One unit of HBM capacity consumes roughly the wafer capacity that could have produced three units of DDR5 capacity. HBM dies need to be larger than ordinary DRAM dies because of how the final packaging works.
> As discussed previously, the ramp of HBM production will constrain industry supply growth in non-HBM products. Industrywide, HBM3E consumes approximately three times the wafer supply as D5 to produce a given number of bits in the same technology node. With increased performance and packaging complexity, across the industry, we expect this trade ratio for HBM4 to be even higher than the trade ratio for HBM3E. We anticipate strong HBM demand due to AI, combined with increasing silicon intensity of the HBM roadmap, to contribute to tight supply conditions for DRAM across all end markets. As the memory industry is still recovering from the challenging environment in 2023, this tight supply environment will help drive the considerable improvements in profitability and ROI (return on investment) that are needed to enable the investments required to support future growth.
>As the memory industry is still recovering from the challenging environment in 2023, this tight supply environment will help drive the considerable improvements in profitability and ROI (return on investment) that are needed to enable the investments required to support future growth.
The problem is that they need actual chips tomorrow. I fear we are reaching the point in the semiconductor industry where, in order to sustain the revenue growth propping up their valuations, they're going to have to start selling future chips that cannot possibly be physically produced.
The problem isn't the price. No matter how much you pay, you can't take delivery of a chip that doesn't exist. Now, you could raise prices to the point where you destroy demand. But that's a tricky window to maneuver through.
> The problem isn't the price. No matter how much you pay, you can't take delivery of a chip that doesn't exist
I’m confused because this reads like a denial of basic economics. If the price is high enough, the chip will be produced for you.
Do you mean because of the production lead time, higher prices won’t result in increased production? Commodities like corn have been managing this for a long time… what’s special about chips?
Corn is heavily subsidized,also has (in the US) federally prvided insurance programs against plunging prices, has paid uselessness(ethanol). Despite all of this corn has caused massive agriculture bankruptcy/takeover when it has failed anyway.
So it's more "what's special about corn". It is also fairly hilarious to claim the parent is denying basic economics and then bring up corn as an example of having successfully managed economics. If the scales were not being thumbed, and "basic economics" were in play, corn would be in very very bad shape.
In the case of DRAM, there is an incredibly long history of these gloom/glut cycles, and they have stayed roughly the same timeframes (~3 years) since the 1990's.
Almost all the ones who have survived this long are either in the same kind of boat as corn - protected in various forms from the downside - or don't increase production and get caught out until they are absoultely forced.
The very temporarily increased profit is not worth going bankrupt for - they make more money long term by being very cautious and know this.
There are a near infinite number of economic studies you could look at (and several sibling comments cite some) - DRAM manufactuers don't chase the price and probably couldn't anymore if they want to.
None of this denies basic economic theory, of course, since economic theory is not exactly "rigorous", even to the degree it could be (IE even the parts that are pure analysis of data rarely reproduce!).
But once you get past econ101, you are taught how markets react to percieved bubbles, how those with near monopolies would rather pocket a windfall now than risk investing in future expansions.
> HBM3E consumes approximately three times the wafer supply as D5 to produce a given number of bits in the same technology node
That's really interesting, and I wonder why? I believe HBM has redundant ECC bits by default, which would add a few %, but other than that, is it just that the yield is much lower due to die stacking? Of course, this is a 2024 document so things may have changed a bit since.
HBM requires stacking the chips. So they need to shave the layers, glue, stack more, shave again. They also require a substrate what is even more wafers. The issue is that a error in the stack means a lot of losses.
In order to get high bandwidth, you want memory as close as possible to the GPU. The more trace lane length = signal loss, bandwidth loss.. HBM is compact, and so you can stack 24GB modules, 8 around a GPU die.
If you tried to do that with normal memory, you need like 64 modules. So a a TON of traces more that all need to be equal length, and because so many = far away from the GPU = less bandwidth.
The issue is like stated above, its a process that waste a ton of wafers. Wafers that can make easily 3x more normal memory.
Intel with "Crescent Island" is trying to make a 160GB card using LPDDR5x memory but the bandwidth is only ~700GB/s.
My own PC died and now I’m locked out of my entire Steam library for the foreseeable future. A $2000 PC is literally a downgrade from what I bought 10 freaking years ago.
This is why I'm so hesitant to use AI. There are other reasons (for example: the journey is the fun part, not the destination; and I want to learn not have something done for me), but the pressure AI puts on memory and storage is too much for me. I do use it occasionally, mostly to keep tabs on it and for small problem solving (how do I do this one piece of a query in MySQL, since I'm not a MySQL DBA), but that's it for now, until this memory/storage issue is fixed.
Unfortunately, I can't not use AI because it provides the best speech recognition I've ever used and, going along with hand problems, it now allows me to write code after a three-year absence from that skill. I do miss it. I'd never do it again professionally in the large, but I have missed writing code a lot.
My solution is to see what four or five-year-old equipment I can buy that will let me run local LLMs. I may only get six or seven tokens per second out of an i7, but it's a start. And best of all, I can turn the machine off when I'm not using it.
IMO, Migrating to small-scale local LLMs would be a significant improvement over using data centers.
> Unfortunately, I can't not use AI because it provides the best speech recognition I've ever used and, going along with hand problems, it now allows me to write code after a three-year absence from that skill.
I'll second this. The combination of Whisper + LLM makes speech recognition fantastic. I occasionally have arm pain from typing, and this is a Godsend.
I don't use it to write code - but in my experience stuff like emails + docs was the greater source of pain (one generally types slower while coding).
> Migrating to small-scale local LLMs would be a significant improvement over using data centers.
This shift is probably inevitable, but it will vary significantly by region depending on prices for electricity. Look at, for example, the difference in fundamental homelab build recommendations between Germans and just about anyone else. Electricity prices in Germany are so high that even a now expensive Raspberry Pi or other ARM board is often preferred over Intel/AMD builds due to low power draw (especially low idle power draw), an effect that adds up for a machine running all the time over years.
With local LLMs and the GPUs to run it, especially if you want a model available to you all the time and can remote into your local network to use it whenever you want, there's no escaping much higher power draws, even at idle. Wherever electricity is expensive, the electric bill can be a prohibitive barrier.
No-Load losses on the power conversion seems to be the real issue and is rarely mentioned and understood even less often. Plus the conversion eff. curve.
Eg: I shaved ~40W off the idle load on a server (250->210W) by doung nothing more than removing the redundant supply
You just reminded me of a funny story about redundant power supply.
A college buddy used to work at Motorola (I'm naming the company because they wouldn't mind this story being shared) back in the late 90's or early 2000's. They had redundant power to their campus, bought from two different companies, coming in on opposite sides of the campus, so that even if some backhoe operator cut a ground-based power line somewhere, they wouldn't lose power.
And yet, one morning, the power went off all across their campus. After a little investigation, they sent pretty much all their employees home at noon and told them "take the afternoon off, don't come back until tomorrow, you wouldn't be able to do any work anyway". Turns out that although the power lines came in at opposite sides of their campus, somewhere a few miles away both of the power lines feeding their campus ended up running through the same underground conduit. And yes, a backhoe had managed to cut that conduit and break both of the lines they depended on at the same time. They had a single, VERY non-obvious, point of failure, and the backhoe had unerringly homed in on that SPoF.
> IMO, Migrating to small-scale local LLMs would be a significant improvement over using data centers.
LLM serving is most efficient when you batch a lot of parallel requests together. Data center solutions also have the advantage of collecting queries from around the globe, so the hardware can be utilized around the clock.
Having everyone serve their own local LLMs would produce a lot more memory demand. Not less. The same memory would be idle most of the time, and when it was used it would be used for 1 person instead of a batch of requests.
There are other reasons to run local LLMs, but solving hardware demand problems is not one of them.
Inefficient per watt, yes, but local inference capacity is greatly underutilized in aggregate. If a model can run on a machine that already exists, that's a bunch of additional chips that don't need to be built.
Seems like we need a USB equivalent standard to RAM sticks. I have lots and lots of old (1-2GB), intercompatible sticks sitting. Sometimes you just need the space and not the speed.
Hey, I apologize for using you as tech support, but I found tons of those M10 sticks you are talking about, but what should I be looking for in terms of an adapter? All I can find are adapters listed as M.2:
For starters, The slowest sticks of DDR(x) are often slower than DDR(x-1). The issue is never capacity, but rather, performance.
The real nonsense? USB? It's a mess. Pick a USB cable and buy it from ANY retailer, let's make it easier, buy a USB-C cable. What is the data rate (depends on cable quality and length), Does it support power delivery? If so, how many watts? (~5W requires a very different cable from ~230W), how do you know from simply looking at the cable? If you buy a cable, how can you tell what it supports by simply looking at the connector? Imagine having a box full of USB-C cables. Could you tell me how fast each of those cables are? (The cables themselves don't! Many don't have any markngs at all, and if they do, the markings could be fraudulent)
RAM does not have that issue. A stick fits or it doesn't. Sure, there are a huge range of speeds, however, that range has a ballpark (JEDEC defines the ballpark, the "cartels" make the memory and push out some faster stuff).
While there are definite exceptions (I was bitten by one recently), you can generally plug in a DDR5 DIMM and expect it to work in the system.
The same cannot be said for USB. Some USB cables ONLY deliver power. Some only work with certain devices. I've USB-C (!!!) cables that only work with the devices they are shipped with, and even more annoyingly, the both may be true! ASUS (!!!) ships MOTHERBOARDS that can only pair with vital hardware via very specific USB-C cables (Strix Hive) and "GLORIOUS" ain't so glorious. Their mouses warn you not to use other USB-C cables, and they are right, depending on the make/model/generation, you can brick your "glorious" peripheral.
No, USB nonsense needs to stay far away from anything else.
Also, the reason this is a huge issue is because the memory makers are cartels. Only a few of them exist, they gang up and bully EVERYONE and nobody has ever invested money to create a competitor due to this, except China, which of course means that the U.S. and portions of the E.U. are insta-banning/trying to insta-ban, even though it is really freaking hard to install spyware on a memory module.
USBism has a lot of elements similar to the "connect your agent to your data with this 'curl | sh' script and do anything. Maybe some friction and incompatibility is useful in some cases.
> Pick a USB cable and buy it from ANY retailer, let's make it easier, buy a USB-C cable.
Okay, let's assume it's a USB-C cable and not just something USB-C shaped that pretends to be one.
> Does it support power delivery?
Yes.
> If so, how many watts?
Always at least 60W (3A). Up to 240W if it's e-marked.
> What is the data rate
Depends whether it's a USB 2.0-only cable (480Mbps) or a complete cable (20Gbps). Possibly more in Thunderbolt or USB4 modes - cables that handle those will be marked (both visually and with an e-mark). In any case, pretty easy to check just by plugging things in.
I don't know why you got downvoted so much. There's no such thing as a 5 watt USB cable and people need to stop acting like there is. 60 is a very good baseline. There's no such thing as a power only USB cable either.
If we're counting fake products then we need to count those RGB sticks that have no RAM in them.
Yeah but DDR3 was faster then and still is now - I can't imagine my M4 Mac having a slightly older version RAM would meaningfully affect me, but another 24GB of it (or whatever I have) would certainly be an improvement.
A better SSD can get up to a passable GB/s rate, but when you're waiting tens of microseconds instead of tens of nanoseconds you can't use it the same way. And a worse SSD is 10x slower in GB/s too, compared to old super cheap RAM.
This will have general inflationary consequences for consumer products (phones, consoles, laptops, etc.). On top of current uncertainties regarding oil and fertilizers, I think 2% inflation in the US and Europe would be a very optimistic target.
Yep. And not in the future, it is here right now. And poor countries will as always be impacted way more than the US or Europe. The AI industry is the very direct cause and should be shamed for handling their rollout in such a destructing way. They deliberately created and grew the feeling of urgency for their own benefit
The average consumer isn't purchasing a lot of products with a lot of RAM every year.
Their 8GB of RAM phones will go up in price a little bit, but people aren't buying phones every year or even every other year.
So if the price of the 8GB of LPDDR went from $40 to $160 and it's all passed on to the consumer buying a new phone every 4 years, that's an extra $30/year in spending.
Most adults I know now use their phone for everything, so they're not buying new computers and laptops. They can wait 3 years for the market to settle before upgrading those, too.
Even I'm a heavy buyer of tech products and RAM, and I would bet you that my family's annual food bill fluctuates by more than what I've had to pay for RAM prices growing.
Almost everything we consume is downstream of RAM pricing. I don’t buy a phone every year, but corporate workforces, streaming servers, cars, factories, etc. will all cost more next year.
Oh, I hope you are right about cars. If one of the outcomes of the memory crunch is that manufacturers start making “dumb cars” again, it would almost make it all worthwhile.
You forgot to tack on the rise in prices of SSDs too, so that's $50 per year at least.
>Most adults I know now use their phone for everything
You mean the cloud for everything. A lot of phone tasks share their computational and storage workloads off in the invisible ether that has actual computers with real costs behind them. Amazon has no problem with bumping up their costs in order to pay for their fleet of servers. I mean, what business doesn't use the cloud these days.
The headline says "sold out". That doesn't mean prices will go up, that means the memory will not be available at all for many companies and the products they make will not be available to consumers.
You don't have to directly buy RAM to be affected. It's just like oil or electricity...if RAM was part of the economic process that you participated in, then that price increase will be passed down to you. If every datacenter in the US suddenly had to pay double for electricity, how would that affect every operation that is operating in that datacenter? Ok now it's not electricity, it's the RAM.
Remember when car companies couldn’t get the chips they needed to make cars? And so now there’s not enough used cars. And so car companies charge way more the new ones and just stopped making cheap cars altogether?
It's inflationary because this demand is backed by Monopoly money. OpenAI is promising to buy half the global supply of RAM with money they don't actually have, pricing out people who would like to make purchases with real money and not $1 trillion in debt.
Hmm. Forget about desktops. I'm programming a microcontroller which is actually reasonably priced and has gotten all its 32Mb of RAM in the same die as the processor, but I'm getting nervous and having a strong urge to stockpile...
Well, that one is likely out of danger - so far PSRAM (I assume that's what you have) is not impacted that much. However, we had to cancel some projects because we couldn't secure enough chips which have DDR on die at a reasonable price. They are 4x now compared to December last year.
Amazon has implemented a password system for RAM delivery. You're supposed to give a password to the deliverer before taking delivery of the RAM, and you're not supposed to send the password over the phone at all (although the Amazon delivery guy still called me and asked for it over the phone).
Its a way to make sure the delivery takes place. A code is sent to recipient and recipient only gives it to the delivery person once item is handed over. Delivery person cant mark item as 'done' without the code. I've seen Uber Eats do the same thing sometimes.
Sometimes I wonder if true reason behind them is not any sort of verification but to maybe increase potential punishment if caught... So effectively nothing...
This is normal for every item above ~100EUR delivered by Amazon here in Austria. Never had it happen that they asked for the code without showing me the package in-person first though.
I've already been burned via amazon a few times recently.
Most recent one I bought a multi-tb hard disk, but got an old multi-gb disk shoehorned into legitimate box. Couldn't just get a replacement - had to buy one for more money since that drive's price had increased.
Depending on your location this will happen for any item with a certain value and presumably risk score. It's not a unique decision they made for RAM. Probably happened to your order because the value was high enough and your location, not because it was RAM specifically
The boom cycle looks like it's going to last a while, but it never lasts forever. If the current memory manufacturers all conspire to not build more capacity, I'm sure someone else will (eg China).
I just bought 16 GB DDR4 for my (somewhat old) Dell tower for $120 + shipping from Techyparts.com. Looking around that seemed like a really decent price (but maybe I'm wrong).
I had previously ordered from Aeonfly, but they cancelled my order--likely because they wanted to raise the price.
Don't worry. Sooner or later this stuff will be mandatory and you'll be obligated to get your state-issued terminal. Social credits will do, if you have no money.
For the skeptical ones: Somewhere people are already being cut off from state and financial institutions without proprietary software with bundled security certificates, applications are bound to collect data on the environment they are used in (other apps) and organizations deliberately limit access to their services or cripple them without their applications (can't do things from a generic web-browser).
Jeff Bezos will be happy to let you rent one from him. He's said this is how he envisions the future of personal computing. You will own nothing, and so on.
"The original retail price of the computer with 4 KiB of RAM was US$1,298 (equivalent to $6,900 in 2025)[21] and with the maximum 48 KiB of RAM, it was US$2,638 (equivalent to $14,020 in 2025)"
That's tech... Moore Law was steadily going until raw CPU got stuck for a few years, switching from single core to multicore then heterogeneous cpus, all those requiring to rewrite software in a scalable manner and then consider newer schedulers switching between performance/efficiency cores and changing CPU freqs.
Also, you have Meta and Google investing in simplified performant versions of their stack for developing countries which is similar.
What's funny though,
I see people that sometimes says it's cheaper to build the app they need in a single prompt than to search for it.
You know that Moore's law is about transistor count? What worked really well for CPU in eg the 1990s was Dennard scaling. That drove the Hz higher. And Dennard scaling is what stopped working.
They're heavily tuned but I think having the generational hypothesis and copying longer-lived objects is still standard? As is allowing a fair amount of garbage to accumulate to increase throughput and reduce copying at the cost of memory footprint?
Posting instead of researching in hopes someone smarter can chime in, because I'm lazy.
Price per GiB is skewed by definition. 10 Years ago, 8GiB was enough and cost between $10 and $20. For the longest time this century, a cheap build was in the $250 to $350 range. Nowadays I'm having trouble with that price range building from used parts.
Honestly, despite being Electron VSCode manages to outspeed every other editor on my machine (by qualitative metrics), even stuff like Zed (which manages to lag on stuff VSCode doesn’t despite performance being its main selling point)
Why? High demand and low supply is very good for shareholders and make indexes go up, their goal is not to meet demand, they want exponential qurartly revunes
I hear what you're saying but memory ultimately is not infrastructure, it's a commodity that's shipped over infrastructure.
The arguments for infrastructure being managed more tightly by the government is because it's a natural monopoly that everything depends on. Memory isn't a natural monopoly, if we found ourselves in a spot where we all depend on 3 companies then that's not a law of nature, we need more companies.
Its not like demand will be lowered by this though, I dont see how this would be bad for ASML (besides PR-China generally not respecting IP and them reverse engineering it)
Don't really need the EUVs for commodity DRAMs and China has plenty of DUVs capable of 10nm. CXMT's DDR5 is known to use 1z nodes (17nm) vs 1c (10nm) used by the big three -- ie, higher cost, less efficiency, less density.
The big three are also expanding their capacity. IMO, it seems to make more sense to sell equipments to those who could make better, efficient use of them. ASML has a backlog of €38.8 billion and is fully booked for 2027.
AI companies would do well to realize that civil disobedience is a thing.
Take what's currently happening to Flock cameras as an object lesson. People are so fed up with Flock that they're cutting their cameras down en masse, sometimes wiping them clean from entire jurisdictions. Tens of thousands of other people in those jurisdictions are cheering them on, even offering alibis for them before they've been caught. Law enforcement doesn't seem too keen to catch them either. You absolutely can be so hated that the law will not be enforced to protect your property.
If AI companies don't slow down and build some good will, AI data centres are probably at risk. They'd probably have been targeted before Flock were they not harder targets. A higher degree of cooperation and organization will be required to sabotage them, but it would be foolish to believe it won't eventually happen.
Monopolizing memory production for another year suggests that lessons have not been learned, the mad build-out will continue, and we're headed for some truly crazy stuff.
You can hear them from quite far away since companies like X use on-site Methane Gas Turbine Generators because they don't want to wait for the grid capacity to catch up.
It doesn't matter. The perception is all datacenters do AI (not necessarily wrong nowadays either) and that's all that matters to the public when they see so many lost jobs and high computer / phone prices.
> Take what's currently happening to Flock cameras as an object lesson. People are so fed up with Flock that they're cutting their cameras down en masse, sometimes wiping them clean from entire jurisdictions.
It seems inevitable that these lost cameras will simply be upgraded to drone fleets. Not an identical replacement, but still "good enough" for the original purposes. Had they anticipated this problem and skipped the first phase, it might have been harder to manufacture consent for privacy invasion. But now anybody who was okay with the original cameras feels like a victim having them torn down, which makes it likelier they'll be okay with drones everywhere.
> Tens of thousands of other people in those jurisdictions are cheering them on, even offering alibis for them before they've been caught. Law enforcement doesn't seem too keen to catch them either. You absolutely can be so hated that the law will not be enforced to protect your property.
At least one third of the country is still okay with them, because they believe it helps rid them of illegals and deter crime, with downsides small enough to ignore, in their mind.
In sum, we don't have as much control over these things as you seem to believe. Capital was always in control with its crude but effective propaganda pipelines to manufacture enough consent. Now imagine those pipelines becoming even more effective thanks to AI. Not looking good.
Who cares about those people? I would argue even they themselves hardly do. And sure, they'll be even more okay with even worse stuff, and that even worse stuff will be even more hated by even more people, whatever the people who are "okay with" (that can mean ignorance, obedience, or evil intent, but it's not justification). If they can be okay with mass surveillance they can be okay with other things, so they're basically a wash, they're just okay with stuff.
> At least one third of the country is still okay with them
Then they can be okay with the majority saying you know what, we're sick of you, you can either stop having your toys or you will be segregated from us and live under your own surveillance. "Being okay with something" isn't some magical wand or a fortress, it's just a limp shrug. It means "pass".
> In sum, we don't have as much control over these things as you seem to believe.
I don't see the connection with the sentence before that, which is a minority being "okay with" something and a majority willing to fight and out there doing it. It's not about controlling what a minority of people are okay with, it's about changing what is done. The people who are okay with stuff can go read a comic book until it's over, how does that not solve anything you raised about them? Or are you saying their being okay with stuff should be respected or even heeded? Couldn't they just make that claim themselves, if they wanted to?
Imagine a raging house fire, people stumbling over each other trying to help, and some guy steps up and goes "WAIT!" they all look at him, sure he has a major contribution for interrupting something like this, a good idea, a plan perhaps; and he goes "... I'm actually okay with this."
> Then they can be okay with the majority saying you know what, we're sick of you, you can either stop having your toys or you will be segregated from us and live under your own surveillance. "Being okay with something" isn't some magical wand or a fortress, it's just a limp shrug. It means "pass".
It's not the majority, though. Roughly speaking, among adults, we have:
* 1/3 who are willing to vote for the red flavor of lesser-two-evils (some of them don't even register it as evil)
* 1/3 who are willing to vote for the blue flavor of lesser-two-evils (some of them don't even register it as evil, including when it literally manufactures consent for the first third to win next time around)
* 1/3 who see the above as a uniparty of capital interests offering the illusion of democracy through "close" elections as the bullshit that it is.
That's a really great point. So much about our world is about perception. I think the datacenter concern is overblown for the most part (except putting them next to houses), but the perception is stronger than the facts.
And if AI is putting people out of jobs and making everything expensive - or that's just the perception - that's gonna set AI back for years.
Yup... And in 2021(5 years ago now) I got 64gb ddr4 for 120 bucks, second hand with a 6 month warranty. I felt like I hit the jackpot then, now that price seems like literal peanuts. I can't wait for the AI bubble to burst.
At this size, I think we have to start worrying about a 2008-type default risk. AI companies will have ordered all the RAM and paid some kind of deposit for that, but actually taking delivery and paying for it is still in the future and the sums of capital involved are extremely large.
Fail to deliver, and they get bankrupted because a court would not simply unwind the deal - a court would demand compensation of the market price of the goods not delivered on delivery day - which will probably be very high.
What we got in 2008 was central banks (eg Fed and ECB) willfully collapsing nominal GDP in their economies. Have a look at the dot-com bust or Black Monday for comparison.
We just had a string of bank failures, like Silicon Valley Bank, but you will not see any impact in the macro-economic data (like unemployment), because the Fed kept nominal spending on track.
In 2008 they didn't. Instead they actively tightened monetary policy by eg introducing interest on excess reserves. The ECB even hiked interest rates.
Companies defaulting on debt doesn't need to bring down the economy.
For the US, you can also see how the construction sector had been winding down for years (eg as measured in construction employment) without an impact on overall unemployment. The crisis was entirely avoidable.
Isn't scalping a term used for tickets? Also, these companies supposedly intend to actually use the RAM they ordered, not reselling at some inflated price.
It's all fair game IMHO, except the elephant in the room which is that the AI company bubble might pop.
I don't see the AI bubble popping any time soon (as in, not in terms of physical datacenter buildout - stock prices could fall). Even if progress hits a brick wall, we still have multiple years of just expanding out the models we have to more people in more industries.
And if anything, I believe progress has been accelerating.
A year ago the SOTA was GPT-5/Opus 4.1/Gemini 2.5 Pro, two years ago Sonnet 3.5.
Looking at Opus 5 for SOTA performance and GPT 5.6 Luna for a viable cheap alternative, AI is much more capable now.
Honorable mention: GPT 5.6 Sol on Cerebras, capacity limited to a few customers, is supposedly serving 750 tok/s.
Compared to 90 tok/s for non fast mode 5.6 Sol, 56 tok/s for Opus 5, and 190 tok/s for 5.6 Luna.
I am very curious about the next generation of models, GPT-6/Astra is rumored to launch still in August. Not sure what is the state of Anthropic's next Fable checkpoint.
If these models also deliver significant improvements, I really do not see how one could seriously still argue among the lines of AI being a scam, and the demand not being there to support the size of the investments.
> I really do not see how one could seriously still argue among the lines of AI being a scam, and the demand not being there to support the size of the investments.
I don’t think most people are saying it is purely a scam (well, some do but I don’t think they are to be taken seriously). What all these talks about circular financing and VC money are saying is that demand cannot sustain the sector long-term, not that there is no demand. People are certainly happy to pay say $20/month for whatever AI chatbot, but would they still pay if it were $200/month, which is closer to the actual costs.
There are several other details that point towards a unsustainable projections:
- measurable benefits from AI-ifying companies are nowhere near what is commonly believed. AI providers are hoping that they can keep the show going until the models are good enough, essentially faking it until they’ve made it, but that is not a given. It is also unstable because it is susceptible to change in public opinion.
- permits for new datacenters are not going to become easier to get as public opinion keeps turning against them. They will have to concentrate in friendly regions, which will add cost (more demand for the same location, plus interconnection for network and electricity, both of which can easily become bottlenecks).
- the electric grids are not ready for all those planned datacenters, so something will have to give. Building more and more on-site diesel generator in times where oil supply is so constrained and random is not very sustainable either.
- eventually the loans will come due and if earnings do not match there will be a, possibly severe, correction.
If we look at the historical example of the dot-com bubble, the crash was not caused by no demand. It was just caused by over-estimated demand and too much money going to a single sector of the economy. We still use the Internet, and it is still hugely important, but the correction was still severe and real people lost real money.
Those are good points, but some arguments are weak:
- Chatbot usage in $20 plans is likely nowhere close to a $200 cost.
Inference cost has come down rapidly, with reports from July claiming OpenAI can now serve all of the logged out ChatGPT traffic on just a few hundred GPUs.
- Measurable benefit: I doubt anything of value is being measured. MS Copilot with GPT 5.5 Instant processing SharePoint files? Developers using AI as a fancy autocomplete under a "I review every line" regime?
I rely on my personal value judgement, based on 30h/week I spend using AI outside of my regular job.
I am developing a mobile app, competing with companies with millions in revenue and entire dev teams. I know it is viable. Others lag in effective adoption, their opinion is likely to change soon with even more capable models.
- data center projects in the US: They look to me to mostly be constructed in the most remote backwater. If even there projects with such moderate environmental impact cannot be realized, that would be an embarrassing policy failure
- the grid: I think that one is true. AFAIK the constraint would be gas turbines, not diesel, and oil supply is not structurally constrained
- the loans: Anthropic is rumored to have become profitable earlier this year because of large growth in enterprise revenue. It does not look so bad to me
>I am developing a mobile app, competing with companies with millions in revenue and entire dev teams. I know it is viable. Others lag in effective adoption, their opinion is likely to change soon with even more capable models.
To add to that, as an ex-software dev that mostly works in a semi-unrelated field now (and can only code as a small part of my job): I think a lot of software devs are underestimating what someone with reasonable technical skills and specialised domain knowledge is able to create with AI.
I honestly wouldn't be surprised if, in 5 years' time, the majority of software used by (e.g.) potato farmers was primarily created by other potato farmers. The code might still be less elegant but I think it will be easier for the potato farmer to iterate with an AI than outsourcing to a dev firm.
On a tangent, because you mentioned five years for widely used software being implemented by domain experts rather than software engineers: this seems like a very conservative timeline.
I have little doubt that a competent potato farmer could possibly implement and sell such software today using Opus 5, Fable, or 5.6 Sol.
Now what I am going to say next may sound a bit crazy, and it is outside of my area of expertise. I hear the recent mathematical breakthroughs made by GPT-6/Astra are field medal worthy discoveries.
What if the current trajectory of improvement holds for another year or two?
Is it impossible that LLMs gain superhuman ability to reason over a large number of constraints so that they can make novel breakthroughs unimaginable to us today?
What I hope to see in five years is not potato farmers writing software, but programmable immune cells that safely kill cancer.
It is also a term use in general trading anywhere traders open and close their positions quickly to make profit off an upswing which might be in part caused by traders doing this. If done with inside knowledge it is illegal, in this case more commonly referred to as “front running” (which is what many suspect happened with numerous ahemfortuitous/serendipitous position changes around announcements of changes in the state of the US vs Iran situation).
> Also, these companies supposedly intend to actually use the RAM they ordered, not reselling at some inflated price.
As well as the companies ordering for the purpose of DC roll-outs and upgrades, there will definitely be some buying purely to sell at a higher price a short time later.
> It's all fair game IMHO, except the elephant in the room which is that the AI company bubble might pop.
[and with reference to the sibling reply to this: “Then we'll have a lot of cheap RAM.”]
I don't think it will pop with a sudden bang, and even if it is it won't affect these memory sales or probably those for 2028. The first real effect will be those using the AI tools suddenly finding themselves needing to pay a lot more for them as companies offering those services deal with reduced access to “new” credit from the circling investment pool shore up their fanciful accounting that way instead. The DC builds/upgrades that are currently planned will happen but more funded by end users and less by the investment cycle. New plans might not be actioned at the same scale, but it will take a while for that to affect component prices as there is a lot of lead time involved.
Water is one example, but it usually goes beyond that. I'm not sure if RAM is ever covered by those laws, but I'm also not willing to look through every state's laws and check.
It was already reported that new RAM production wouldn't be available until at least the end of 2027, and when it does come back online there'll be a backlog of orders, so prices and availability won't be stable until at least 2029. That's assuming there aren't more global economic and political issues delaying it.
China RAM entered the US market in quantity a week ago west coast. It’s slow buggy trash but it’ll cool the market until failure rates are known. Europe got its first shipment 3 weeks ago.
The lines that would have made consumer-level stuff are being reduced in favor of the higher profit margin/demand products. That means more supply of those other products (but still not enough) and lower supply of the consumer-level stuff (driving up prices).
Isn't DDR4 like, old and obsolete? From my knowledge, you can't just take a fab that produces DDR4 and make it produce DDR5 or higher. It's kind of a leap that requires replacing a lot of stuff.
I would expect that old DDR4 machinery to end up somewhere else, still chugging along. Maybe someone buys the old production and keeps chugging low quality DDR4 or something.
To me, it sounds like that didn't happened. Fabs that were producing legacy DDR4 are still producing it, just selling it higher and surfing on the consumer price increase of DDR5. As DDR5 became more expensive, people looked to build old AM4 PCs with the older memory, and the market just raised the price so that option would also cost more. IMHO DDR4 price increase has nothing to do with production.
There is some overlap. The last DDR4 fabs were typically on a 16nm process, the first DDR5 fabs were also on a 16nm process.
There shouldn't be much issue converting those 16nm DDR4 fabs to DDR5, it's really just an issue of swapping out the masks. That's assuming they haven't all already been converted years ago.
Even when there isn't an overlap, moving to the next process node usually isn't so much about replacing machines, but adding more machines. To oversimplify, you might not be able to convert one older DDR4 fab to DDR5, but sometimes you might be able to convert two older DDR4 fabs to one DDR5 fab.
Such conversions might not be the most cost-effective option in the long run (that unmodified fab could have kept selling DDR4 for years), but conversions might be significantly faster than building new fabs.
[not an expert so I could be totally wrong] I was also just talking about the physical space. Even if the machines on the entire line have to be replaced it still probably makes sense for a company to transition in situ - displacing the old line. There's the lease for the land and building, people's commutes are structured around a certain physical location, air conditioning and electricity are already there, etc... It's not just the big expensive machines that make up the fabrication - it's also the entire structure that makes that production happen.
Even if they need to replace every single lithography machine, there are a lot of other machines that probably don't need to be replaced for any given upgrade.
Like everything related to it being a clean room; All the logistics robots which move wafers around; All the machines that deposit layers of material like CVD/AVD/Sputtering; The various machines that do the actual etching; The furnaces; The inspection and quality control equipment;
Also, absolutely everything to do with dicing and packaging. But that's usually already at another factory (often a completely different company) due to how process agnostic it is. Same applies to growing and prepping the silicon wafer.
I am guessing there is some fluidity between these markets. If consumers get priced out of the DDR5 market, the consumer DDR4 price increases as well. It is the same with cars, if the price goes up for new cars, it usually increases for used car market as well. Same with energy, price goes up for oil, all things beeing equal, other energy carries will increase to some degree.
That's it. Which is insane. We used to buy used DDR4 from obsolete servers in buckets before it was going to become e-waste. Now even that one is expensive.
So.. memory futures? Sounds like a signal of an efficient and mature market to me, manufacturers sell their stock way beforehand, keep production steady, and wholesalers can hold stock and absorb risk in exchange of profit.
Nothing prevents wholesalers from selling their memory or memory futures at a loss, as that is a risk that is justified by equal possibiltiy of the memory appreciating in excess of inflation.
Why AI companies wont build their own RAM but prefer to pay multiple times more than what they sold pre AI boom? (And ram business was still profitable). They already make custom processors
AI companies aren't fabbing processors. They're buying capacity at fabs to make their own designs. To make more RAM, they'd need to invest in making more DRAM fabs, which is a slow process.
Lame: Fight the war on general computation by spending decades pushing elaborate technical standards for locking down bootloaders, try to convince society why we need device attestation and hardware-bound keys, gradually establish devices without root access as the norm, get into constant skirmishes with free software and privacy advocates, get drawn into unpleasant political quagmires, etc etc
One unit of HBM capacity consumes roughly the wafer capacity that could have produced three units of DDR5 capacity. HBM dies need to be larger than ordinary DRAM dies because of how the final packaging works.
> As discussed previously, the ramp of HBM production will constrain industry supply growth in non-HBM products. Industrywide, HBM3E consumes approximately three times the wafer supply as D5 to produce a given number of bits in the same technology node. With increased performance and packaging complexity, across the industry, we expect this trade ratio for HBM4 to be even higher than the trade ratio for HBM3E. We anticipate strong HBM demand due to AI, combined with increasing silicon intensity of the HBM roadmap, to contribute to tight supply conditions for DRAM across all end markets. As the memory industry is still recovering from the challenging environment in 2023, this tight supply environment will help drive the considerable improvements in profitability and ROI (return on investment) that are needed to enable the investments required to support future growth.
https://investors.micron.com/static-files/4550f98c-1054-4847...
>As the memory industry is still recovering from the challenging environment in 2023, this tight supply environment will help drive the considerable improvements in profitability and ROI (return on investment) that are needed to enable the investments required to support future growth.
The problem is that they need actual chips tomorrow. I fear we are reaching the point in the semiconductor industry where, in order to sustain the revenue growth propping up their valuations, they're going to have to start selling future chips that cannot possibly be physically produced.
The alternative is they could charge more. They could. And people would pay it.
The problem isn't the price. No matter how much you pay, you can't take delivery of a chip that doesn't exist. Now, you could raise prices to the point where you destroy demand. But that's a tricky window to maneuver through.
For consumers they’re already destroying demand
> The problem isn't the price. No matter how much you pay, you can't take delivery of a chip that doesn't exist
I’m confused because this reads like a denial of basic economics. If the price is high enough, the chip will be produced for you.
Do you mean because of the production lead time, higher prices won’t result in increased production? Commodities like corn have been managing this for a long time… what’s special about chips?
What is your actual argument?
It takes a billion dollars and years to produce a new chip fab line. So no, it cannot just be produced for you on a dime.
Corn is heavily subsidized,also has (in the US) federally prvided insurance programs against plunging prices, has paid uselessness(ethanol). Despite all of this corn has caused massive agriculture bankruptcy/takeover when it has failed anyway.
So it's more "what's special about corn". It is also fairly hilarious to claim the parent is denying basic economics and then bring up corn as an example of having successfully managed economics. If the scales were not being thumbed, and "basic economics" were in play, corn would be in very very bad shape.
In the case of DRAM, there is an incredibly long history of these gloom/glut cycles, and they have stayed roughly the same timeframes (~3 years) since the 1990's.
Almost all the ones who have survived this long are either in the same kind of boat as corn - protected in various forms from the downside - or don't increase production and get caught out until they are absoultely forced.
The very temporarily increased profit is not worth going bankrupt for - they make more money long term by being very cautious and know this.
There are a near infinite number of economic studies you could look at (and several sibling comments cite some) - DRAM manufactuers don't chase the price and probably couldn't anymore if they want to.
None of this denies basic economic theory, of course, since economic theory is not exactly "rigorous", even to the degree it could be (IE even the parts that are pure analysis of data rarely reproduce!).
It could be tricky. Or, since supply (in the short term) is pretty well established, they could just auction off chips.
If you can only make X number of widgets, you must increase the price until demand stabilizes at X.
This is what consumers are always told. Prices go up as demand goes up or scarcity increases.
But once you get past econ101, you are taught how markets react to percieved bubbles, how those with near monopolies would rather pocket a windfall now than risk investing in future expansions.
> HBM3E consumes approximately three times the wafer supply as D5 to produce a given number of bits in the same technology node
That's really interesting, and I wonder why? I believe HBM has redundant ECC bits by default, which would add a few %, but other than that, is it just that the yield is much lower due to die stacking? Of course, this is a 2024 document so things may have changed a bit since.
HBM requires stacking the chips. So they need to shave the layers, glue, stack more, shave again. They also require a substrate what is even more wafers. The issue is that a error in the stack means a lot of losses.
In order to get high bandwidth, you want memory as close as possible to the GPU. The more trace lane length = signal loss, bandwidth loss.. HBM is compact, and so you can stack 24GB modules, 8 around a GPU die.
If you tried to do that with normal memory, you need like 64 modules. So a a TON of traces more that all need to be equal length, and because so many = far away from the GPU = less bandwidth.
The issue is like stated above, its a process that waste a ton of wafers. Wafers that can make easily 3x more normal memory.
Intel with "Crescent Island" is trying to make a 160GB card using LPDDR5x memory but the bandwidth is only ~700GB/s.
My own PC died and now I’m locked out of my entire Steam library for the foreseeable future. A $2000 PC is literally a downgrade from what I bought 10 freaking years ago.
What the freaking hell.
What part of your PC died?
This is why I'm so hesitant to use AI. There are other reasons (for example: the journey is the fun part, not the destination; and I want to learn not have something done for me), but the pressure AI puts on memory and storage is too much for me. I do use it occasionally, mostly to keep tabs on it and for small problem solving (how do I do this one piece of a query in MySQL, since I'm not a MySQL DBA), but that's it for now, until this memory/storage issue is fixed.
Unfortunately, I can't not use AI because it provides the best speech recognition I've ever used and, going along with hand problems, it now allows me to write code after a three-year absence from that skill. I do miss it. I'd never do it again professionally in the large, but I have missed writing code a lot.
My solution is to see what four or five-year-old equipment I can buy that will let me run local LLMs. I may only get six or seven tokens per second out of an i7, but it's a start. And best of all, I can turn the machine off when I'm not using it.
IMO, Migrating to small-scale local LLMs would be a significant improvement over using data centers.
> Unfortunately, I can't not use AI because it provides the best speech recognition I've ever used and, going along with hand problems, it now allows me to write code after a three-year absence from that skill.
I'll second this. The combination of Whisper + LLM makes speech recognition fantastic. I occasionally have arm pain from typing, and this is a Godsend.
I don't use it to write code - but in my experience stuff like emails + docs was the greater source of pain (one generally types slower while coding).
> Migrating to small-scale local LLMs would be a significant improvement over using data centers.
This shift is probably inevitable, but it will vary significantly by region depending on prices for electricity. Look at, for example, the difference in fundamental homelab build recommendations between Germans and just about anyone else. Electricity prices in Germany are so high that even a now expensive Raspberry Pi or other ARM board is often preferred over Intel/AMD builds due to low power draw (especially low idle power draw), an effect that adds up for a machine running all the time over years.
With local LLMs and the GPUs to run it, especially if you want a model available to you all the time and can remote into your local network to use it whenever you want, there's no escaping much higher power draws, even at idle. Wherever electricity is expensive, the electric bill can be a prohibitive barrier.
Smaller solar setups might become a hedge against high power bills in the future it seems.
No-Load losses on the power conversion seems to be the real issue and is rarely mentioned and understood even less often. Plus the conversion eff. curve.
Eg: I shaved ~40W off the idle load on a server (250->210W) by doung nothing more than removing the redundant supply
You just reminded me of a funny story about redundant power supply.
A college buddy used to work at Motorola (I'm naming the company because they wouldn't mind this story being shared) back in the late 90's or early 2000's. They had redundant power to their campus, bought from two different companies, coming in on opposite sides of the campus, so that even if some backhoe operator cut a ground-based power line somewhere, they wouldn't lose power.
And yet, one morning, the power went off all across their campus. After a little investigation, they sent pretty much all their employees home at noon and told them "take the afternoon off, don't come back until tomorrow, you wouldn't be able to do any work anyway". Turns out that although the power lines came in at opposite sides of their campus, somewhere a few miles away both of the power lines feeding their campus ended up running through the same underground conduit. And yes, a backhoe had managed to cut that conduit and break both of the lines they depended on at the same time. They had a single, VERY non-obvious, point of failure, and the backhoe had unerringly homed in on that SPoF.
> IMO, Migrating to small-scale local LLMs would be a significant improvement over using data centers.
LLM serving is most efficient when you batch a lot of parallel requests together. Data center solutions also have the advantage of collecting queries from around the globe, so the hardware can be utilized around the clock.
Having everyone serve their own local LLMs would produce a lot more memory demand. Not less. The same memory would be idle most of the time, and when it was used it would be used for 1 person instead of a batch of requests.
There are other reasons to run local LLMs, but solving hardware demand problems is not one of them.
Yeah, local LLMs are probably an order of magnitude less efficient at a fixed level of "intelligence" if not more.
Inefficient per watt, yes, but local inference capacity is greatly underutilized in aggregate. If a model can run on a machine that already exists, that's a bunch of additional chips that don't need to be built.
Seems like we need a USB equivalent standard to RAM sticks. I have lots and lots of old (1-2GB), intercompatible sticks sitting. Sometimes you just need the space and not the speed.
Grab the cheapest Intel Optane M10 drive from eBay with a 1x PCI-E adapter from them too. 16GB of Swap for less than $20.
The M10 are over produced and only PCI-E 3.0 2x lane devices but 16GB of DDR5 is going to cost over $200.
You can do the same with USB-C M.2 adapters but the adapters cost more and eww USB for Swap, hope the device never drops.
Hey, I apologize for using you as tech support, but I found tons of those M10 sticks you are talking about, but what should I be looking for in terms of an adapter? All I can find are adapters listed as M.2:
https://www.amazon.com/GLOTRENDS-Adapter-Aluminum-Heatsink-P...
I believe the intention is for you to put it in the M.2
That's a 4 lane adapter, it'll work if you have a 4x, 8x, or 16x PCI-E slot open. This is a 1 lane adapter: https://www.ebay.com/itm/128010189845?var=429206655218
RAM over “USB” seems dangerously close to an SSD.
In any case DDR3 isn’t nearly as bad price-wise and would have much better energy efficiency than stacking 1–2 gig sticks.
No, we absolutely do not: https://en.wikipedia.org/wiki/USB
For starters, The slowest sticks of DDR(x) are often slower than DDR(x-1). The issue is never capacity, but rather, performance.
The real nonsense? USB? It's a mess. Pick a USB cable and buy it from ANY retailer, let's make it easier, buy a USB-C cable. What is the data rate (depends on cable quality and length), Does it support power delivery? If so, how many watts? (~5W requires a very different cable from ~230W), how do you know from simply looking at the cable? If you buy a cable, how can you tell what it supports by simply looking at the connector? Imagine having a box full of USB-C cables. Could you tell me how fast each of those cables are? (The cables themselves don't! Many don't have any markngs at all, and if they do, the markings could be fraudulent)
RAM does not have that issue. A stick fits or it doesn't. Sure, there are a huge range of speeds, however, that range has a ballpark (JEDEC defines the ballpark, the "cartels" make the memory and push out some faster stuff).
While there are definite exceptions (I was bitten by one recently), you can generally plug in a DDR5 DIMM and expect it to work in the system.
The same cannot be said for USB. Some USB cables ONLY deliver power. Some only work with certain devices. I've USB-C (!!!) cables that only work with the devices they are shipped with, and even more annoyingly, the both may be true! ASUS (!!!) ships MOTHERBOARDS that can only pair with vital hardware via very specific USB-C cables (Strix Hive) and "GLORIOUS" ain't so glorious. Their mouses warn you not to use other USB-C cables, and they are right, depending on the make/model/generation, you can brick your "glorious" peripheral.
No, USB nonsense needs to stay far away from anything else.
Also, the reason this is a huge issue is because the memory makers are cartels. Only a few of them exist, they gang up and bully EVERYONE and nobody has ever invested money to create a competitor due to this, except China, which of course means that the U.S. and portions of the E.U. are insta-banning/trying to insta-ban, even though it is really freaking hard to install spyware on a memory module.
USBism has a lot of elements similar to the "connect your agent to your data with this 'curl | sh' script and do anything. Maybe some friction and incompatibility is useful in some cases.
> Pick a USB cable and buy it from ANY retailer, let's make it easier, buy a USB-C cable.
Okay, let's assume it's a USB-C cable and not just something USB-C shaped that pretends to be one.
> Does it support power delivery?
Yes.
> If so, how many watts?
Always at least 60W (3A). Up to 240W if it's e-marked.
> What is the data rate
Depends whether it's a USB 2.0-only cable (480Mbps) or a complete cable (20Gbps). Possibly more in Thunderbolt or USB4 modes - cables that handle those will be marked (both visually and with an e-mark). In any case, pretty easy to check just by plugging things in.
I don't know why you got downvoted so much. There's no such thing as a 5 watt USB cable and people need to stop acting like there is. 60 is a very good baseline. There's no such thing as a power only USB cable either.
If we're counting fake products then we need to count those RGB sticks that have no RAM in them.
> Sometimes you just need the space and not the speed.
Your computer already supports that. It's called "swap".
Yeah but DDR3 was faster then and still is now - I can't imagine my M4 Mac having a slightly older version RAM would meaningfully affect me, but another 24GB of it (or whatever I have) would certainly be an improvement.
CPUs have L1, L2, L3+ caches too after all.
A better SSD can get up to a passable GB/s rate, but when you're waiting tens of microseconds instead of tens of nanoseconds you can't use it the same way. And a worse SSD is 10x slower in GB/s too, compared to old super cheap RAM.
Thats like saying I need a car with hotswappable engines. Sure. But you wouldn't want that car anymore.
This will have general inflationary consequences for consumer products (phones, consoles, laptops, etc.). On top of current uncertainties regarding oil and fertilizers, I think 2% inflation in the US and Europe would be a very optimistic target.
Yep. And not in the future, it is here right now. And poor countries will as always be impacted way more than the US or Europe. The AI industry is the very direct cause and should be shamed for handling their rollout in such a destructing way. They deliberately created and grew the feeling of urgency for their own benefit
2% was already optimistic once we decided we were buying too much from China.
Literally put a 50% tax on everything and then said it'll only raise prices 2%
And once we also decided that we needed to attack Iran.
Yes, but the impact will be small.
The average consumer isn't purchasing a lot of products with a lot of RAM every year.
Their 8GB of RAM phones will go up in price a little bit, but people aren't buying phones every year or even every other year.
So if the price of the 8GB of LPDDR went from $40 to $160 and it's all passed on to the consumer buying a new phone every 4 years, that's an extra $30/year in spending.
Most adults I know now use their phone for everything, so they're not buying new computers and laptops. They can wait 3 years for the market to settle before upgrading those, too.
Even I'm a heavy buyer of tech products and RAM, and I would bet you that my family's annual food bill fluctuates by more than what I've had to pay for RAM prices growing.
Almost everything we consume is downstream of RAM pricing. I don’t buy a phone every year, but corporate workforces, streaming servers, cars, factories, etc. will all cost more next year.
Oh, I hope you are right about cars. If one of the outcomes of the memory crunch is that manufacturers start making “dumb cars” again, it would almost make it all worthwhile.
At the low end, where the vast majority of people are, even in the US, an extra $120 will double the price.
You forgot to tack on the rise in prices of SSDs too, so that's $50 per year at least.
>Most adults I know now use their phone for everything
You mean the cloud for everything. A lot of phone tasks share their computational and storage workloads off in the invisible ether that has actual computers with real costs behind them. Amazon has no problem with bumping up their costs in order to pay for their fleet of servers. I mean, what business doesn't use the cloud these days.
The headline says "sold out". That doesn't mean prices will go up, that means the memory will not be available at all for many companies and the products they make will not be available to consumers.
You don't have to directly buy RAM to be affected. It's just like oil or electricity...if RAM was part of the economic process that you participated in, then that price increase will be passed down to you. If every datacenter in the US suddenly had to pay double for electricity, how would that affect every operation that is operating in that datacenter? Ok now it's not electricity, it's the RAM.
Remember when car companies couldn’t get the chips they needed to make cars? And so now there’s not enough used cars. And so car companies charge way more the new ones and just stopped making cheap cars altogether?
Why are “demand and supply driven price increases” considered inflationary?
Isn’t it just… prices?
It's inflationary because this demand is backed by Monopoly money. OpenAI is promising to buy half the global supply of RAM with money they don't actually have, pricing out people who would like to make purchases with real money and not $1 trillion in debt.
Any price increase can be considered inflationary. They usually don’t come down by all that much
Prices rising is inflation.
Hmm. Forget about desktops. I'm programming a microcontroller which is actually reasonably priced and has gotten all its 32Mb of RAM in the same die as the processor, but I'm getting nervous and having a strong urge to stockpile...
Well, that one is likely out of danger - so far PSRAM (I assume that's what you have) is not impacted that much. However, we had to cancel some projects because we couldn't secure enough chips which have DDR on die at a reasonable price. They are 4x now compared to December last year.
Amazon has implemented a password system for RAM delivery. You're supposed to give a password to the deliverer before taking delivery of the RAM, and you're not supposed to send the password over the phone at all (although the Amazon delivery guy still called me and asked for it over the phone).
can u provide some context? what is this trying to solve and how does it solve it? very curious.
Its a way to make sure the delivery takes place. A code is sent to recipient and recipient only gives it to the delivery person once item is handed over. Delivery person cant mark item as 'done' without the code. I've seen Uber Eats do the same thing sometimes.
What’s wrong with a signature?
It's to prevent theft by the delivery person or by the wrong recipient
Signatures are terrible, and especially it's hard to automatically confirm they're correct even if you have a sample on file.
Signatures are completely pointless. Literally draw squiggles as anyone else can. Nobody ever checks or compares them to the original, how can they?
Sometimes I wonder if true reason behind them is not any sort of verification but to maybe increase potential punishment if caught... So effectively nothing...
Wow that's actually a really good idea to prevent delivery theft by the delivery person!
This is quite common for all high value deliveries across all platforms in India.
This is normal for every item above ~100EUR delivered by Amazon here in Austria. Never had it happen that they asked for the code without showing me the package in-person first though.
They did the same for a Samsung Fold phone delivered to me last month (in the US).
I've already been burned via amazon a few times recently.
Most recent one I bought a multi-tb hard disk, but got an old multi-gb disk shoehorned into legitimate box. Couldn't just get a replacement - had to buy one for more money since that drive's price had increased.
Depending on your location this will happen for any item with a certain value and presumably risk score. It's not a unique decision they made for RAM. Probably happened to your order because the value was high enough and your location, not because it was RAM specifically
This is a shame, because now I cannot order from Amazon as orders go to building reception and nobody is going to remember the pin to take delivery.
Keep an eye on this page → https://ca.pcpartpicker.com/trends/price/memory/
Note to the momentarily shocked: this is in Canadian dollars.
It's not just currency, Amazon Canada have some jumps recently not in US version:
https://pcpartpicker.com/trends/price/memory/
Holy moly that's just unbelievable when shown in graph form.
also: https://duckduckgo.com/?q=MU&ia=web
The boom cycle looks like it's going to last a while, but it never lasts forever. If the current memory manufacturers all conspire to not build more capacity, I'm sure someone else will (eg China).
CXMT is now producing as much as one of the big three. Only for China, of course.
They are producing about a third of what Micron (smallest of big three) produces, and plan on being on par with them in ~4 years.
That’s less that China needs to buy from the rest of the world, so it’s still a win.
I just bought 16 GB DDR4 for my (somewhat old) Dell tower for $120 + shipping from Techyparts.com. Looking around that seemed like a really decent price (but maybe I'm wrong).
I had previously ordered from Aeonfly, but they cancelled my order--likely because they wanted to raise the price.
How am I going to buy a computer that connects to AI datacenters now?
I'm sure some enterprising company will find a way sell you a computer subsidized by your data usage or maybe by forfeiting your first born
Don't worry. Sooner or later this stuff will be mandatory and you'll be obligated to get your state-issued terminal. Social credits will do, if you have no money.
For the skeptical ones: Somewhere people are already being cut off from state and financial institutions without proprietary software with bundled security certificates, applications are bound to collect data on the environment they are used in (other apps) and organizations deliberately limit access to their services or cripple them without their applications (can't do things from a generic web-browser).
Yes and if it's true it's usually downvoted or flagged on HN
> For the skeptical ones
The words after this point make you less convincing, not more, as they only serve to confirm the egregious slippery slope of your reasoning.
Do you call any negative prognosis based on extrapolation of an observable downward trend a slippery slope?
Jeff Bezos will be happy to let you rent one from him. He's said this is how he envisions the future of personal computing. You will own nothing, and so on.
I believe you're just supposed to stream your video and audio up through your Meta Glasses.
Or lease a laptop from Apple if faceputer's aren't as much your style.
Time to go dig an Apple II out of the landfill
From the "Apple II (original)" Wikipedia article:
"The original retail price of the computer with 4 KiB of RAM was US$1,298 (equivalent to $6,900 in 2025)[21] and with the maximum 48 KiB of RAM, it was US$2,638 (equivalent to $14,020 in 2025)"
All those devs who were saying "ram is cheap" BTFO.
That's tech... Moore Law was steadily going until raw CPU got stuck for a few years, switching from single core to multicore then heterogeneous cpus, all those requiring to rewrite software in a scalable manner and then consider newer schedulers switching between performance/efficiency cores and changing CPU freqs.
Also, you have Meta and Google investing in simplified performant versions of their stack for developing countries which is similar.
What's funny though, I see people that sometimes says it's cheaper to build the app they need in a single prompt than to search for it.
You know that Moore's law is about transistor count? What worked really well for CPU in eg the 1990s was Dennard scaling. That drove the Hz higher. And Dennard scaling is what stopped working.
Moore's law is alive and well.
It also was just an observation, not a strict rule set in stone
Bad times for the two-space copying garbage collectors.
Are there any of those left? The popular VMs for Java, Javascript and dotnet are all very mature heavily tuned designs.
They're heavily tuned but I think having the generational hypothesis and copying longer-lived objects is still standard? As is allowing a fair amount of garbage to accumulate to increase throughput and reduce copying at the cost of memory footprint?
Posting instead of researching in hopes someone smarter can chime in, because I'm lazy.
My benchmark:
1) .net app, one text field, one button; private bytes 22mb, working set 27mb
2) native app, two text fields, three buttons; private bytes 1.2mb, working set 7.3mb
3) electron, about:blank, 1.2GiB
electron, about:blank, 90mb
Just tested
Seconded, the only 1.2 GB electron app I’ve made had a ~500MB ML model’s weights embedded in it
Thanks to AI, we can now one-shot convert electron to native apps ;-)
Recipe for delicious pizza:
1. Make extra greasy pizza. Nauseating-level of grease.
2. Put pizza through De-Greasinator 5000.
Sounds efficient.
The grease is vaporised and inhaled by one randomly chosen Memphis resident, who permanently gets chronic asthma.
"it converted my gambling app into a three-armed bandit!"
Yeah would someone like to explain why there's an octopus in the break room?
Maybe they should start with Claude.
https://dam.stanford.edu/memory-prices.html
Pretty sure those devs were saying the same thing way before 2017 as well, which seems to be ~ the last time RAM was this expensive.
Now RAM in the cloud, now you're really paying the Java premium.
According to your link, DDR5 today has the same price in real US $ as DDR3 was in 2014, during a brief spike towards high prices.
Since that peak in 2014 until today, the prices were lower than now. The peak in 2017 was more than 10% lower than today.
To reach permanently higher prices than today, we must go backwards until 2011.
So we have already regressed at least 12 years into the past, but more likely 15 years, and it is unknown how much more we will regress.
> we have already regressed at least...
It seems worth noting here that DDR5 RAM is faster than DDR3 RAM, and was not even available until 2020.
Price per GiB is skewed by definition. 10 Years ago, 8GiB was enough and cost between $10 and $20. For the longest time this century, a cheap build was in the $250 to $350 range. Nowadays I'm having trouble with that price range building from used parts.
OK I’ll take the bait.
Were devs supposed to optimize RAM for a shortage that might come? I’m guessing you had enough foresight to stockpile RAM?
What’s implied here is that devs were supposed to optimize for user resources, not for developer convenience.
Maybe we can have AI rewrite those programs and optimize for low ram usage
They were supposed to optimize for the situation in which they weren't the only program running on a computer, and they failed to do it.
Can you blame them? Users didn't care. Heck, developers didn't care - look at how popular vscode is.
Revealed preferences aside, every dev comment section for 10 years HAS been full of people loudly insulting Electron.
Honestly, despite being Electron VSCode manages to outspeed every other editor on my machine (by qualitative metrics), even stuff like Zed (which manages to lag on stuff VSCode doesn’t despite performance being its main selling point)
Time to bring back 32 bit OS, that way you only need 4GB RAM per computer!
I saw on Twitter that TSMC has made 1B worth of chips for Apple that they cannot package because of missing memory.
(PS: It is time to let ASML sell to CXMT.)
> (PS: It is time to let ASML sell to CXMT.)
Why? High demand and low supply is very good for shareholders and make indexes go up, their goal is not to meet demand, they want exponential qurartly revunes
It’s temporary revenue, it’s not as if those businesses are suddenly fundamentally better.
To me it seems like they’re just riding the AI hype out, expecting a crash after, and not willing to risk a lot of capital on someone else’s bubble.
Temporary revenues for decades to come.
That’s very much up for debate.
True - but what is happening now is very destructive to a lot of important companies.
Indeed. There's a reason the US government has historically stopped air traffic / railroad / longshore strikes: cascading effects.
If memory is now upstream of so many companies, then maybe a higher regulatory involvement would be appropriate (like power companies).
Specifically, and it could fix all of this, capping the max % of their supply they're allowed to sell into the AI space.
I hear what you're saying but memory ultimately is not infrastructure, it's a commodity that's shipped over infrastructure.
The arguments for infrastructure being managed more tightly by the government is because it's a natural monopoly that everything depends on. Memory isn't a natural monopoly, if we found ourselves in a spot where we all depend on 3 companies then that's not a law of nature, we need more companies.
The government interferes with many commodities. Food, oil, rare-earth elements, sugar, airplane frames, etc
Its not like demand will be lowered by this though, I dont see how this would be bad for ASML (besides PR-China generally not respecting IP and them reverse engineering it)
You still have to actually meet the demand. It's not good if you don't sell any products.
I just thought… didn’t Intel start as a memory chip company?
Couldn’t intel repurpose some of its fabs (maybe the old ones?) to produce memory?
I’m no expert but… seems like an easy win?
You want a bunch of 4 kilobyte DRAM chips with an interface that predates DDR1, for your Commodore 64?
I thought CXMT already have plenty of capacity - Apple are currently talking to the US government for blessing to buy from them.
Other manufactures like Micron do have aggressive expansion plans, but fabs take a long time to build and bring online - multiple years.
CXMT is sold out too.
https://www.guru3d.com/story/cxmt-dram-capacity-reportedly-s...
Don't really need the EUVs for commodity DRAMs and China has plenty of DUVs capable of 10nm. CXMT's DDR5 is known to use 1z nodes (17nm) vs 1c (10nm) used by the big three -- ie, higher cost, less efficiency, less density.
The big three are also expanding their capacity. IMO, it seems to make more sense to sell equipments to those who could make better, efficient use of them. ASML has a backlog of €38.8 billion and is fully booked for 2027.
We are hammering our foot by investing in AI at this point.
AI grows and grows and becomes god and kills us all.
How did it all start? Oh we invested all our money in AI.
Queue the blog posts:
We rewrote all our Python code in Go because the RAMs too damn high!
Turns out AI-maxxing implicates citizens-minning.
AI companies would do well to realize that civil disobedience is a thing.
Take what's currently happening to Flock cameras as an object lesson. People are so fed up with Flock that they're cutting their cameras down en masse, sometimes wiping them clean from entire jurisdictions. Tens of thousands of other people in those jurisdictions are cheering them on, even offering alibis for them before they've been caught. Law enforcement doesn't seem too keen to catch them either. You absolutely can be so hated that the law will not be enforced to protect your property.
If AI companies don't slow down and build some good will, AI data centres are probably at risk. They'd probably have been targeted before Flock were they not harder targets. A higher degree of cooperation and organization will be required to sabotage them, but it would be foolish to believe it won't eventually happen.
Monopolizing memory production for another year suggests that lessons have not been learned, the mad build-out will continue, and we're headed for some truly crazy stuff.
Except nobody knows where the AI data centers are actually located.
You can hear them from quite far away since companies like X use on-site Methane Gas Turbine Generators because they don't want to wait for the grid capacity to catch up.
It doesn't matter. The perception is all datacenters do AI (not necessarily wrong nowadays either) and that's all that matters to the public when they see so many lost jobs and high computer / phone prices.
And how does the public know there is a datacenter inside a random building?
Either it's being proposed or other people tell them about them. It's not hard to find out where datacenters are.
The local government would need to know, and that would come out in town meetings.
Browse Google Maps aerial view and look for the backup generators.
Every produce storage facility going to have them, won't it?
No. It’s not economical for produce to have backup generators. Better to just move produce or let it go.
"That's not a datacenter. It's just a regular old building consuming 20+MW of electricity"
here's one
https://www.datacentermap.com/usa/michigan/ann-arbor/related...
> Take what's currently happening to Flock cameras as an object lesson. People are so fed up with Flock that they're cutting their cameras down en masse, sometimes wiping them clean from entire jurisdictions.
It seems inevitable that these lost cameras will simply be upgraded to drone fleets. Not an identical replacement, but still "good enough" for the original purposes. Had they anticipated this problem and skipped the first phase, it might have been harder to manufacture consent for privacy invasion. But now anybody who was okay with the original cameras feels like a victim having them torn down, which makes it likelier they'll be okay with drones everywhere.
> Tens of thousands of other people in those jurisdictions are cheering them on, even offering alibis for them before they've been caught. Law enforcement doesn't seem too keen to catch them either. You absolutely can be so hated that the law will not be enforced to protect your property.
At least one third of the country is still okay with them, because they believe it helps rid them of illegals and deter crime, with downsides small enough to ignore, in their mind.
In sum, we don't have as much control over these things as you seem to believe. Capital was always in control with its crude but effective propaganda pipelines to manufacture enough consent. Now imagine those pipelines becoming even more effective thanks to AI. Not looking good.
> anybody who was okay with the original cameras
Who cares about those people? I would argue even they themselves hardly do. And sure, they'll be even more okay with even worse stuff, and that even worse stuff will be even more hated by even more people, whatever the people who are "okay with" (that can mean ignorance, obedience, or evil intent, but it's not justification). If they can be okay with mass surveillance they can be okay with other things, so they're basically a wash, they're just okay with stuff.
> At least one third of the country is still okay with them
Then they can be okay with the majority saying you know what, we're sick of you, you can either stop having your toys or you will be segregated from us and live under your own surveillance. "Being okay with something" isn't some magical wand or a fortress, it's just a limp shrug. It means "pass".
> In sum, we don't have as much control over these things as you seem to believe.
I don't see the connection with the sentence before that, which is a minority being "okay with" something and a majority willing to fight and out there doing it. It's not about controlling what a minority of people are okay with, it's about changing what is done. The people who are okay with stuff can go read a comic book until it's over, how does that not solve anything you raised about them? Or are you saying their being okay with stuff should be respected or even heeded? Couldn't they just make that claim themselves, if they wanted to?
Imagine a raging house fire, people stumbling over each other trying to help, and some guy steps up and goes "WAIT!" they all look at him, sure he has a major contribution for interrupting something like this, a good idea, a plan perhaps; and he goes "... I'm actually okay with this."
> Then they can be okay with the majority saying you know what, we're sick of you, you can either stop having your toys or you will be segregated from us and live under your own surveillance. "Being okay with something" isn't some magical wand or a fortress, it's just a limp shrug. It means "pass".
It's not the majority, though. Roughly speaking, among adults, we have:
* 1/3 who are willing to vote for the red flavor of lesser-two-evils (some of them don't even register it as evil)
* 1/3 who are willing to vote for the blue flavor of lesser-two-evils (some of them don't even register it as evil, including when it literally manufactures consent for the first third to win next time around)
* 1/3 who see the above as a uniparty of capital interests offering the illusion of democracy through "close" elections as the bullshit that it is.
>upgraded to drone fleets
I wouldn't do that unless you want your airspace filled with birdshot
That's a really great point. So much about our world is about perception. I think the datacenter concern is overblown for the most part (except putting them next to houses), but the perception is stronger than the facts.
And if AI is putting people out of jobs and making everything expensive - or that's just the perception - that's gonna set AI back for years.
All of you that are excited for the m5 mac studios are going to be so bummed.
Yup... And in 2021(5 years ago now) I got 64gb ddr4 for 120 bucks, second hand with a 6 month warranty. I felt like I hit the jackpot then, now that price seems like literal peanuts. I can't wait for the AI bubble to burst.
This hopefully helps mitigate the risk of putting more production capacity in place, which the Big 3 were not willing to do so far.
At this size, I think we have to start worrying about a 2008-type default risk. AI companies will have ordered all the RAM and paid some kind of deposit for that, but actually taking delivery and paying for it is still in the future and the sums of capital involved are extremely large.
Also there is a ton of debt flying around, its not just equities hurting VC funds, real banks are facing risks now.
That's the reason RAM companies aren't expanding. Worst case, they can sell their normal production to gamers.
Fail to deliver, and they get bankrupted...
Do deliver and they don't get paid because the buyer purchased with money that doesn't exist.
Lose-lose situation!
But they don't get bankrupted, because it's an ordinary year for them, only with different customers.
Fail to deliver, and they get bankrupted because a court would not simply unwind the deal - a court would demand compensation of the market price of the goods not delivered on delivery day - which will probably be very high.
What kind of default risk?
What we got in 2008 was central banks (eg Fed and ECB) willfully collapsing nominal GDP in their economies. Have a look at the dot-com bust or Black Monday for comparison.
> willfully collapsing nominal GDP
What on earth are you talking about?
The credit risk was a very real problem; Kaupthing, Anglo Irish, RBS, Lehman etc.
We just had a string of bank failures, like Silicon Valley Bank, but you will not see any impact in the macro-economic data (like unemployment), because the Fed kept nominal spending on track.
In 2008 they didn't. Instead they actively tightened monetary policy by eg introducing interest on excess reserves. The ECB even hiked interest rates.
Companies defaulting on debt doesn't need to bring down the economy.
For the US, you can also see how the construction sector had been winding down for years (eg as measured in construction employment) without an impact on overall unemployment. The crisis was entirely avoidable.
See https://www.cato-unbound.org/2009/09/14/scott-sumner/real-pr... for a bit more background.
Isn't scalping illegal in some countries?
Isn't scalping a term used for tickets? Also, these companies supposedly intend to actually use the RAM they ordered, not reselling at some inflated price.
It's all fair game IMHO, except the elephant in the room which is that the AI company bubble might pop.
I don't see the AI bubble popping any time soon (as in, not in terms of physical datacenter buildout - stock prices could fall). Even if progress hits a brick wall, we still have multiple years of just expanding out the models we have to more people in more industries.
And if anything, I believe progress has been accelerating.
A year ago the SOTA was GPT-5/Opus 4.1/Gemini 2.5 Pro, two years ago Sonnet 3.5.
Looking at Opus 5 for SOTA performance and GPT 5.6 Luna for a viable cheap alternative, AI is much more capable now.
Honorable mention: GPT 5.6 Sol on Cerebras, capacity limited to a few customers, is supposedly serving 750 tok/s.
Compared to 90 tok/s for non fast mode 5.6 Sol, 56 tok/s for Opus 5, and 190 tok/s for 5.6 Luna.
I am very curious about the next generation of models, GPT-6/Astra is rumored to launch still in August. Not sure what is the state of Anthropic's next Fable checkpoint.
If these models also deliver significant improvements, I really do not see how one could seriously still argue among the lines of AI being a scam, and the demand not being there to support the size of the investments.
> I really do not see how one could seriously still argue among the lines of AI being a scam, and the demand not being there to support the size of the investments.
I don’t think most people are saying it is purely a scam (well, some do but I don’t think they are to be taken seriously). What all these talks about circular financing and VC money are saying is that demand cannot sustain the sector long-term, not that there is no demand. People are certainly happy to pay say $20/month for whatever AI chatbot, but would they still pay if it were $200/month, which is closer to the actual costs.
There are several other details that point towards a unsustainable projections:
- measurable benefits from AI-ifying companies are nowhere near what is commonly believed. AI providers are hoping that they can keep the show going until the models are good enough, essentially faking it until they’ve made it, but that is not a given. It is also unstable because it is susceptible to change in public opinion.
- permits for new datacenters are not going to become easier to get as public opinion keeps turning against them. They will have to concentrate in friendly regions, which will add cost (more demand for the same location, plus interconnection for network and electricity, both of which can easily become bottlenecks).
- the electric grids are not ready for all those planned datacenters, so something will have to give. Building more and more on-site diesel generator in times where oil supply is so constrained and random is not very sustainable either.
- eventually the loans will come due and if earnings do not match there will be a, possibly severe, correction.
If we look at the historical example of the dot-com bubble, the crash was not caused by no demand. It was just caused by over-estimated demand and too much money going to a single sector of the economy. We still use the Internet, and it is still hugely important, but the correction was still severe and real people lost real money.
Those are good points, but some arguments are weak:
- Chatbot usage in $20 plans is likely nowhere close to a $200 cost.
Inference cost has come down rapidly, with reports from July claiming OpenAI can now serve all of the logged out ChatGPT traffic on just a few hundred GPUs.
- Measurable benefit: I doubt anything of value is being measured. MS Copilot with GPT 5.5 Instant processing SharePoint files? Developers using AI as a fancy autocomplete under a "I review every line" regime?
I rely on my personal value judgement, based on 30h/week I spend using AI outside of my regular job.
I am developing a mobile app, competing with companies with millions in revenue and entire dev teams. I know it is viable. Others lag in effective adoption, their opinion is likely to change soon with even more capable models.
- data center projects in the US: They look to me to mostly be constructed in the most remote backwater. If even there projects with such moderate environmental impact cannot be realized, that would be an embarrassing policy failure
- the grid: I think that one is true. AFAIK the constraint would be gas turbines, not diesel, and oil supply is not structurally constrained
- the loans: Anthropic is rumored to have become profitable earlier this year because of large growth in enterprise revenue. It does not look so bad to me
>I am developing a mobile app, competing with companies with millions in revenue and entire dev teams. I know it is viable. Others lag in effective adoption, their opinion is likely to change soon with even more capable models.
To add to that, as an ex-software dev that mostly works in a semi-unrelated field now (and can only code as a small part of my job): I think a lot of software devs are underestimating what someone with reasonable technical skills and specialised domain knowledge is able to create with AI.
I honestly wouldn't be surprised if, in 5 years' time, the majority of software used by (e.g.) potato farmers was primarily created by other potato farmers. The code might still be less elegant but I think it will be easier for the potato farmer to iterate with an AI than outsourcing to a dev firm.
On a tangent, because you mentioned five years for widely used software being implemented by domain experts rather than software engineers: this seems like a very conservative timeline.
I have little doubt that a competent potato farmer could possibly implement and sell such software today using Opus 5, Fable, or 5.6 Sol.
Now what I am going to say next may sound a bit crazy, and it is outside of my area of expertise. I hear the recent mathematical breakthroughs made by GPT-6/Astra are field medal worthy discoveries.
What if the current trajectory of improvement holds for another year or two?
Is it impossible that LLMs gain superhuman ability to reason over a large number of constraints so that they can make novel breakthroughs unimaginable to us today?
What I hope to see in five years is not potato farmers writing software, but programmable immune cells that safely kill cancer.
Perhaps I am an optimist.
Then we'll have a lot of cheap RAM.
Not sure if HBM can be used as regular DDR.
> Isn't scalping a term used for tickets?
It is also a term use in general trading anywhere traders open and close their positions quickly to make profit off an upswing which might be in part caused by traders doing this. If done with inside knowledge it is illegal, in this case more commonly referred to as “front running” (which is what many suspect happened with numerous ahemfortuitous/serendipitous position changes around announcements of changes in the state of the US vs Iran situation).
> Also, these companies supposedly intend to actually use the RAM they ordered, not reselling at some inflated price.
As well as the companies ordering for the purpose of DC roll-outs and upgrades, there will definitely be some buying purely to sell at a higher price a short time later.
> It's all fair game IMHO, except the elephant in the room which is that the AI company bubble might pop.
[and with reference to the sibling reply to this: “Then we'll have a lot of cheap RAM.”]
I don't think it will pop with a sudden bang, and even if it is it won't affect these memory sales or probably those for 2028. The first real effect will be those using the AI tools suddenly finding themselves needing to pay a lot more for them as companies offering those services deal with reduced access to “new” credit from the circling investment pool shore up their fanciful accounting that way instead. The DC builds/upgrades that are currently planned will happen but more funded by end users and less by the investment cycle. New plans might not be actioned at the same scale, but it will take a while for that to affect component prices as there is a lot of lead time involved.
Buying and selling at market price is rarely illegal. I think in the USSR it might be?
It was legal in USSR for USSR to scalp though, depends on point of view.
Ha ha, the USSR was a silly thing. Anyway, what was the topic about?
I think it was about the failures of centrally planned 5-year plans of RAM sales.
In the US, selling at the (new higher) market price during a declared state-of-emergency is illegal in most states and territories.
Are talking about selling marked up water during a state of emergency? Don’t think that would apply here.
Water is one example, but it usually goes beyond that. I'm not sure if RAM is ever covered by those laws, but I'm also not willing to look through every state's laws and check.
It was already reported that new RAM production wouldn't be available until at least the end of 2027, and when it does come back online there'll be a backlog of orders, so prices and availability won't be stable until at least 2029. That's assuming there aren't more global economic and political issues delaying it.
But Micron is overpriced and new capacity is about to come online right? Right???
China RAM entered the US market in quantity a week ago west coast. It’s slow buggy trash but it’ll cool the market until failure rates are known. Europe got its first shipment 3 weeks ago.
Honest question: is this the first time this thing happens?
Besides the AI craze, to me it sounds kinda reasonable. Most large buyers (apple, public clouds, OEMs) do forecasting and orders in advance, right?
According to semiconductor industry stock market analysts, memory gets really expensive about once a decade on average so far.
Memory prices are cyclical. But this is bigger and longer than most cycles.
The fact that consumer-level DDR4 is getting more expensive makes absolutely no market sense to me.
The lines that would have made consumer-level stuff are being reduced in favor of the higher profit margin/demand products. That means more supply of those other products (but still not enough) and lower supply of the consumer-level stuff (driving up prices).
Isn't DDR4 like, old and obsolete? From my knowledge, you can't just take a fab that produces DDR4 and make it produce DDR5 or higher. It's kind of a leap that requires replacing a lot of stuff.
I would expect that old DDR4 machinery to end up somewhere else, still chugging along. Maybe someone buys the old production and keeps chugging low quality DDR4 or something.
To me, it sounds like that didn't happened. Fabs that were producing legacy DDR4 are still producing it, just selling it higher and surfing on the consumer price increase of DDR5. As DDR5 became more expensive, people looked to build old AM4 PCs with the older memory, and the market just raised the price so that option would also cost more. IMHO DDR4 price increase has nothing to do with production.
There is some overlap. The last DDR4 fabs were typically on a 16nm process, the first DDR5 fabs were also on a 16nm process.
There shouldn't be much issue converting those 16nm DDR4 fabs to DDR5, it's really just an issue of swapping out the masks. That's assuming they haven't all already been converted years ago.
Even when there isn't an overlap, moving to the next process node usually isn't so much about replacing machines, but adding more machines. To oversimplify, you might not be able to convert one older DDR4 fab to DDR5, but sometimes you might be able to convert two older DDR4 fabs to one DDR5 fab.
Such conversions might not be the most cost-effective option in the long run (that unmodified fab could have kept selling DDR4 for years), but conversions might be significantly faster than building new fabs.
[not an expert so I could be totally wrong] I was also just talking about the physical space. Even if the machines on the entire line have to be replaced it still probably makes sense for a company to transition in situ - displacing the old line. There's the lease for the land and building, people's commutes are structured around a certain physical location, air conditioning and electricity are already there, etc... It's not just the big expensive machines that make up the fabrication - it's also the entire structure that makes that production happen.
Even if they need to replace every single lithography machine, there are a lot of other machines that probably don't need to be replaced for any given upgrade.
Like everything related to it being a clean room; All the logistics robots which move wafers around; All the machines that deposit layers of material like CVD/AVD/Sputtering; The various machines that do the actual etching; The furnaces; The inspection and quality control equipment;
Also, absolutely everything to do with dicing and packaging. But that's usually already at another factory (often a completely different company) due to how process agnostic it is. Same applies to growing and prepping the silicon wafer.
> Isn't DDR4 like, old and obsolete?
I'm posting this from a machine that uses DDR3 and is still perfectly capable of meeting my needs.
>Isn't DDR4 like, old and obsolete?
I built a new gaming PC with DDR4 two months ago and it's pretty great. For servers RAM speed usually matters even less.
Because the prices of ddr5 are causing people to use substitutes
I am guessing there is some fluidity between these markets. If consumers get priced out of the DDR5 market, the consumer DDR4 price increases as well. It is the same with cars, if the price goes up for new cars, it usually increases for used car market as well. Same with energy, price goes up for oil, all things beeing equal, other energy carries will increase to some degree.
Not sure if that's what you've meant...
That's it. Which is insane. We used to buy used DDR4 from obsolete servers in buckets before it was going to become e-waste. Now even that one is expensive.
If memory makers can fill their order books with only a single type of memory (whatever that is), then that would be preferable to them, I suppose.
So.. memory futures? Sounds like a signal of an efficient and mature market to me, manufacturers sell their stock way beforehand, keep production steady, and wholesalers can hold stock and absorb risk in exchange of profit.
Nothing prevents wholesalers from selling their memory or memory futures at a loss, as that is a risk that is justified by equal possibiltiy of the memory appreciating in excess of inflation.
Why AI companies wont build their own RAM but prefer to pay multiple times more than what they sold pre AI boom? (And ram business was still profitable). They already make custom processors
Too slow, bum and bust cycles, dependencies one layer down are also booked up.
AI companies aren't fabbing processors. They're buying capacity at fabs to make their own designs. To make more RAM, they'd need to invest in making more DRAM fabs, which is a slow process.
Lame: Fight the war on general computation by spending decades pushing elaborate technical standards for locking down bootloaders, try to convince society why we need device attestation and hardware-bound keys, gradually establish devices without root access as the norm, get into constant skirmishes with free software and privacy advocates, get drawn into unpleasant political quagmires, etc etc
Game: Just buy up all the RAM...