The RAM shortage is bringing back DDR4 (theverge.com)

This is a supply constraint and requires no assumption about demand - HBM packaging capacity constrains the amount of wafers that can go towards HBM (regardless of price in the near term), and excess wafer capacity will be dumped into DDR5 and crash DDR5 margins.

We’re doing a similar thing now too, building tons of extra power capacity and links to data centers.

After the crash no one will need to build a new data center for 20 years, and there will be plentiful power connections

That’s not exactly true, a lot of GPUs are sitting in warehouses because they are lacking prerequisites to being installed, in some cases this is even just power.

https://www.tomshardware.com/tech-industry/data-centers-in-n...

That’s one big difference. The other big difference is that capital back then was mainly spent by startups with no or little traction.

Today, the "old" corporations like google, meta, amazon are also spending a lot of money.

> ai capacity is being used right now today.

A lot of its current use is just basic questions that could be handled by local llms just as well, yet they are being processed on GB200s on the other side of the world.

This does not seem to be the case. The very first VeraRubin NVL72 systems are barely installed and won't be running real workloads until beginning of '27. Colossus II appears to be the largest user of Blackwell Ultra NVL72s and those are also only recently on-line in September of this year with the remaining 3/4 of them not until end of the year. The delays in datacenter deployment are very long. Even their GB200s didn't come on-line until January of '26. Hopper H100/200s still dominate.

The other hyperscalers are even further delayed. CoreWeave has some VeraRubin, but not in volume (as I understand).

https://en.wikipedia.org/wiki/Colossus_(data_center)

https://www.moduledge.com/blog/nvidia-hopper

> In general. ai capacity is being used right now today.

Not according to this article[1], which says:

"I believe we are now in an inevitable overbuild situation, one with no neat, tidy Dot-Com Bubble-style exit story. Demand for NVIDIA GPUs — and those from Broadcom, AMD and other semiconductor companies — is driven by speculative capital believing that the AI industry will become magnitudes larger than it is today, largely driven by the fact that everybody believes there’s far more demand for compute capacity than actually exists. ... NVIDIA has created a remarkable illusion perpetuated by the media — that GPU sales are a direct measurement of the actual demand for AI compute, rather than a measurement of how a few companies are willing to invest in an idea two years in advance, using circular financing as a means of creating the sense that you must buy these GPUs now, or you’ll miss out on the future. ... At the very least, hyperscalers are going to be burdened with brutal depreciation charges or onerous write-offs for years to come, whether their capacity turns into revenue or not. ... I don’t see how 90%+ of NVIDIA’s sales ever end up generating a single dollar of revenue, and considering the amount of project financing-backed data center debt deals, there’s very little that exists to protect investors if AI compute demand never arrives."

[1] - https://www.wheresyoured.at/wherere-all-the-ai-chips/

You have described the market milking process very accurately.

Theres machine learning algorithms in the mix, and it all works perfectly smoothly, until it doesn't.

Smoothly here, being defined as the markets emotional feelings about it all, not, the price went up.

The price went up because it could, as you pointed out.

To add onto this: it is generally not possible for suppliers to sell below market rate. If DDR5 is going for $400/16GB, and you want to sell 16GB for $200, people are going to buy your cheap RAM until you run out and resell it for more money. This is called arbitrage, and it's the market rewarding others for being greedy where you weren't. The only way to make the price go down is to make the market's price go down: produce so much supply that you extinguish all of the arbitrage-driven demand.

There are some wrinkles to this. If CXMT's RAM was somehow inferior, then the market could sustain two different prices for two different products. But it's difficult to separate the market like this without just making unusable trash. The one separation we do have - DDR5 vs HBM - works against the consumer's favor. One layer down in the supply chain, RAM manufacturers have to buy silicon wafers to etch circuits into, and HBM takes up treble more wafer space than DDR5 per gigabyte. In fact, this is specifically the scarcity that allowed Sam Altman to engineer a RAM crisis, by buying up a bunch of wafer supply that he could then redirect to more AI-optimized HBM.

We can also infer from all of this that CXMT - while producing lots of memory - does not have the scale necessary to actually quench the inflated demand. Either because their yields are shit, or because China wants to run a business and made the same calculation as the other RAM producers that the AI bubble would pop before they could get additional equipment online to service demand.

That's what we mean when we say "milking". Just because you can explain something in economic terms does not make it justifiable.

If you're growing onions in the backyard, you don't need a router.

If you're writing emails, you're gonna need to send them through a router or two before someone reads them.

> I don't believe GPUs are a new asset as Jensen Huang claims.

He sells GPU's, so taken in context, he's promoting the financing of GPU's, by anyone who is listening to his sales pitch. He travels in CEO circles, so, he's talking to other CEO's, and the banking industry.

The phrasing of it also directly pushes back against the articles accusing the AI industry of circularly financing itself to inflate everyones balance sheet.

And the other thing this idea of it being an asset that you invest in, like a house presumably, is that it primes all of us for 2 future realities: price increases and minimal performance gains ( "its useful for longer so finance it" ).

As to economic value proposition to any individual or group, it boils down to how they are going to use it to recoup the expense.

Don't just take the sales guys word for it, ever.

There have been many videos lately, on whats happening in the automotive market. The industry has been adding more and more tech, to the point where its now financially extractive via mandatory dealer interaction aka fees, and people are becomming fed up with it all, wanting something simpler for less money.

Translate this to the big fat GPU card industry, and how long is it before AMD/Intel add just enough AI ops into every CPU that noone cares anymore and they start asking themselves "Why am I buying 96GB of RAM and a GPU, and 96GB RAM and a CPU ? I'm buying ram twice ! It's a conspiracy !"

Unified architectures seem to be inevitable, GPU's would no longer need seperate ram, then shrinkflation is gonna hit your video card, and you're gonna pay GPU+RAM prices for just a GPU. The upside to this is that you could fit more GPU's in your box and they won't need 2.5 slots each. For now, they are hoping the DGX Spark type boxes fill the market demand, and allow them to sell higher priced products. Unified architectures would dent demand for those.

Just my 2 cents on it all. Enjoy the ride.

>Routers don't do productive work.

I'm sorry, what?! Communications is a valuable thing. How is all that AI slop going to hit the Internet if there's no routers?

Without routers, your email isn't going anywhere, it doesn't matter who wrote it.

For what it’s worth, I’m willing to buy as much RAM from everyone else at 50% of what these companies are selling it at. If everyone works together, we can cut the market price by selling it to me. I will provide the much needed capital in dollars to shift the price and others can provide the RAM to me.

I’m willing to spend a lot to drive down the price. As a show of good faith, I promise not to actually pay any more than 50% of what you can get on the used market.

The important question isn’t so much “are people trying to get the best deal possible for themselves?” but rather “how well is this market functioning?” Or phrased another way, “how competitive is this market?”

This particular market is pretty clearly not functioning well, and there are textbook Econ 101 reasons why a market may not function well. Of those textbook reasons, I suspect the most obviously applicable one here is large barriers to entry.

The reason why the memory cartel fucked everyone over isn't exactly the pricing. It's the supply. It's that the established manufacturers all agreed on "AI demand is a fluke, we wouldn't rush to online new manufacturing capacity" back in 2022.

A calculated move. If they're right and the AI demand is a fluke, they don't overinvest in expensive manufacturing infra, they don't get saddled with more capacity than they need, and there is no memory glut eating their margins in the future. If they're wrong and the AI demand isn't a fluke, they get to pad their margins big time, and the supply bottleneck isn't their problem. As long as they shake hands and bet on "fluke" in sync, the downside is very limited for all of them.

They shook hands, they bet in sync, and they were wrong. Now the supply bottleneck is everyone's problem.

Potentially even theirs - if Chinese vendors can use it to wedge themselves into the market when it's at its most lucrative, and invest in growth. Which they almost certainly will.

Huh. This is a crime against humanity. The planet is burning.

Who is "they"? Certainly we have some concentration of consumers of this hardware. Are you saying they'd rather destroy the hardware to protect their supplier's market rather than allow the secondary market to function as normal? Maybe this is written into contracts? We're in a very sad state of affairs if that's the case.

A company can either generate a small new revenue stream from on-selling old gear, or they can pay to destroy and dispose of it.

Which option is a profit-oriented, not-totally-profitable AI outfit going to choose?

The liquidators will squeeze every penny out. All that ram will hit someone who can take the chips out and repackage for consumers.

Aren't servers depreciating assets owned by the company and tracked on the balance sheet?

If so, on a liquidation event, I'd imagine they'd have to be sold to compensate the shareholders.

Lots of precedent for that: In the EU WEEE has incentivized companies to crush working electronics rather than selling them, because it is way less bureaucracy to hand something to a certified recycler than it is to legally sell the stuff. And once something has been entered into the WEEE system as waste it's pretty much impossible to legally recover it: the recyclers are literally not allowed to simply pick working equipment out of the pile and sell it. Technically WEEE is supposed to enable reuse, but it's way easier and less bureaucracy to just crush stuff. The necessary per-item paperwork, testing and certification (it has to actually work, broken/non-working equipment cannot be removed from the WEEE system, because it is waste, legally) and logistics required for reuse exceed the residual value anyway.

That's why used enterprise/industrial electronics are way less common on EU eBay compared to other places, despite the industrial base being so large (or used to being so large). You can actually see when WEEE came into force because very little used T&M equipment entered circulation in the EU after the early 2000s.

Another factor is lease agreements, where companies don't buy equipment and the lessor "recycles" (i.e. crushes) the equipment to avoid creating a used/second hand market as an additional benefit.

That's a witty joke, we used to make them all the time ...

Energy is a commodity so the price will be very similar for the same goods. What they tack on as taxes might be different.

Me too.

EU average is just under 29c/kWh (non-contract price) [0]

In the UK I got 27.5c/kWh (+ 58c/day fixed charge) on an 18-month contract a month ago

[0] https://ec.europa.eu/eurostat/statistics-explained/index.php...

Regardless if the market allows consumers access to negative solar prices or not, consumers can attach solar panels to their servers (independent of the grid), so the LLM is there when the sun is shining.

It's fun to think about running datacenter hardware at home, until you get the electricity bill.

I've been extremely pleased with my R320 in a NAS role. Idles at 40 watts, is quiet, has ECC, and the drives stay cool.

Maybe that's still too much power for some people, but I'm happy with the tradeoffs.

My homeserver is essentially a top of line gaming PC from 2012 with a Quadro RTX 4000 from 2018. It was not top of the line in 2013 when I bought everything on sale but after many inexpensive ebay upgrades, it's now what I wish I could have afforded back then. Most of the time, it's perfectly fine since it's main job is jellyfin. I'm sure I could be saving some electricity with newer hardware but with the current market, it's just not worth it. I'm missing AVX2 support which Claude Code has recently made a requirement but now I just have Claude ssh into it from another PC that is a few years newer.

Typing this from an Asus X99 WS 3.1 with a E5-2696v4 . I got 256GB of RAM, sees only 128GB but I am sure with some overvoltage and lowering the ram speed it could see all 256GB.

if anybody got a tip, i'd be more than glad to listen, please.

I just bought a Xeon E5-2690 v4 for $16 and a Sparkle A770 for like $350. Going to try some local models on there. I think it will be plenty good for my use cases.

Yeah, I have an old X99 old asrock Fatality (lol) board, micro ATX with a Xeon E5 2640 and 128GB of ECC DDR4 running 24/7 as a Truenas server with 80TB of storage in it. It seemed like an overkill when I put it together two years ago when I replaced my 5820 CPU that had been my daily desktop but now with storage and everything prices exploding it has turned into a very useful thing. I still don't like the fact that it sits at around 130 watts. But given that it has so many hard drives in it, it's not that bad. My whole basement cloud now feeds into it.

What's up with the "V"? I've seen a couple of your comments that have a lone V at the end of the comment or paragraph.

One or two DDR3 ramdisk cards were produced but they’re basically non-existent today. There are DDR4/5 ones (CXL), but they they’re restricted to specific enterprise motherboards.

> Yep. To the point it makes you wonder if the "AI bubble" is totally misunderstood. It's less "AI bubble" and much more "Monopolize computing". Buy up computers until most companies have no choice but to use cloud.

Unless prices get categorically higher, I don't think that's likely. Like 10x more than it is now.

Companies were buying computers for all their employees in the 90's, where a mediocre laptop could pretty easily cost two grand in 1990s money. My mom just bought a middle-of-the-road laptop for about a thousand dollars in today-money, and that was a consumer buying it; companies can generally spend more. "No choice but to use the cloud" feels like it's way too strong of a conclusion to draw.

A "monopoly" grab from a conspiracy spanning all of the major tech firms? You really think that's the most likely explanation? Not just that the firms are competing in AI and all realize they need massive compute?

Prices are up a bit but there are plenty of "free-programmable computers" from a dozen or more manufacturers on the market. That's not going away.

> the business plan people got convicted for trying in ancient Persia, 5000 years ago.

Excuse my ignorance, do you have more details on this?

Of course, when we talk about utilizing AI as an efficient tooling, it is well expected to see the very outcomes you desire. But the point here is the disconnect between the perceived productivity gains and their corresponding measuring metrics. Remember the Goodhart's Law[1]? It becomes fairly unreliable when you try to measure productivity by a set of standards instead of actually producing productivity. When it comes to future about this issue, I'd hope to see the progress actually happening in open source, decentralized AI instead of the current way of using it.

[1]: https://en.wikipedia.org/wiki/Goodhart%27s_law

Yeah, I talk to normies and nobody still understands the utility of just having a ChatGPT desktop app being able to basically drive your computer and do whatever needs to be done.

Well is there any productivity gain from chatbots? So far, I mostly see a deluge of misleading content disguised as documentation for things that don't even exist. If anything, the amount of slop out there is slowing me down…

I can't imagine it can even feel productive to anyone, unless you use agents to automate re-rolling your slop machines, but then you're just going to end up with an unmaintainable ball of mud. Either way, it's definitely net negative.

Kind of same outcome as if you just generated lots and lots of random code (perhaps at AST level; but don't tell your Claude, that'd trick it into making itself obsolete tech!) and compiled it in parallel till it passed tests, but this time around, the computer also tells you that you hit the nail on its head, that you're totally right <s>and that you're such a good puppy. agent uses tool `clicker`</s>

How so?

What? All a shortage does is ensure those with deep pockets control the entire ecosystem. Bye bye personal computing. The people with money don’t want you to own anything, they’d rather you rent everything from them. How is that not an undesirable outcome?

To put the AI buildout in perspective, it exceeds the GDP percentage that was put into the railroads or interstate highway system.

It's actaully "super" in the sense super intelligence is causing the shortage

My guess is it may happen after the election in November. Trump has everything tied up in a pointless war in Iran right now, so it would be a good time for China to make a move, before the 2028 election gives us a more competent government again.

Yeah but TBH an old server with DDR3 ECC is still way sufficient to serve something like 99.9% of all the servers out there's jobs.

That's the plain truth: most sites/servers simply serve very very very little and have not anywhere near the need for 256 GB of RAM, let alone 512 GB.

These old servers aren't sufficient for the Amazon/Apple/Microsoft/Meta/Google/Oracle of this world, but for 99.9% of the rest of the world, they're more than plenty.

And the world and technology shall keep on marching: we'll have DDR6 and as soon as that happens, suddenly all the hyperscalers shall consider DDR5 to be oh-so-lame and dumped by the truckloads in landfills.

We may have two more painful years ahead for RAM but we'll then have plenty of used enterprise/datacenters gear on the market.

Most games do not, unless you want the best graphics. But then you’re out of the 90% use cases

DDR4 have similar latency as DDR5 and that's what matters in all games, while raw bandwidth only matters in some games

You were ripped hard. $150 is 32GB of 2133/2400 DDR4. DDR3 should be like $50 for 2x16GB set.

Sorry.

One would think just some classes and forms of RAM increased in price. I have 128 GB (2 x 64GB DDR5 SO-DIMM) in my home NAS. Home NAS. Those nice ugreen DXP8800 Plus thingamabobs. It'd be over 2k euros for RAM alone now. Ridiculous. Add to that two SN850X 4TB SSDs, and 8 (+1 unpacked spare) IronWolves Pro and it's now a serious worth.

No end in sight.

I know the only manufacturer of such 64GB modules: Crucial (Micron), but Micron doesn't sell new chips to Crucial, thus almost no chance to get them anymore (though few months ago there were some). Kingston (using Micron, Hynix and Samsung) still does not have 64GB.

Soon Zen 6 should become available (well, or not so soon). With 7200 MTs memory controller? I haven't seen such JEDEC modules at all. Are they possible without the additional clock driver (like on 6400 CUDIMMs)?

Lenovo was selling AMD HX Pro with 96 GB DDR5 and 2TB of SSD end of last year 1600. Same model now 6300.

I have laptop, desktop, and server RAM. Which did you want? I can send those pictures.

"hosting GPUs for local llms" meaning pci-express bandwidth and memory bandwidth are not issues for them, and CPU features probably also don't make a difference.

I think it'd be mostly for engrams. It seems LLMs are moving in this direction, with 50GB in qwen 3.8 flash, and 200GB in deepseek 4.1 flash.

My $20 xeons 2696v2 and 2670 are tool old for real offloading, and sata ssd drives are likely a bit slow for ssd streaming :)

For consumer use DDR4 is fine and price is the real obstacle as is availability. So if they can ramp up DDR4 production that would be far more useful than limited quantities of DDR5.

I can’t blame them for making back their costs for any manufacturing and maybe a little extra for profit selling at the current rate. Let’s see what they do though :)

Could totally be one of those situations where the current cartel outsource all their production for MoW outside the US to them and then just pocket the investment money lol

However … without the US datacenter buildout ramping continuously … the ram will eventually make its way from warehouses to the market lol. We really need a way to use server dram in consumer boards… but idk if that will happen.

Although it would take a ton of engineering and probably a new chip … a custom motherboard would that did that would be legit haha :) would solve a lot of issues.

100% agree. This is generally a good news story. The e-waste issue is one of the darker aspects of tech which is mostly ignored. We should promote brands like Fairphone and Framework that are trying to make a difference here. Reuse more of your kit!

It's good if you're a tech guy in silicon valley with a good salary who used to upgrade every year to the latest pro max.

It's terrible if you're a labourer in India who buys phones under 10k INR (~100 USD) when your salary is 20k INR per month.

Unfortunately there are way more in the latter category than the former, and most of them won't know what ram is. They'll just be paying more and getting less going forward for a few years.

(Source, the salaries and phones I've seen of the staff at office. And we're in a major region you've likely heard of, not some tiny rural village.)

I was reading that oil from the Middle East had to go to South Korea to be refined before being shipped as diesel to Australia.

Cost of hiring an oil tanker has ballooned massively (like 40 fold) since America decided to bomb its supply chain, so these long supply chains are massively expensive.

The Republican Party has decided that running a world of global free trade was no longer desired. America wont pay the price for another decade or two, but it will.

Well capitalism is responsible for that.

I think in a few hundred years we'll have learnt what a disgusting, poisonous system it is.

I have bridge to sell you if you think that is likely.

How many prompts include "make it as efficient as possible"? Or will efficiency clash with shipping as many features as possible(not even humanly anymore).

> That too sounds like an actually good development all things considered

It would be good if we would use less resources.

But we are not.

It's just that some people don't get the newer resources.

Old e-readers were already usable. Generating extruded story-like product on them when they're already perfectly capable of displaying the vast majority of human literature doesn't make them any more usable.

Also making old hardware unaffordable too.

Produce rots in the fields because there was no way to get a profit from it

Look at CrowPanel, you can get nice e-ink devices made for development, at a reasonable price. I have a small 2.13in esp32 board, I spent some time yesterday making it into a pocket flashcard device for my German courses. The energy consumption is minimal, small form factor, easy to flash via esptool or arduino utils

I remember getting a Radeon card on sale for about $180AUD in 2012ish. And 4 years later looking for an upgrade and discovering the cheapest higher performing card was at least $300 Aud.

I eventually upgraded to an RX 570 in 2018, and hit the same problem 5 years later.

So as far as loe to mid range cards are concerned you can wait 5 years and still have to pay more to replace it for an upgrade.

As far as I can tell, in terms of MSRP, 980ti was $649 at launch. 1080 and 1080ti were both $699 when they hit the market.

$699 in 2015 (when the 1080 was new) is ~$975 today, adjusted for inflation.

$975 is pretty close to the launch MSRP of a 5080, which is "just" $999. (We can't actually buy them for $999, of course, but we couldn't usually get 1080 cards at MSRP during the mining frenzy a decade ago either.)

This is the crux of a lot of the anti-AI sentiment especially in the gaming market. They feel entitled to access to the latest chips at a subsidized price below what industry is willing to pay. Many of them then take it a step further and claim some conspiracy (like a few commenters here).

I get why they're upset (it's not fun to be crowded out) but it's a fundamentally selfish position.