomerate if not for this supply shock.
Or it just be that demand really does seriously exceed supply. It's not fast to bring supply online, it's costly and risky to do so if the demand will not be sustained for a long time. IMO, there's no conspiracy. I have been telling folks for the last 3 years that the price of AI compute is not coming down for a good 5+ years. Possibly 10 years too. Every human in the world will need AI, the same way we all got on the internet and then mobile. However, outside of humans, for every human there's going to be ten or hundreds of agents also all demanding compute.
What is the cost of AI tokens? You can't just calculate cost of inference. There is a cost to train, there is a cost for staff to build it all, the money investors provided for it all. When you run a business, all those costs must be recovered. Sure, it takes $2 in material cost to make a pizza, but what's the cost of pizza? You have to take into account, energy, rent, staff, insurance, and profit too. Tokens/$ is subsidized if you are talking about APIs
The big AI companies like OpenAI and Anthropic may have rushed into it too fast, and in effect wasted a lot of money on early model generations that never paid for themselves, early model architectures that were not cost-efficient, experiments that never panned out, etc.
The AI pure-play companies like OpenAI and Anthropic are presumably in worst financial position, since others like Google, Microsoft, Amazon and Meta (also Alibaba, Baidu, Bytedance, Tencent, Xiaomi) have generally been more conservative and are treating AI as an incremental revenue source not a highly leveraged all-in bet.
> Or it just be that demand really does seriously exceed supply. It's not fast to bring supply online, it's costly and risky to do so if the demand will not be sustained for a long time
This is exactly it. There's a lot of conspiracy theory in this thread without realizing that it's just what happens when an insanely large amount of capital is deployed: market distortion.
It's not just RAM, it's things like gas turbines and power transformers. I've seen suggestions that it's even affecting housing starts because the construction effort is pulling in workers, especially electricians.
However unlike you I'm not of the opinion that all that can be built, in that short timeframe. Even less so that it can be operated without pushing us over the 2C global warming threshold.
> extremely high prices for a few years that might have the effect of changing the arithmetic for potential competitors that wouldn't have done it before.
The big problem is, competitors need money. Say a fab costs 10 billion $ a piece, that is a big chunk of investment needed that absolutely needs to be paid back because even a big bank consortium will have a hard time shouldering this kind of single risk.
Now, the fab needs around three years to build... which means if there suddenly comes a huge amount of new supply or alternatively, the AI bubble pops, prices will crash, the companies won't be able to service the loans, and the banks will be sitting on a very expensive piece of real estate and machinery.
Particularly in semiconductors, this "bust and boom" cycle, "pork cycle" or however you want to call it, is very very well known. Banks and other investors have been through that one more often than you can count, that's the reason why there are so few producers left after all.
The only place I think can reasonably pull it off to start a competitor is China because their economic policy is focusing on decades in the future and the CCP is willing (and able) to shoulder the risk - however China can't acquire EUV machinery due to sanctions and will likely take a few years until they can pull it off on their own, by the time TSMC et al will already be at the next generation.
Micron has several fabs and fab upgrades coming online or under construction right now. An expansion in Boise and Virginia. Plus a huge megafab project in New York that's going to be the biggest in the world AFAIK.
All the DRAM makers have new fabs or plant expansions underway right now - indicating they believe the demand increase to be sustainable.
How many years ago did they start construction on these projects? How many years before that did they greenlight the project? I'm quite skeptical the timelines are short enough to line up with the AI boom.
That's the death trap of silicon boom-bust cycles that have claimed so many in the industry. Demand surges, you start expanding, the bust arrives, you go out of business.
Micron is putting up over $100 billion for this expansion wave IIRC. You can't afford that unless you are certain the increase in demand is sustainable.
You want to assure investors that investment will pay off so you want to talk up your order book and market outlook.
You also want to take as much profit as you can now so if things turn against you it doesn't mean bankruptcy. A pile of cash fixes many problems :)
Oil is sky high, a recession might be near. Also, optimizations (like the current KV one) can yield a drastic reduction in RAM needs. It's a risk.
If I was trying so stabilize (just stabilize) funding for humongously expensive fabs: I'd be giving markets reassurances that my business will continue to be high margins for the mid-future.
Not enough people are prepared if this isn't a bubble. I think he's betting on this being a bubble which would imply building more supply would be a mistake, because they'll end up with a factory and no more customers.
It's starting to look more and more like we will always need more memory.
I honestly expected that there would be more competition entering the memory market by now. I know it takes a lot of time and capital, but Europe has plenty of semiconductor know-how with ASML and ARM, and a strong desire to be less dependent on other parts of the world, so investment in a European semiconductor industry would make a lot of sense to me. And these high memory prices sound like the perfect opportunity.
What are those construction times in China? Not just for nuclear, but also memory or semiconductor plants, probably less than what it is in the West.
The three-headed A memory cartel should enjoy their brief moment in the sun. Because for the five years it will be over, and when I say over, it’ll be over for good…
It takes money and time. I think it’s telling we haven’t seen additional investment yet. Maybe those with the money also see this bubble ending soon? Otherwise a commodity with 80%+ margins would have competitors coming in from all sides.
If I understand this comment correctly, it's saying that Micron and the others would be better off selling RAM at below the price they could get, in order to keep competitors from deciding to enter the market.
Such collusion might well be a violation of anti-trust law, but in any case, it wouldn't work. If they decide to sell at below-market price, they must (by definition) be preventing some customers from buying by some non-price mechanism (eg, only selling to companies run by other members of the CEO's family). Potential competitors would then see an opportunity to sell to these excluded customers.
Not exactly, I think it's saying that Micron would be better off increasing supply themselves (and thus lowering prices) than waiting for a competitor to come and provide more supply to meet the higher demand.
If there was enough existing competition, that might be the right move, because if they don't increase capacity, a competitor probably would in order to gain more market share. But I think the barrier of entry for making RAM is high enough that Micron can probably get away with keeping supply low (and thus prices high) for at least a few years before a new competitor appears. Assuming demand stays high that long.
They have two massive DRAM fabs (56000 m^2 cleanroom size each) that they started construction on in 2022 and 2023 in Idaho, with production starting in 2027 and 2028.
The point they were stating was that if they kept up the high prices, that leaves room for a low priced competitor, likely from China, to occupy the lower end (and larger) share of the market. Like what BYD did to the EV market everywhere except the US. Or what Chinese mobiles did to the smartphone market. Eventually they become so powerful that they start going upmarket, trying to catch everyone in the higher end segments too. And by the time the Western companies think they can retaliate, too bad, China's got them by the bollocks on raw material.
But the reason that semiconductors are so vulnerable to boom/bust cycles (going back decades) is that takes a very long time to build the fab- you have nothing for years, and then the fab completes and enormous numbers of chips get made. So if you get demand wrong either direction you have problems. Either you flood the market or prices go way up.
I don't think it is possible to beat the fab construction delays. This isn't about things being slow in the US: TSMC took four years to go from site selection to some production at their Fab 22 plant in Taiwan (it is currently only 60% built out, is my understanding). And TSMC is best in the world at getting frontier (logic) semiconductor fabs built, it took them slightly longer (five years) for Fab 21 in Arizona. Logic fabs are not the same as memory fabs, but they are close enough to be a good comparison for timeline.
Basically, unless you started fab construction (specifically to make memory chips) in 2023 you cannot change production for 2027 and 2028. The only thing that will affect prices between now and 2028 is demand, production can't go up because the fabs to make those chips don't exist. If you started now you might get memory chips in 2031, and can you beat 40 billion dollars that the demand will still be there? (A current generation logic fab costs about 20 billion. By 2031 a logic fab will cost about 40 billion.)
> (...) that takes a very long time to build the fab- you have nothing for years, and then the fab completes and enormous numbers of chips get made. So if you get demand wrong either direction you have problems. Either you flood the market or prices go way up.
The problem is that if you do not expand your output, others will. When that happens, you will gain a competitor. Then when the crunch happens, your competitor may survive but you may not.
Therefore, if you do not increase output to meet demand them you might very well be signing the end of your company.
> You’re missing their point: by the time competition “catches up” and starts to compete the odds are that demand will have shifted.
That's besides the point, because demand shifter for everyone in the market. You included.
So your inability to supply that demand resulted in you creating the conditions for competition to show up, and that competition can very likely outcompete you out of the market.
> So your inability to supply that demand resulted in you creating the conditions for competition to show up, and that competition can very likely outcompete you out of the market.
Doubtful. Your competition started building out additional production roughly when you did. Lets say you got a 1 year head start on shovels in the ground.
Those competitors have decades of experience producing RAM. You do not. Even if you get your facility on-line to first wafer a year earlier than the incumbents, it's unlikely you will beat them to full production output as even extremely experienced operators take 6-18mo to get a fab from first production to expected nominal output.
The only way to "win" this is to have started before the supply crunch happened back in 2021 or 2022 - or more likely much earlier. Like CXMT did.
If you were able to find a bunch of investors to give you $40-50B today, already had a site ready to start construction literally tomorrow, by the time you are on-line you will be competing with new capacity currently being built by Samsung, Micron, and SK Hynix - not to mention CXMT. The only way you will be profitable is if the current demand exists through 2031 and beyond.
I know I certainly wouldn't be investing in such a project, since the entire bet is that RAM prices 5 years from now will still be at roughly what they are today. Which is simply a bet that the current AI bubble and infrastructure build-out is still ongoing.
Let's say you start a fab today. By the time it gets online, and is able to produce commercial quantities of RAM- something like 2031- you will be competing against at least two brand new fabs from Micron, at least three more brand new ones from SK Hynix, and I'm not even sure how many from Samsung. By the time your new fab is online there will be much more production from all of the existing players who have already mastered the process of producing RAM and so will have an advantage over you, who is learning. The supply crunch is entirely short term, for which there are no fixes in a universe where the laws of causality apply. This is the limitations of physical reality that software can largely avoid, but that all other forms of engineering are limited by.
Its not about lowering their prices - they shouldn't have abandoned their regular long term customers (e.g. normal people) to just chase AI fever drem money bullshit.
They would still make a lot of money but without making a lot of bad blood & making a massive opening for chinese competition.
"Sorry, we will stop selling food for the next 3 years - but worry not, those few cannibals that survive till 2029 will have the best & most affordable food from our brand new factory that we are building!"
> I think it's saying that Micron would be better off increasing supply themselves (and thus lowering prices) than waiting for a competitor to come and provide more supply to meet the higher demand.
It's a bit too late. Chinese manufacturers have finally started making DDR5 and have raised enough cash in an IPO to start ramping up production in a major way.
> Shares of chipmaker Changxin Technology Group soared nearly 466% Monday in their debut on Shanghai’s tech-heavy STAR Market, making CXMT the most valuable China-listed company.
The Hefei-based firm had raised 57.92 billion yuan ($8.6 billion) after pricing its IPO at 8.66 yuan per share, making it Asia’s biggest so far this year.
> Yes, one of those markets with infinite demand where supply has no impact on price. Those ones.
In economics, "demand" doesn't mean "I want this", it means paid, funded demand. 100s of millions of "free" LLM users aren't funded demand, they're subsidized. This leads to front-running and scalping of the silicon market at an unreal scale, it also monopolizes access to silicon and compute. Nothing happens by chance here.
I mean, we can assume the people who make the memory love getting huge prices for it. But beyond that, I just dont see why someone who invest several multi billion dollars to squeeze into a potential window that may or may not last 3 years. Especially with how volatile the whole AI bubble is.
> Especially with how volatile the whole AI bubble is.
There's nothing volatile about the AI bubble, it's been going on for a while regardless of cost and that will continue because of politics - which is rather transparent.
> I just don't see why someone who invest several multi billion dollars to squeeze into a potential window that may or may not last 3 years.
Interesting, Chinese companies don't seem to know this, they're investing bigly in RAM and other silicon despite being sanctioned and tariffed up the wazoo. Should we tell them that it's a crunch for no more that 3 years so they don't waste themselves for nothing?
The more interesting question is why Chinese RAM is hugely tariffed, if not for preserving the crunch and its monopoly pricing?
> I mean, we can assume the people who make the memory love getting huge prices for it.
Your mistake is to think of the RAM makers instead of the systemic reasons for the problem, these would have to include tariffs, sanctions, subsidies, etc that drive market pricing, RAM manufacturers and their potential competitors.
> Interesting, Chinese companies don't seem to know this, they're investing bigly in RAM and other silicon despite being sanctioned and tariffed up the wazoo.
It's also possible they're preparing for the possibility that they get cut-off and so profit is not the primary motivation. As you mention, they're already sanctioned...
Also, CMXT started in 2016 so that gives some indication of the timelines involved. Even in Taiwan it's like a year and a half just to get the building made and these are experts operating in their own country working 24/7. That timelines doesn't include equipment into it, which is also in shortage, and getting lines qualified.
they've been basically building their own PC universe. They are not specifically trying to compete with America. They're trying to disown any reliance on America.
While Trump and Business minds of America focus on pumping and dumping our stock market, they're 10 to 20 years out on what they want.
It's fucking myopic to think China cares two shits about America.
> Oracle just sent a force majeur to their data center in texas essentially telling them they can't expect to get paid for several years...
Is that what you're talking about?
"Oracle has issued a force majeure notice regarding its New Mexico AI data center project, known as Project Jupiter. This legal move is intended to allow Oracle to delay payments if the project does not meet its expected completion date in 2028."
This has nothing to do with memory prices and everything to do with too much cash chasing too few physical goods used in construction and power generation, which is a systemic problem upstream of RAM overpricing.
> And you're trying to say someone should go out and invest 5 years
Did I say that? Where? I talked about Chinese silicon makers who work in a very different system. As to why we have a monopolized market, check [1].
> You arn't paying attention.
Attention is all you need - not me though, I've got enough of it. Check again the last paragraph of my previous comment to understand why.
The person I originally replied to used pretty absolute verbiage for a problem they're apparently not very well researched about, which makes me disinterested in using hand-holding kind phrasing to someone who is boldly misleading others.
Wouldn’t shortages produce basically the same incentive structure regarding competition or national security requirements? Like the demand for RAM is actually way up, so if they don’t raise prices enough the inventory will just quickly become depleted.
This is a bit like complaining that toilet paper manufacturers did not dramatically scale up production in the first months of COVID-19. But had they invested in additional factories and production lines, they would be now holding the bag.
The memory market isn't all that different, except it's even more capital-intensive. Roughly half of HN is convinced that the current AI boom is unsustainable even though the technology itself is amazing. So why should memory makers take the bullet here?
Ultimately, as much as this sucks, having an inexpensive gaming rig is not a human right. People with money decided that all that money should go to AI, and we have to suffer through the downstream consequences of that.
If someone enters as competitor because of high prices/profit, they'll want for them to stay that way. It won't make it better for the consumer if the supply is constricted in such a way.
I think this is a point where governments should intervene and perhaps start fabs that would produce RAM at cost for working class and SMEs to use.
He's not prioritizing high prices. Demand is exploding and supply can't keep up. Everybody in the world that can is already trying to ramp up production.
What would the alternative be? If you just kept prices lower, you'd have shortages. There would still be the same motivation—if not more!—to spin up competing production.
And to build up a large amount of new capacity yourself, you need to invest a massive amount of capital into something relatively risky.
> I would be very worried if I created any opening (even on the very long term) for competition to justify itself being built up to meet demand
The interesting question is, why aren't other companies looking at this from the other direction?
For example, AMD now has a trillion dollar market cap. They make consumer and enterprise GPUs with integrated GDDR/HBM and have got to be looking at APUs that do the same going forward. They don't benefit from the price of their complement being high. They sold off their logic fabs at a point when they were half dead and had fallen behind, but you don't have to be TSMC to make DRAM. A large DRAM fab costs around $25 billion, i.e. 2.5% of their market cap. Why aren't they building two of them? Best case supply is still tight when it opens and they have a huge win, worst case the market has crashed by then but they still have a productive asset worth several billion dollars and paid for with cash from the time when it was highly available to them. And since they have internal demand, the amount the price would have to crash before the investment is a net loss is significantly more than it is for other companies.
Likewise Apple, Google, Amazon, etc. who have even more capital. What are these companies doing? They have the option to spend an amount of money they can easily afford to write off if it goes south, to get a huge win if supply continues to be tight.
Even if they hold the same opinion as Micron’s CEO, spinning up such a fab to reasonable yields would take a couple of years.
It’s a known dilemma in farming. Your crops are losing, so you decide to switch them to crops that are profitable, but so do your six neighbors and by the time the new crops start producing, you have flooded the market and everyone’s losing again.
Google and Amazon have tied up all their capital (and then some) in data centre investments, they don’t have more money to spend on fabrication. Plus they have a lot of experience running data centres, and zero experience running fabs.
The Big 3 (Samsung, Hynix, Micron) have been playing this game for 30+ years.
There used to be many more DRAM companies but most of them went bankrupt when DRAM was cheap because the big 3 will intentionally drive the price down to bankrupt the competition.
Now that they have a tri-opoly, they're milking the profits.
New competition will emerge but they don't have the capital of the big 3 so they likely won't survive the next DRAM winter.
You seem to be adding a lot of intentionality to price swings that seem expected when supply is relatively fixed while demand fluctuates. (Maybe you’re right, but it’s not the simpler explanation.)
The supply isn't fixed. A new plant takes 3-5 years to build and cost tens of billions. DRAM companies don't always know where the demand will be when the next plant completes. 2 years ago there was too much supply so DRAM price hit rock bottom. So DRAM companies held back on investing in new plants so here we are now.
All three of those companies were convicted in a price fixing scandal.
I feel like it's "obvious" that it's not particularly in the interests of these companies to increase production in the current environment. Except perhaps in the long term "we don't want to get nationalized" or something, I suppose.
Personally I do not think think that they are dragging their feet extremely intentionally, but I do think they are structurally incentivised to do so.
Historically the RAM industry has done the opposite (ramped up production aggressively due to high prices) only for that production to all come online during a glut.
This process can be fatal for a company as your expensive fixed investment hits a demand cycle that can turn overnight.
So it’s not about being a moustache twirling villain and conniving to keep prices high, it’s just not responsible or sensible to ramp up production massively based on a temporary AI bubble that could pop at any moment
There is no competition, only the illusion of it. All these big companies and owners are already invested in each other. If you have the multiple billions needed to build your own chip fab you are already getting a slice of the current pie. Why undercut yourself?
I thought that the bulk of memory is already "technically bought" by the AI hyperscalers in the form of the big (not so) "hidden" debt? As in: the supply should already be known.
Micron reportedly sold $100bn worth of DRAM last financial year, and reportedly has $40bn long term commitments. $40bn is more than their entire fiscal 2025 revenue, so it's a lot, but it's clear it'll be a minority of their available supply.
Make no mistake here: they also plan to retain these high prices. It's like the mafia now. There should be instant-jail time to cure these CEOs of their greed. Like, go to prison for a week, to help refresh your mind and stop thinking in terms of milking the customers dry.
If/when the bubble pops demand will fall so hard they'll have little ability to keep prices too high. Especially with Chinese companies pushing out a lot more product by then.
If you follow the money though much of this future “demand” is on fulling orders from the AI labs and at the moment the AI labs don’t have the cash to pay that bill. A rather important detail in this whole mix.
So either the labs need to start printing money, now, or there’s going to be a whole bunch of excess chips hitting the market in 2027-28. Someone in the middle is going to get left taking delivery of all these chips but unable to collect payment for them. And someone else is going to get left holding the bag on loans to house those chips. That’s the storm ahead.
Do you not believe that Anthropic and/or OpenAI are going to be able to IPO? Once they do, that's a huge source of capital for them that could absolutely be used to fund hardware purchases.
Based on current reporting the IPO wouldn’t generate anywhere near enough cash to pay for the bills coming due. Just between those two they have over a trillion $ of commitments coming due that need cash payment. Thats actual dollars they need to pay and neither the IPO nor current cash flows appear to be generating that kind of money to pay these bills when they come due.
It’s a big problem which neither big lab has a great answer for right now.
The payments coming due for contracted memory purchases are spread out over 5 to 10 years, not 1 to 2 years as implied by your "excess chips hitting the market in 2027-28".
For memory specifically, the counterparties are mainly "Google, Microsoft, Amazon, Meta, even Apple" like my sibling comment says (plus creditors most of us have never heard of making loans secured by data centers OpenAI and Anthropic plan to rely on).
Going public should make OpenAI and Anthropic a more attractive target for new large loans, again mostly secured by data centers.
If demand for AI services flatlines or if an international agreement stops the labs from training any new models (for safety reasons), then memory prices, GPU prices and inference prices will probably fall drastically, but OpenAI and Anthropic would probably survive that.
As long as they are taking deliveries, they need to pay for them somehow. And given the very visible shortage, the amount of stuff they ordered is massive & the payouts to the manufacturers will be similarly bug.
I don't think Micron & co in the memory cartel will be just shipping stuff on loan till 2028.
As an end user it’s difficult to not be bitter about the situation, particularly when you have execs coming out and saying, “yep these prices sure are terrible, better buy now before they get even worse!”
I’ve seen people say, “well that’s just the free market, there’s a customer that’s willing to pay more than you,” but that argument doesn’t hold water when industry incumbents are actively obstructing entrance of new competitors into their primary markets. Micron, etc can’t both refuse to carve off more of their capacity for consumer-destined NAND and also refuse to allow others to fill the vacuum they’ve so conveniently created. Either you’re serving the market or you’re not; you don’t get to have your cake and eat it too under properly functioning, healthy capitalism.
319 comments
[ 5.3 ms ] story [ 154 ms ] threadBut agreed that as an adult with kids and job and family etc the wait is not ideal. Luckily, even in the 80s much c64 software was available on cartridges, and today there are even faster FPGA solutions :).
Effectively: "Now's the time to invest..."
... yet.
That is why the big AI corps all went public this year or are trying to do so. They are valued large enough that they will be included in the popular indices (NASDAQ, S&P 500, ...) from the start, which means an awful lot of money will flow into them each month automatically from all the retirement savings.
When the AI bubble bursts, Nvidia is going to be Enron.
Because it would immediately go to scalpers anyway. Like, say the 5090 - I'm seeing prices of almost 7000€ for that thing [1]. Assuming without AI it would be priced at 1500€ - I could make about 3 months of rent just by immediately selling the GPU to the next reseller who then peddles it to some AI training farm. And in this economy, oh hell yes I would.
> When the AI bubble bursts, Nvidia is going to be Enron.
Nah. NVIDIA was a money printer way before AI was a thing. Post-collapse they will still be a money printer, just at lower speed.
The ones who are going to get fucked over hard when the AI bubble pops is pensioners, the generation who is just entering the workforce (2007ff had an aftereffect in wage depression lasting for up to a decade), and everyone who took up investments to service the AI demand (utilities, construction companies and the likes).
[1] https://www.pccomponentes.de/msi-geforce-rtx-5090-ventus-3x-...
No they wouldn't, tens of thousands of unique users would be in line with dozens or hundreds of scalpers, if they're all allowed to buy one every 180 days per payment method, the end users actually get a fighting chance. Today they have no chance.
There is enough profit to pay people to buy stuff and resell it.
I've watched vans dropping of an endless stream of people in the parking lot of my local Costco to buy out all the cases of Pokemon cards within 20 minutes of the "random" drop they put on the floor. They go in, buy the max quantity of the item with cash provided by their keeper, bring it back to the van with a receipt and change, and get the $20-30 per item pay they were promised. There were 2 or 3 'legit' buyers out of hundreds of obvious straw buyers.
If the profit per item is $4-5k, expect this to go into turbo mode and be done in a marketplace/automated fashion and go viral. Just sign up to your favorite scalper's platform, submit your information, get a quote and a prepaid label and drop the box you got from nVidia off at USPS the next day. That "tens of thousands" of individuals will turn into "millions" really quick once it gets known there is rent money involved by clicking a few buttons and it goes viral on all the social media platforms.
The only way out from that is manufacturer enough cards to meet demand, and at that point the retail channel doesn't really matter.
AI bots (probably powered by NVIDIA chips) would gobble those up in seconds.
The exception is CXMT, which uses ASML DUV machines in a less efficient process.
Was that also capital looking for returns?
I would not cheer the AI tulip craze effect on memory as a benefit of capitalism as it has several obvious unchecked violations of well regulated markets.
So they must use older equipment, which has low productivity for state-of-the-art memories. Because of the "sanctions" there are now Chinese companies that have progressed significantly in semiconductor manufacturing equipment, but the effect of this will also take some more time to manifest.
Nonetheless, if the 3 big memory producers will not be able to increase the availability of memory until 2028, then until then CXMT will grab with certainty a major part of the market (they have already grown quickly until 10% of the market, from a negligible percent previously).
https://pcpartpicker.com/trends/price/memory/#ram.ddr5.5600....
In the long term, the big comapnies are pushign to have consumers to adopt "terminals" back in old days.
- we use crappy devices, while cloud got beefy machines.
The amount of money spend on the AI hype train is crazy. Meanwhile we're wasting fab time making chips that might never be powered on. It's absolutely insane that there are more money to be made on hardware for AI that may never be used, rather than producing a product that consumers and businesses need right now.
The only chance of memory demand going down would be breakthroughs in model size reduction.
So on the one hand you have all the sota model makers doing investor expectation management in saying wet need a slowdown for safety (may or may not be true, but also means they don’t spend on the next training cycle before IPO? Could be making their books look better too?) and on the other we have a sense that the models aren’t yet at the stopping place where we can just not train another cycle and use distillations of the current generation?
Source/citation?
RAM production is easier to transfer same with flash
But sell it to consumers, where there is pent up demand
That is a very commodify market and as such I expect power cost per query ends up setting the pricing.
Would it ever? Demand for inference (and I suppose the same applies for training) _increases_ when models get cheaper.
I suppose real risk is that one player aggressively cuts margins (say, AWS) and others follow suite. However the release of increasingly more capable, exclusive closed-weight models prevents this to an extent.
Only way to keep prices this high would be a multinational cartel of RAM makers agreeing never to compete on price. One defector kills it. Not likely you could keep a cartel like this given that the players are all on different sides of geopolitical tensions. China's not going to want to help keep Taiwan, Japan, Germany, or the USA in the money.
You are going to get busy. and design and engineer around the cartel. Is there one company in the world that might do that? A company that has the engineering, design and the money with the capacity to execute over time?
Because that would be smart. And they have the packaging tech to add compute to it.
Building a DRAM company from scratch is hard, expensive, and risky. You're burning a fortune to get started, and the result is a commodity product. But the AI bubble allows CXMT to have terrible yields, sell a legacy product, and still make a healthy profit! This is literally a once-in-a-lifetime opportunity to catch up.
And yes, right now that means selling DRAM for 10% below the crazy AI-inflated prices the incumbents charge. But sooner or later either demand will drop or supply will catch up. CXMT, like many other Chinese companies, is undoubtedly willing to operate at razor-thin margins to increase market share. Either the other manufacturers move down as well, or they'll be made irrelevant by CXMT stealing all their customers.
https://www.tomshardware.com/pc-components/dram/chinese-cxmt...
And you are now encouraging other large pocket tech companies inside and outside China to give thought to taking memory in-house if that means making and selling devices at a reasonable price in a timely manner.
this would be the time to revive and evolve some of that Made In Germany Engineering ...
Save for one area, which is mostly because it was annoying to write and check with miri, which is unsafe code. It can loop through testing edge cases at my direction quickly.
But it can write code better than the average programmer without doubt. It’s only at the edge you can tell the places where it suffers. It has improved a lot, though. That rate of progress doesn’t guarantee it will capture all of that remaining edge but it’s still impressive.
It is also possible they all pivot to copying Neural Link and upselling the whole world on brain implants that the oligarchs can control.
We've had AI "bubbles" / hype-fests before (expert systems, Japanese 5th generation systems, and I don't expect this will be the last.
There will also be technology bubbles and hype-fests in the future outside of AI - it's human nature. Hard to predict what they will be of course (the future is like that).
"Crypto is the last 'big thing' because it'll revolutionize currency and everyday transactions"
"Metaverse is the last 'big thing' because everything you own will be virtual in the future'
Realistically I see AI as no different.
It wasn't the first SV-techbro hype technology, and it won't be the last.
https://en.wikipedia.org/wiki/South_Sea_Company#Quotations_p...
I think supply will be tight until at least the AI market crashes, which could take years or happen within months. Nobody can really predict when this will happen but it's fairly clear (IMHO) that the revenue projections needed to justify the investments by OpenAI, Anthropic and SpaceX simply don't exist.
Without the AI boom, the economy is in recession and has been for awhile. We're in an affordability crisis and the energy crunch is coming.
It makes me consider whether there would be retaliation in later years against the US intellectual property system in countries other than China. While RAM is a tangible product, the areas it serves aren't really.
Brazil and the UK (well, technically Japan) caused 2 natural rubber price spikes that first led to the UK smuggling rubber tree seeds to be imported to Malaysia, where the UK took over global natural rubber production, marginalizing Brazil.
Then Japan conquered the Malaysia during WW2, at which point everyone started developing artificial rubber and natural rubber was reduced to a much smaller market.
Is this really a categorically different thing if we're talking about $85 million vs. $12 million?
You could say that about pretty much any employee - and you'd be completely wrong. What makes you right in the case of CEOs and boards?
https://www.cnbc.com/2026/04/30/us-ceo-pay-grew-20-times-fas...
Without what the US government did and still does, the prices of memory would have never increased so much and all the world would not have been forced to pay so much for memory.
A few years ago, the US government has added the Chinese vendors of DRAM and of flash memory on their sanctioned entity list exactly at the time when Apple and other major computer manufacturers were evaluating the replacement of products from Micron and the like with the Chinese alternatives. This was prevented by the "sanctions".
There is absolutely no doubt that the so-called sanctions did not have any other purpose than removing the competition for Micron.
The official reason is absolutely ridiculous, i.e. that they are among the companies that provide products to the Chinese military.
If that would have been the reason, then 100% of all existing Chinese companies from any domain of activity would have to be put on the "sanctioned entity" list, because any of them would sell to the military of their country, if given an opportunity, like all companies from USA will provide products and services to the US military, given the opportunity.
While the Chinese memory makers will obviously also sell to the military, almost all their production goes into consumer products, like smartphones and laptops, which are the main products affected by the US "sanctions" (which are no true sanctions, because true sanctions against China would have been accompanied and conditioned by some political requests, which is not the case for any of the US "sanctions").
There is also no doubt that without the US "sanctions" the Chinese memory makers would have bought the latest semiconductor manufacturing equipment and they would have had the production capacity to flood the market with their products when Micron, Samsung and Hynix stopped providing enough memory for anyone who is not among their top 5 customers.
Thus their is no doubt that the US government has stolen a lot of money from my own pocket when I was forced to update a couple of old computers this year, by artificially restricting the amount of memory that is produced and preventing competition in this market.
I would guess that the reason for any downvotes is that even if what you wrote is completely true it isn't really relevant, because the current problem is not lack of competition. The current problem is a huge increase in demand over a short time.
Even if there were no restrictions on China and they had latest generation fabs, they would not have had latest generation fabs just sitting around idle. They would be unable to meet the current increased demand for exactly the same reason everybody else can't.
If that is true (and given the reasons you provide below, and the general quality of your commenting, I assume it is) then the situation is similar (if less militarized) to that of the oil sector, where the cost increase is likely higher by perhaps three orders of magnitude.
It wouldn't be surprising to also see Japan make a comeback in memory - they are investing heavily (both privately & government) in semiconductor manufacturing.
There are 3 companies
(Who said that we can only talk like a pirate one day a year?)
Where does he say he'd going to make sure prices stay high?
Micron is working on increasing that supply. The started building a massive fab for DRAM in 2022 in Idaho. It should start production in 2027. They started a second massive fab for DRAM next to it in 2023 with production scheduled for 2028.
- Oreo CEO, suggesting you to buy every Oreo you can get
More like 5-10
2 years is not remotely enough time to spin up a competitor. It's not even enough time for a current leading competitor to spin up another fab.
If you are a newcomer you are talking a decade or more before mass production. If you are HK Hynix perhaps 5 years or so.
The majors all have large fabs under construction, but the majority of new production from those facilities won't be coming fully on-line until 2028 at the earliest. Even from the date of first wafer shipped most of these facilities take 6-12mo to reach mass production level output.
Chinese DRAM is interesting only because they started building capacity (expertise) a decade ago. Only now are the expected to have any material impact on the total volume of the market. And if they want to gain even more market share, they will be under the same ~5 years to spin up a brand new fabrication plant. Perhaps a bit less because they are China, but still multiple years.
But it is enough time for customers to consider competitors they may have avoided before.
So maybe their priced-out customers give Chinese DRAM a serious look, make some compromises, and never go back to Micron.
Micron started building it's new fab in 2022. They are expecting their first chips to be out by 2027 (maybe). The projected cost, $50B.
This is a fundamental problem which the free market cannot handle. It takes multiple years and billions for competitors to spin up. No sane bank or business person would try and get into cutting edge memory fabrication. That means the players we have are the players we'll likely always have (With MAYBE china entering the arena).
The free market only works when the required upfront capital for a new player to enter a business is relatively low. In industries where the players are static or shrinking, you have a free market failure.
A thing I heard from a friend who works in at one of the largest providers of production machines in this space: they have the machines and for some slices of the market they could basically ship them to a new fab next week. The problem is more in the installation (since the installs themselves are complicated), where they're basically limited by not having enough people to go around and _do the installs_.
It's really funny to think about how the bottleneck is so deep in the chain.
Silicon Valley is littered with SuperFund sites of bankrupted companies who conveniently couldn't afford to clean up the toxic mess they left behind when they became the vessel full of debt after the corporate shell game moved on.
The first two fabs there will require 17 million gallons per day of water and when fully operations in the early 2040s the facility will need 48 million gallons per day. That is more than existing infrastructure can handle so the project also includes a new pipeline extension to Lake Ontario.
That water requires extensive treatment before it can be discharged. It will be pretreated on site, then go to a new $1 billion industrial wastewater treatment plant being for this project, and then finally discharged into the Oneida river.
The energy requirements when fully operation are 1850 MW drawn continuously 24/7 every day of the year. That's 16 TWh a year.
For comparison that is a little above the annual electricity use of Ft. Worth, TX or Charlotte, NC, and little below Indianapolis, IN. It is about 14% under San Francisco, CA. A little over half of Chicago, IL, about 1/3 of Los Angeles, CA, and about 30% of NYC.
That's way beyond the existing grid capacity, so lots of infrastructure upgrades there are taking place.
Or maybe they believe the demand is not sustainable, and they don't want to fund $10B ghost fabs.
The three greedy memory companies are changing the equations in tech industry and it won’t be in their favor in the long term.
Hard to know who to believe, on one side there are the absurdly high prices and statements from companies and on the other hand there are strangers posting on HN that it's actually all fine.
The risk to any of the manufacturers, especially those with most market share, would seem to be if they are not all acting in sync. Micron's consumer brand, Crucial, had a great reputation, but now they have exited that business in favor or devoting all their capacity to currently higher margin products... will this come back to bite them later if they want to re-enter the consumer business?
The memory manufacturers have been through bubbles before.
I don't deny that AI is "real", in that there is certainly something there, but I would hate to be the guy betting billions or even trillions of dollars that we're not in a demand bubble well in excess of what is justified by that tech at the moment and expanding capacity is a great idea. Of course, I would also hate to be the guy saying "no we shouldn't expand capacity" when it turns out that, yes, we should have, but the staggering profits being made in the meantime would cushion that blow pretty well.
From the bottom of my heart, I hope that's the future.
I'm not convinced that will be the case. The AI companies don't want you to have local control. They want you to subscribe to a service that they can change at any time.
I grit my teeth when I type this, but (god help us) I think Apple is maybe the best (least bad?) hope here. They are the only big player with the hardware chops and without a current vested interest in getting you addicted to monthly AI subscriptions. I'm not saying this is highly likely... I'm just saying that out of the current major players, they're the ones with the ability and motivation to move in this direction in the near future.
The next best hope is probably just nVidia or AMD catering to the consumer market once again, after the AI bubble bursts or the datacenter market reaches saturation.
The outside hope is that some startup makes a business out of burning open weight LLMs to silicon like ChatJimmy... although ChatJimmy was bought up by AMD AFAIK.
The relevant questions are: 1) Where are the economies of scale in the technology stack? 2) What's "good enough" to consumers, and how does that stack up with the relevant computing power needed? 3) What are the transaction costs along various system boundaries?
I think that the biggest force keeping inference in large centralized services is simply that provides a better product for the average user who doesn't care about local control (and the average user doesn't care about local control; indeed, for most people it's a misfeature, as then they have to administer their own hardware). HN is full of nerds that want to own their own stack; they're willing to put up with a little loss of capability to run Qwen 3.8 locally. But from the folks I know that have tried it vs. Claude vs. Codex vs. Antigravity, the local models are still pretty weak compared to what you can get by paying for a service. Most people will just pay for the service until the performance becomes indistinguishable and the price becomes less.
The vast majority of people are uninterested in local vs non-local models (not just LLMs but ANNs in general), only the services the models are providing. Even if performant local models can actually be served on device quickly, if the frontier labs don't want to go that way, they simply can and there's little threat the market can be taken from them that way.
If you can serve local models, the 'advantage' is that they can be run on the users device without that incurring any cost to you thus allowing flexibilty in pricing, but if such models can be made then whatever cost would be incurred to labs running on their own servers would be a rounding error and also inconsequential to them. That is to say, they can also compete with you on pricing. So unless someone comes along with something the frontier labs can't figure out or replicate, then it doesn't matter. Open AI might be forced to drum down prices but only in situations where costs have also been drummed down. In such scenarios, margins may even become better for them.
Even if you made the US the freest market ever, you have the small problem of your free-market hero startups needing to compete with state-sponsored models from China.
(Also, free markets depend on savvy consumers to function properly. We assume that consumers will not choose lemonade that poisons them, for instance. This works well for lemonade stands -- one assumes consumers will steer clear of the lemonade stand ringed by corpses -- but works less well when the product is more complex and people don't understand it. You know, like AI.)
That's true. But will anyone have any money to pay for the LLMs? The AI market is currently being heavily subsidised by, ultimately, everyone else. But the economy in general is looking extremely dire at the moment, and that seems unlikely to last.
Can you elaborate on this? What about the economy is looking dire?
And from what I read in the media, a lot of the economic indicators are backing this up: discretionary spending is down, private debt levels are rising, etc. All signs that the slack in the system is disappearing.
Micron would probably not even exist, or be a sticker brand for some Chinese conglomerate if not for this supply shock.
They are a lot more reticent this time around.
What does that even mean ?
The cost of AI itself (tokens/$ for a given level of intelligence) has been falling rapidly, and shows no sign of stopping.
The AI pure-play companies like OpenAI and Anthropic are presumably in worst financial position, since others like Google, Microsoft, Amazon and Meta (also Alibaba, Baidu, Bytedance, Tencent, Xiaomi) have generally been more conservative and are treating AI as an incremental revenue source not a highly leveraged all-in bet.
This is exactly it. There's a lot of conspiracy theory in this thread without realizing that it's just what happens when an insanely large amount of capital is deployed: market distortion.
It's not just RAM, it's things like gas turbines and power transformers. I've seen suggestions that it's even affecting housing starts because the construction effort is pulling in workers, especially electricians.
However unlike you I'm not of the opinion that all that can be built, in that short timeframe. Even less so that it can be operated without pushing us over the 2C global warming threshold.
The big problem is, competitors need money. Say a fab costs 10 billion $ a piece, that is a big chunk of investment needed that absolutely needs to be paid back because even a big bank consortium will have a hard time shouldering this kind of single risk.
Now, the fab needs around three years to build... which means if there suddenly comes a huge amount of new supply or alternatively, the AI bubble pops, prices will crash, the companies won't be able to service the loans, and the banks will be sitting on a very expensive piece of real estate and machinery.
Particularly in semiconductors, this "bust and boom" cycle, "pork cycle" or however you want to call it, is very very well known. Banks and other investors have been through that one more often than you can count, that's the reason why there are so few producers left after all.
The only place I think can reasonably pull it off to start a competitor is China because their economic policy is focusing on decades in the future and the CCP is willing (and able) to shoulder the risk - however China can't acquire EUV machinery due to sanctions and will likely take a few years until they can pull it off on their own, by the time TSMC et al will already be at the next generation.
All the DRAM makers have new fabs or plant expansions underway right now - indicating they believe the demand increase to be sustainable.
Micron is putting up over $100 billion for this expansion wave IIRC. You can't afford that unless you are certain the increase in demand is sustainable.
You want to assure investors that investment will pay off so you want to talk up your order book and market outlook.
You also want to take as much profit as you can now so if things turn against you it doesn't mean bankruptcy. A pile of cash fixes many problems :)
If I was trying so stabilize (just stabilize) funding for humongously expensive fabs: I'd be giving markets reassurances that my business will continue to be high margins for the mid-future.
It's starting to look more and more like we will always need more memory.
The three-headed A memory cartel should enjoy their brief moment in the sun. Because for the five years it will be over, and when I say over, it’ll be over for good…
Such collusion might well be a violation of anti-trust law, but in any case, it wouldn't work. If they decide to sell at below-market price, they must (by definition) be preventing some customers from buying by some non-price mechanism (eg, only selling to companies run by other members of the CEO's family). Potential competitors would then see an opportunity to sell to these excluded customers.
If there was enough existing competition, that might be the right move, because if they don't increase capacity, a competitor probably would in order to gain more market share. But I think the barrier of entry for making RAM is high enough that Micron can probably get away with keeping supply low (and thus prices high) for at least a few years before a new competitor appears. Assuming demand stays high that long.
They have two massive DRAM fabs (56000 m^2 cleanroom size each) that they started construction on in 2022 and 2023 in Idaho, with production starting in 2027 and 2028.
I don't think it is possible to beat the fab construction delays. This isn't about things being slow in the US: TSMC took four years to go from site selection to some production at their Fab 22 plant in Taiwan (it is currently only 60% built out, is my understanding). And TSMC is best in the world at getting frontier (logic) semiconductor fabs built, it took them slightly longer (five years) for Fab 21 in Arizona. Logic fabs are not the same as memory fabs, but they are close enough to be a good comparison for timeline.
Basically, unless you started fab construction (specifically to make memory chips) in 2023 you cannot change production for 2027 and 2028. The only thing that will affect prices between now and 2028 is demand, production can't go up because the fabs to make those chips don't exist. If you started now you might get memory chips in 2031, and can you beat 40 billion dollars that the demand will still be there? (A current generation logic fab costs about 20 billion. By 2031 a logic fab will cost about 40 billion.)
The problem is that if you do not expand your output, others will. When that happens, you will gain a competitor. Then when the crunch happens, your competitor may survive but you may not.
Therefore, if you do not increase output to meet demand them you might very well be signing the end of your company.
That's besides the point, because demand shifter for everyone in the market. You included.
So your inability to supply that demand resulted in you creating the conditions for competition to show up, and that competition can very likely outcompete you out of the market.
Doubtful. Your competition started building out additional production roughly when you did. Lets say you got a 1 year head start on shovels in the ground.
Those competitors have decades of experience producing RAM. You do not. Even if you get your facility on-line to first wafer a year earlier than the incumbents, it's unlikely you will beat them to full production output as even extremely experienced operators take 6-18mo to get a fab from first production to expected nominal output.
The only way to "win" this is to have started before the supply crunch happened back in 2021 or 2022 - or more likely much earlier. Like CXMT did.
If you were able to find a bunch of investors to give you $40-50B today, already had a site ready to start construction literally tomorrow, by the time you are on-line you will be competing with new capacity currently being built by Samsung, Micron, and SK Hynix - not to mention CXMT. The only way you will be profitable is if the current demand exists through 2031 and beyond.
I know I certainly wouldn't be investing in such a project, since the entire bet is that RAM prices 5 years from now will still be at roughly what they are today. Which is simply a bet that the current AI bubble and infrastructure build-out is still ongoing.
They would still make a lot of money but without making a lot of bad blood & making a massive opening for chinese competition.
But greed is greed...
It's a bit too late. Chinese manufacturers have finally started making DDR5 and have raised enough cash in an IPO to start ramping up production in a major way.
> Shares of chipmaker Changxin Technology Group soared nearly 466% Monday in their debut on Shanghai’s tech-heavy STAR Market, making CXMT the most valuable China-listed company.
The Hefei-based firm had raised 57.92 billion yuan ($8.6 billion) after pricing its IPO at 8.66 yuan per share, making it Asia’s biggest so far this year.
https://www.cnbc.com/2026/07/27/cxmt-china-market-debut-chip...
The Memory cartel do not care about silly things like laws
In economics, "demand" doesn't mean "I want this", it means paid, funded demand. 100s of millions of "free" LLM users aren't funded demand, they're subsidized. This leads to front-running and scalping of the silicon market at an unreal scale, it also monopolizes access to silicon and compute. Nothing happens by chance here.
Eventually we got the same ghost town after this AI bubble madness.
These are billion dollar+ funded operations that take multi-years just to see if you can compete at the levels required.
All these comments show up every time, like there's a memory faerie land where you just put on your memory faerie hat and walla! Memory!
So bizarre.
There's nothing volatile about the AI bubble, it's been going on for a while regardless of cost and that will continue because of politics - which is rather transparent.
> I just don't see why someone who invest several multi billion dollars to squeeze into a potential window that may or may not last 3 years.
Interesting, Chinese companies don't seem to know this, they're investing bigly in RAM and other silicon despite being sanctioned and tariffed up the wazoo. Should we tell them that it's a crunch for no more that 3 years so they don't waste themselves for nothing?
The more interesting question is why Chinese RAM is hugely tariffed, if not for preserving the crunch and its monopoly pricing?
> I mean, we can assume the people who make the memory love getting huge prices for it.
Your mistake is to think of the RAM makers instead of the systemic reasons for the problem, these would have to include tariffs, sanctions, subsidies, etc that drive market pricing, RAM manufacturers and their potential competitors.
It's also possible they're preparing for the possibility that they get cut-off and so profit is not the primary motivation. As you mention, they're already sanctioned...
Also, CMXT started in 2016 so that gives some indication of the timelines involved. Even in Taiwan it's like a year and a half just to get the building made and these are experts operating in their own country working 24/7. That timelines doesn't include equipment into it, which is also in shortage, and getting lines qualified.
https://www.tomshardware.com/tech-industry/building-a-chipma...
While Trump and Business minds of America focus on pumping and dumping our stock market, they're 10 to 20 years out on what they want.
It's fucking myopic to think China cares two shits about America.
And you're trying to say someone should go out and invest 5 years, multiple billions in investment to catch up to today memory prices?
You arn't paying attention.
Is that what you're talking about?
"Oracle has issued a force majeure notice regarding its New Mexico AI data center project, known as Project Jupiter. This legal move is intended to allow Oracle to delay payments if the project does not meet its expected completion date in 2028."
This has nothing to do with memory prices and everything to do with too much cash chasing too few physical goods used in construction and power generation, which is a systemic problem upstream of RAM overpricing.
> And you're trying to say someone should go out and invest 5 years
Did I say that? Where? I talked about Chinese silicon makers who work in a very different system. As to why we have a monopolized market, check [1].
> You arn't paying attention.
Attention is all you need - not me though, I've got enough of it. Check again the last paragraph of my previous comment to understand why.
[1] https://news.ycombinator.com/item?id=49927315
Staying in the consumer market would have lowered the market price.
I mean they did stop selling to consumers
Ah, sorry, _not_ charging the highest price you could get at a market should be an antitrust case?
Absolutely no chance. You cant force a company to optimize for its so called (in economics) >producer surplus<.
How should this work? Minimum price control by gov?
Wouldn’t shortages produce basically the same incentive structure regarding competition or national security requirements? Like the demand for RAM is actually way up, so if they don’t raise prices enough the inventory will just quickly become depleted.
The memory market isn't all that different, except it's even more capital-intensive. Roughly half of HN is convinced that the current AI boom is unsustainable even though the technology itself is amazing. So why should memory makers take the bullet here?
Ultimately, as much as this sucks, having an inexpensive gaming rig is not a human right. People with money decided that all that money should go to AI, and we have to suffer through the downstream consequences of that.
I think this is a point where governments should intervene and perhaps start fabs that would produce RAM at cost for working class and SMEs to use.
And to build up a large amount of new capacity yourself, you need to invest a massive amount of capital into something relatively risky.
The interesting question is, why aren't other companies looking at this from the other direction?
For example, AMD now has a trillion dollar market cap. They make consumer and enterprise GPUs with integrated GDDR/HBM and have got to be looking at APUs that do the same going forward. They don't benefit from the price of their complement being high. They sold off their logic fabs at a point when they were half dead and had fallen behind, but you don't have to be TSMC to make DRAM. A large DRAM fab costs around $25 billion, i.e. 2.5% of their market cap. Why aren't they building two of them? Best case supply is still tight when it opens and they have a huge win, worst case the market has crashed by then but they still have a productive asset worth several billion dollars and paid for with cash from the time when it was highly available to them. And since they have internal demand, the amount the price would have to crash before the investment is a net loss is significantly more than it is for other companies.
Likewise Apple, Google, Amazon, etc. who have even more capital. What are these companies doing? They have the option to spend an amount of money they can easily afford to write off if it goes south, to get a huge win if supply continues to be tight.
It’s a known dilemma in farming. Your crops are losing, so you decide to switch them to crops that are profitable, but so do your six neighbors and by the time the new crops start producing, you have flooded the market and everyone’s losing again.
$50B for Micron.
> Why aren't they building two of them?
They don't have in-house expertise to do that.
I feel like it's "obvious" that it's not particularly in the interests of these companies to increase production in the current environment. Except perhaps in the long term "we don't want to get nationalized" or something, I suppose.
Personally I do not think think that they are dragging their feet extremely intentionally, but I do think they are structurally incentivised to do so.
3 businessmen are in jail, they ask each other what they are there for.
Businessman 1: "I made my prices higher than my competitors and was jailed for price gouging!"
Businessman 2: "I made my prices lower than my competitors and was jailed for price dumping!"
Businessman 3: "I made my prices the same as my competitors and was jailed for price collusion!"
This process can be fatal for a company as your expensive fixed investment hits a demand cycle that can turn overnight.
So it’s not about being a moustache twirling villain and conniving to keep prices high, it’s just not responsible or sensible to ramp up production massively based on a temporary AI bubble that could pop at any moment
Why the need for speculation?
Make no mistake here: they also plan to retain these high prices. It's like the mafia now. There should be instant-jail time to cure these CEOs of their greed. Like, go to prison for a week, to help refresh your mind and stop thinking in terms of milking the customers dry.
So either the labs need to start printing money, now, or there’s going to be a whole bunch of excess chips hitting the market in 2027-28. Someone in the middle is going to get left taking delivery of all these chips but unable to collect payment for them. And someone else is going to get left holding the bag on loans to house those chips. That’s the storm ahead.
It’s a big problem which neither big lab has a great answer for right now.
These labs made a bold bet that cost trillions, but those trillions in research are a box of concrete around their feet now
For memory specifically, the counterparties are mainly "Google, Microsoft, Amazon, Meta, even Apple" like my sibling comment says (plus creditors most of us have never heard of making loans secured by data centers OpenAI and Anthropic plan to rely on).
Going public should make OpenAI and Anthropic a more attractive target for new large loans, again mostly secured by data centers.
If demand for AI services flatlines or if an international agreement stops the labs from training any new models (for safety reasons), then memory prices, GPU prices and inference prices will probably fall drastically, but OpenAI and Anthropic would probably survive that.
I don't think Micron & co in the memory cartel will be just shipping stuff on loan till 2028.
I’ve seen people say, “well that’s just the free market, there’s a customer that’s willing to pay more than you,” but that argument doesn’t hold water when industry incumbents are actively obstructing entrance of new competitors into their primary markets. Micron, etc can’t both refuse to carve off more of their capacity for consumer-destined NAND and also refuse to allow others to fill the vacuum they’ve so conveniently created. Either you’re serving the market or you’re not; you don’t get to have your cake and eat it too under properly functioning, healthy capitalism.