[legitster]

What's happening right now is very different than what happened back during the dot com crash. When they were doing price fixing it was because there was a glut of supply and demand was tanking. Prices were falling and they coordinated to keep prices from falling less.

Right now demand for DRAM is extremely high bordering on endless. Prices are going up. The incentive for one of the big players to undercut the other on cost even just a little bit to pick up market share is extremely lucrative.

It would also be dumb to cut production when prices are high because you increase the incentive for one of the outside players to suddenly ramp up production and jump in the market.

Not saying they aren't coordinating in other ways (following each other's leads on price hikes and availability). But again the context here is literally the opposite as last time.

[sc68cal]

> The incentive for one of the big players to undercut the other on cost even just a little bit to pick up market share is extremely lucrative.

I would argue that the DRAM price fixing scandal actually demonstrates that the industry operates like a cartel. During times of low demand and high supply, they will coordinate to protect prices, and then during spikes in demand (or alleged spikes in demand) they coordinate to keep the price from dropping.

https://en.wikipedia.org/wiki/DRAM_price_fixing_scandal

[ChadNauseam]

> then during spikes in demand (or alleged spikes in demand) they coordinate to keep the price from dropping.

Why would they need to coordinate to keep the price from dropping during a spike in demand? a spike in demand will obviously not be expected to lower prices regardless of collusion

[zozbot234]

Yes, the industry is capacity limited so if there's a true spike in demand, prices will be high even absent any collusion. Especially if previous investment in expanding capacity has been lacking for many years.

[jagged-chisel]

To keep someone from undercutting others’ prices and causing a competitive pricing war.

Realistically, it wouldn’t be a meaningful drop to the consumer. But it’ll affect some executive’s ability to buy another unnecessary trinket.

[zozbot234]

If the industry is at capacity (which it plausibly is, especially since HBM memory is made in the same facilities) then no one can physically "undercut" anyone else. Collusion works by artificially restraining supply of some valuable good; if there was genuine collusion at play, we'd probably be seeing companies make less of the expensive HBM (to push its price even higher; note that patent and other IPR restrictions can in fact have this effect, to some extent) and more of the comparatively cheap DRAM!

[amy_petrik]

>Why would they need to coordinate to keep the price from dropping during a spike in demand?

They wouldn't you're right.

But I would expect for them to follow the sorts of behavior we've observed in other markets - egg prices, gasoline prices. When a spike occurs, even if as brief as a lightning strike, they will only very slowly drop prices, when in a purely capitalistic world the price drop ought to be equally fast - suggestive that the slow drop is a mutual agreed upon collusion. After all, it's in all sellers best interest to game that "consumers temporarily agreeable to scalping prices" as hard as possible, Nash equilibrium or whatever amongst sellers. Many such cases and more vicious and brutal punishments for such behavior would serve to benefit the common man, the final point and benefit of capitalism

[protocolture]

>they will only very slowly drop prices

Of course. Price drops only really come through in response to competition.

Long term high prices invite further investment. Investment arrives and wants a quick return. Fastest way to return is to sell above cost but below market. Established players respond. Yada yada. I once met a former projector salesman who was unbelievably angry that someone, I think Acer, came along and destroyed the ~2000 AUD hard price floor that projectors once commanded, which dropped the whole market and his commissions along with it.

Even when collusion is government endorsed instead of outlawed, the same rules apply. See the bromkonvention. You need the new player willing to take the 10% margin to hurt the bottom line of the guy taking 150%.

>when in a purely capitalistic world the price drop ought to be equally fast

No the price can only drop as fast as supply and competition can catch up. For an industry with high input capex costs, thats extremely slowly. I would think some banks would be keen to take a risk on a new RAM fab based on the demands coming from AI, but also I would personally not take the bet that AI will be in this state in 5 years time. So assuming Banks and other lenders are as skeptical as I am, they wouldn't lend, or would request a bigger entity guarantee the loan.

>brutal punishments for such behavior

Brutal punishments for failing to ramp up production? Or for not lowering prices for no reason? I really dont understand

[zajio1am]

> when in a purely capitalistic world the price drop ought to be equally fast

That is not obvious. When demand is higher than supply, it is clearly good move to raise prices. But when demand is lower than supply it us not clear than lowering prices would raise volumes to compensate for lower margins.

[tshaddox]

If there are a small number of sellers and barriers to entry are high, we should expect at the bare minimum "tacit collusion":

https://en.wikipedia.org/wiki/Tacit_collusion

[lpribis]

> Following the plea agreement he was sentenced to 8 months in prison and fined US$250,000.[6] Lee was subsequently promoted to President of Samsung Germany in 2009, and then President of Samsung Europe in 2014.

Lmao he got rewarded for taking the fall.

[flanderizedNed]

Was in the room working for a hardware company during the offshoring phase of high tech manufacturing

Offshoring was 100% because of antitrust concerns. Copyright landlords and hardware manufacturers were concerned with Americans and their morals also having the skills to create endless competition.

American workers compensation prize was endless hustle jobs to distract them from political revolution.

The olds don't care if the kids end up unskilled knowledge serfs, fuzzy VHS tapes of outdated academics. The olds will be dead.

[hyperpape]

The problem with citing cartels and greed is...when prices are low, is it because the industry is momentarily altruistic? Did the industry wake up in late 2025 and think "holy shit, I've been being nice, but it's time to turn over a new leaf and start gouging people?"

I mean, don't get me wrong, they are greedy. But that's been true for years. What's changed is the market.

[fl1pmoniz]

What has changed is for years they never had an excuse as good as AI , what they do isn't creating demand where it does not exist, it is making sure demand doesn't drop by colluding on prices and limiting production between all the companies

[woleium]

i thought samsung took a fab offline as well?

[ffsm8]

I do not have any particular insight into the market, I was just extremely confused reading this in particular

> Right now demand for DRAM is extremely high bordering on endless. Prices are going up. The incentive for one of the big players to undercut the other on cost even just a little bit to pick up market share is extremely lucrative.

Given this first statement, the last one makes no sense.

Either the demand is limited, giving sellers that undercut the ability to move more products... Or the demand is higher then production, making it nonsensical to undercut anyone, because you'll sell out anyway, no matter if you're cheapest or not

[lazide]

If demand is high and supply is fixed (generally the case) you can also pull a DeBeers, grab everyone by the balls, and charge 3x (or more) the already high prices everyone would already get when you sell out your inventory.

[energy123]

To the extent that demand is price inelastic, which in practice means a lack of substitutes and the product serving an essential need.

[lazide]

With the AI boom ongoing, plenty of people with deep pockets are desperate. Perfect time for some gouging.

[tliltocatl]

Demand for diamonds is rather elastic, yet DeBrees did it.

[bakugo]

> Right now demand for DRAM is extremely high bordering on endless

Why is demand extremely high right now when it wasn't a couple of months ago? What changed since then that caused it to triple overnight?

[TrainedMonkey]

Fundamentally because demand and supply curves are lagged. This drives DRAM flood and drought cycles. Right now high DRAM prices are driving fab investment up and demand down. In 3-5 years new fabs will cause DRAM glut that will drive prices down. Lower prices will stimulate demand via things like doubling DRAM size in consumer electronics as companies compete on getting the number bigger. Eventually demand curve will eclipse the supply and we will end up in the DRAM drought again.

This time things are further complicated by the fact that the world is investing a sizable chunk of GDP into building RAM hungry data centers in hopes of building a god which will convert the rest of the world into data centers.

[ghurtado]

> in hopes of building a god which will convert the rest of the world into data centers

I'm as pro AI as it comes, and I love your way of putting it. Very poignant.

[tonyhart7]

if its what it takes then so be it

which company achieve AGI first would be a quadrillion company

[Dylan16807]

If you make something that's as smart as a sloppy intern and costs ten thousand dollars an hour to run, you're not going to make any money off of it.

[jpdb]

But if you make something that's as smart as a sloppy intern and costs even a small amount less, you're making an unbelievable amount of money off it.

[vablings]

The biggest issue here is that it hurts smaller consumers. Scaling up a fab to produce ram takes 5+ years, people generally buy new hardware ~5 years so this lifecycle of hardware for some people is now locked out. Let's say you was due for an upgrade this year you now might be priced out of the market for the next 5 years and basically due to poor business practices

There is a similar issue happening in the manufacturing space where metal foundries are basically "full" up on allocation for other customers and will refuse to sell to you unless your purchase order is six digits otherwise you pay a hefty premium which once again drives capital towards larger corporations. Compounded by a stagnant jobs market the means that scarcity is just going up and up and nobody is re-investing to meet consumer demands because the market is poisoned by speculation to the absolute extreme

[ffsm8]

Do people really still upgrade so often? I mean it made sense pre 2015 for desktops, and pre 2020 for laptops... But since... Not much has changed from a performance standpoint.

Even gpus hardly advanced since 2022 (4090) and the next generation is at least 1++ years off. Likely 2-3... And it's unclear wherever it will actually be an upgrade or more of the AI shenanigans they released with the 50 generation.

[pimeys]

Well, I just updated my 5950x with 128 GB to AMD Strix Halo with the same amount of much faster RAM. It is noticeable faster, but what's better: the whole computer is tiny and sips energy.

I'm very happy I ordered this in the summer, framework delivered it to me early this month. I wonder will these machine just be out of stock now or the price goes up a lot...

[pabs3]

I'm still on a 2013 desktop I found in a dumpster, its mostly fine still, apart from ballooning memory use in browsers.

[o11c]

Laptops are pretty fragile, especially with how a lot of people treat them. Replacement is due to breakage, not obsolescence.

[1718627440]

If you have an old enough Laptop, you can just replace parts.

[brennanpeterson]

This isn't true. It used to be, as a new fab would appreciably add quantity. At 1M wspm in 2015, a new 100k fab at the most modern node would add effectively 20-30% capacity, and usually multiple.players at once, since all had cash.

Now, the relative shrink is tiny, so capacity adds are just wspm, in effect, and that gives 5%.

Put differently, you cannot invest your way out of the shortage, or into meaningful share....so you take profit.

[gertlex]

apparently wspm is Wafer Starts Per Month

guess I'm interpretting "1million wspm; add 10%; was effectively a 20-30% capacity increase" in 2015.

Not sure where 5% then comes from. Guessing "relative shrink" is referring to process size (5, 4, 2 nm, whatever) not linearly corresponding to density of transistors, etc.

[davoneus]

I've heard it was Sam Altman and OpenAI basically buying every wafer available from both Samsung and SKHynix at the start of October.

Neither company know of the other purchase until it was a done deal.

[Panzer04]

Kind of funny to see manufacturers get screwed by their own opaque pricing policies for once.

All well and good when you're dictating terms with dozens of buyers, but probably not so much when a single buyer is dictating terms to a couple of sellers.

[Scaevolus]

The CEO of OpenAI and OpenAI didn't coordinate?

[ndriscoll]

Samsung and SK Hynix. OpenAI made deals to buy almost half of manufactured DRAM in the world. There's speculation that after this was announced, other companies started making their own deals in a panic, further driving demand.

[lucb1e]

> after this was announced, other companies started making their own deals in a panic, further driving demand.

Oh ffs it's like the toilet paper thing. I was amazed how long that continued despite credible sources saying there is no shortage, just insane demand from the loonies that don't believe it would return to normal instantly if they would just stop buying more and more extras because "see, it's out again!"

[zozbot234]

This is almost certainly what's going on right now in the retail market. OTOH it's also a semi-rational response to volatility and uncertainty as to future wholesale prices, due to, e.g. the projected build-out of future AI datacenters. As with any durable good, whenever the price might be expected to rise in the future, people will want to hoard stockpiles and the expected price rise will be brought forward to the present.

[lysace]

Is DRAM built using the same process/node as GPUs?

[alephnerd]

> Why is demand extremely high...

Data Center projects that were announced a couple months ago are now beginning to be built out.

Additionally, there have been some supply chain issues the past few years due to trade wars between the US, SK, and China [0], along with the earthquake that hit Taiwan last year [1].

Generally, you feel the pain of supply chain issues within 6-18 months of the initial incident, which is where we are now at because stockpiles have been reduced significantly.

[0] - https://www.digitimes.com/news/a20230310PD204/chip-war-memor...

[1] - https://www.reuters.com/technology/micron-flags-hit-its-dram...

[loeg]

It didn't triple overnight. Contracts for 2026-2027+ hyperscaler orders get negotiated gradually over time and when those contracts are N% higher than last year, ~all supply is spoken for.

[ares623]

Maybe Some nerds joked at lunch that if we hoard all the RAM we cut it off from competitors. They saw what COVID did to supply chains and thought they’d be so smart if they could simulate it.

[bigiain]

DRAM 2025 as toilet paper from 2020. Seems plausible.

[gishh]

No no no, it’s all just the bullwhip effect, remember? Remember when prices went fucking nuts during Covid and then dropped?

Oh wait, they never fucking dropped.

Still waiting, all you bullshit, er, bullwhip truthers out there.

[0manrho]

People pulling their heads out of their ass as to how to actually deploy these systems at scale (AKA to do this effectively, you need to do more than just throw pallets of GPU's at it, such as properly considering Topologies of both NVMe-over-Fabric and PCIe roots/lanes [0]) combined with advances in various technologies (eg RDMA, CXL, cuDF/BaM/GPUD2S/etc) that meaningfully enhance how system ram can be integrated and leveraged are a big part of it.

Also we're hitting that 5 years after DDR5 being readily available which means that a lot of existing enterprise hardware that was on DDR4 is going EOL and being replaced with DDR5 which, given many platforms these days have many more channels available than previously, results in more DRAM being bought than was previously used per node and in total. A lot of enterprise was still buying new DDR4 into 2023 as it was a more affordable way to deploy systems with lots of PCIe lanes which was more important than any the costs associated with the performance gain from DDR5 or related CPU's. (Also, early days DDR5 wasn't really any faster than DDR4 with how loose the timing was unless you were willing to pay a BIG premium)

Regarding the hype of the day: AI specifically, part of it is the rise of wrappers and agents and inference in general that can run on CPU's/leverage system ram. These usecases aren't as sensitive to latency as the training side of things as the network latency from the remote user to the datacenter means latency hits due to hitting the CPU ringbus(infinity fabric, QPI, whatever you want to call it) results in a much less significant share over the overall overhead, and the cost/benefit/availability concerns there has also increased the demand for non-GPU AI compute and RAM.

I wouldn't rule out corruption/price fixing (They've done it before) but I have no evidence of this. Wouldn't surprise me, but I don't think this is it (unless this problem persists for several quarters/years)

There's some geopolitics and FOMO (Corporate keeping up with the joneses) and economics that goes into this as well but I can't really speculate on that specifically, that's not really my area of expertise. Suffice to say, it's kind of like a bank run where it's not so much that the demand itself hit the curve of the hockey stick, but it was gradually increasing until it hit a threshold that was starting to cause delays in delivery/deployments. Given how important many companies view being on the cutting edge here, this lead to sudden spike in volume customers willing to pay premiums for early delivery to hit deployment deadlines, artificially inflating demand and further constraining supply, which just fed back into that feedback loop pushing transient demand even higher.

0: Yes NVMe NAND flash is different than DRAM flash, but the systems/clusters that host the NVMe JBOD's tend to use lots of sysRAM for their index/metadata/"superhot" data layer (think memcached, Redis, the MDS nodes for Lustre, etc), and with the advent of CXL and SCM you can deploy even more DRAM to a cluster/fabric than what is strictly presented by the CPU/mobo's memory controllers/channels. This is not driving overall market volume, but is a source of fierce competition for supply at the very "top" of the DRAM/Flash market.

TL;DR: Convergence of a lot of things driving demand.

[imtringued]

>People pulling their heads out of their ass as to how to actually deploy these systems at scale (AKA to do this effectively, you need to do more than just throw pallets of GPU's at it)

Yeah something people don't understand is that the models have become so big it can take minutes to load them from SSDs. If you need to restart a CUDA process for whatever reason, you'd rather want to load the model files from RAM. This means for every GB of VRAM, you also need a GB of system RAM. Then there are things like prefix caching and multi user KV caching. Users generally don't do all their requests one after the other in a short window of time. This means you are better off freeing up VRAM after a minute has passed. If the user sends another request, using the system RAM as cache is still more energy, time and VRAM efficient than recalculation. DDR5 based DRAM is incredibly cheap.

>Regarding the hype of the day: AI specifically, part of it is the rise of wrappers and agents and inference in general that can run on CPU's/leverage system ram.

It has more to do with the dominance of mixture of expert models. Due to expert sparsity the required memory bandwidth drops quite significantly. It is possible to run gpt-oss 20B on a computer with 32GB RAM, a segment that used to be reserved for enthusiasts and developers has now become the mainstream amount of RAM on desktop PCs and mini PCs.

Yeah so if I had to summarize, the problem is that DDR4 is EOL, shifting demand to DDR5. AI demand means people want more than 16GiB RAM (there is actually a flood of used 2x8 GB kits in the laptop DDR5 market). DRAM manufacturers switch to supplying AI data centers and stopped resupplying retailers. Retailers are running out of inventory, leading to a sharp rise in prices. Early production of DDR6 will be out in 2026 with consumer availability in 2027, so there is zero incentive to expand production for DDR5.

[kvemkon]

> DDR4 is EOL

11 years after specification released.

DDR4 makes it possible since 5 and half years to build 128GB AMD PC using 4x 32GB DDR4-3200 1.2V JEDEC modules. Only since half a year it is at last possible to build AMD PC with 128GB DDR5-5600 RAM. Because Ryzen DDR5 controller cannot operate with 2 sticks sharing the same channel at JEDEC speed/voltage.

DDR4 just cannot be EOL already because of unbuffered ECC! Even today it is not possible to have 128GB with 2 DDR5 sticks on AMD. Only 96GB.

And actually I'd expect 256GB (DDR5-8800 1.1V JEDEC) to be possible with DDR5. We're now only from 4000 to 5600 5(!) years since the specification was published. What if AMD achieves this only few months before DDR5 EOL?..

> DDR6 will be out in 2026 with consumer availability in 2027

... or not at all?..

[Spooky23]

I think implying that the DRAM makers are nefarious is nuts. What would you do in their shoes?

You have to think about what’s happening in the context that its 2025. The US is different… there is no functional financial regulator. The CEO of OpenAI is trying to corner the market and be too big to fail. We’re seeing crazy stock swings with Oracle (whose CEO quit after the September stock event), AMD, Intel, etc. Much of the action is fueled by the circular relationships with OpenAI. Add the drama of malevolent players like Elon & POTUS.

If you’re the CEO of Micron, you’re in a tough spot. You have constrained capacity, delays in your new fab buildout that puts it out of play, and a market that is lucrative, but may be a massive Enron type event.

The people making the shovels need to hold tight. There’s no reason to collude. The smart move is to keep on trucking.

[m4rtink]

Or they don't want to be the ones holding the bag when all the AI companies that ordered RAM from them end up bankrupt and insolvent ?

The stuff they ordered might even be so custom to be useless to be sold by the creditors, leaving the RAM manufacturers hanging dry with unpaid invoices and idle capacity.

So they just do the old cartel thing and wait it out - with the added benefit of selling the same stuff they have been making for twice the price!

[casey2]

Supply and demand "theory" being used yet again to justify obvious price fixing.

[blackoil]

If fixing, Why now? Why this much?