Average DRAM price in USD over last 18 months (opens in new tab)

To those who haven't heard how colossal the size of OpenAIs contracts are.

900,000 wafers monthly. Tom's hardware estimates that is equal to 40% of global dram production capacity.

https://www.tomshardware.com/pc-components/dram/openais-star...

I find it very hard to believe that two South Korea based Chaebol memory producers would have no wind of this from the other.. more likely is that they both agree that it is not in their or the nation's interest to take on too much of a risk profile... while also balancing the risk of appearing to be colluding.

One difference that strikes me with the .com bubble is that I don't remember the .com companies having sustained multi-billions losses / cash burn. They were not profitable but this is quite different. If (or when) the music stops, won't OpenAI go bust immediately? That's quite a counterparty risk those companies are taking.

And when demand eventually crashes, all the new production capacity is left without buyers to sell to, so maybe it does not even make sense to create it.

You have to wonder whether they realy intend to use the ram, or they just spotted an opportunity to corner an essential market and choke/extort in resale.

They are not going to repeat the Windows 8 debacle again are they? (Over producing in response to perceived demand only to not get the demand when they are ready to sell) Didn't that bankrupt some memory manufacturers?

> Supply and demand.

Fake demand - fueled by ponzi debt financed by other debt.

Oh, forgot to mention - fake supply too. China was doubly-excluded from the RAM market by using both equipment export bans and tariffs - ensuring the supply is frozen solid.

I'm pretty sure this isn't a coincidence or due to incompetence.

That process usually takes years.

Here the price hike was pretty instant as secondary effect of DDR5 evasion in two waves. I July and now in October.

There is usually no shortage of old working PC components as they also are avalable used and tested from people decommissioning and upgrading systems. These are not some rare parts in normal market situations.

I made a habit of maxing out motherboards a year or two before upgrading to an new platform. This was always dirt cheap until like 5 years ago.

Yeah the cheapest time to buy old tech is always just when the new stuff has come out. That's when suppliers are trying to shift old stock at cheaper margins.

You can take a look at the 5800X3D and how it was at its cheapest about 2 years ago when AMD was winding down production and Zen 4 had been launched.

I wonder if semi-reliable RAM could be made to work for training. After all gradient descent already works in a stochastic environment, so maybe the noise from a few flipped bits doesn't matter too much.

Sorry but I am looking for reliability. Do you have any stick that fails 100% of the time?

Have you tried underclocking or loosening the timings?

I've come to the opinion that for the vast majority of apps I've built, it could all be built using HTML + CSS (all built server side). I can sprinkle in little bits of interactivity using something like HTMX. And I'll have a website that is very easy to optimise, has phenomenal backwards compatibility, and gets rid of a whole class of issues associated with SPAs.

I often regret in my career not pushing back more on "requirements" that ended up requiring a more complicated app, whereas the customer would have been happier with a simpler solution.

I agree, most websites allow complex flows. But I suspect, that most loads don't ever touch those. There is probably an automated way to deliver just a flat page, and maybe even allow for the top 5 interactions without loading all the frameworks and libraries.

Apple could make it happen. For some reason when an iPhone won't load something, people blame the thing instead of the phone. If they made Safari show an error page when a page used more than 256MB of RAM, suddenly the problem would disappear overnight.

True, but I don't want to go back to that. Bloated web pages suck. But hi-res, true color images and videos don't suck.

Obligatory mention: https://www.downloadmoreram.com

Your mistake is assuming there is some correlation with usefulness and size.

The JS Gmail UI from 15 years ago was just as functional as the one today.

Websites that are supposed to be simple lists end up bloated and laggy because of really poor JS that makes one request per item iteratively to populate a list.

I’d argue that in many of these instances, less is far more.

I want my car to just be really good at being a car, reliably get me from A to B. A Bluetooth connection to the stereo system is nice, but I don’t need a freaking 20” phablet right next to my face when I’m driving.

When I go to a website, I’m usually looking for information, to read something. I don’t often want fancy scroll and animations, I just want clear readable text free of distractions.

More and more these two examples seem to be going away, we’re losing the plot of what the point of these things are.

You can do a lot with little, it just requires investing more in development which understandably most companies are uninterested in. Besides, plenty of websites are bloated as all hell. Why does a newspaper website, for example, have to be very much more than plain html?

Actually a plain car would be great

I crank the window up and down 3x faster than the little button

And I could adjust my damn seat before electricity is available... sigh

I just checked our local apple store website and going from 48gb of ram on a M4 MacBook Pro it's 1500€ to get 128gb. Seems about the same :E

I've been meaning to get more components for a diy NAS since atleast the last year and just been pushing it lazily. I'm literally kicking myself now when I actually started looking up deals for this black friday.

The kit I bought last year for €200 is around €800 now. Crazy!

Been watching 2x32GB DDR5-6000 go from €360 to €920 on a local retailer over the last two weeks

They even have one SKU of 2x32GB DDR5-6400 that’s gone to… €4480

> while also making OpenAI look good on paper

On paper this makes OpenAI look like absolute assholes. Like they have realized that all of their potential competitors will be memory constrained and have poured billions of dollars into making sure that happens instead of using that money to improve their own product.

If manufacturing fake demand (warehousing 900,000 memory wafers per month?) doubles and triples customer prices, it would decrease real demand, hurting the manufacturers after the artificial demand ends.

Fake demand? Either they’re selling the RAM or they’re not. They don’t make money by pretending to sell into fake demand. They make money by selling chips. A sale is a sale.

No, it's not collusion to ask for more money from OpenAI if you hear that they are trying to buy 40% of the world's supply. Increased demand leads to higher prices, that's normal.

OpenAI, by doing simultaneous deals, hid the true demand from the suppliers, thus lowering their price and raising everyone else's.

Every normal market interaction could be described by scary monopoly-sounding words. You heard about you're friend's salary in the same industry, and ask for a higher salary next time you negotiate. You're colluding with your friend to price gouge your employer!

In reality, so-called collusion is normal and unobjectionable. But when price surges happen, often due to factors outside of the seller's immediate control, people look for any reason to find an ethical dimension and find how to place blame, because this is more convenient. Things that were normal become abnormal and suspicious. It is in consumers economic self-interest to act in this way, because it often secures favors to them from various economic policies that they don't normally get when the market is "normal". This is no less a form of collusion than what sellers might do to secure their economic advantage. But a key difference for these anti "gouging" policies is that it gives consumers a special privilege and makes market pricing less able to fulfill its social functions.

They have, rather famously, been caught doing financial shenanigans before. [0]

[0]: https://en.wikipedia.org/wiki/DRAM_price_fixing_scandal

A competitive auction would not be collusion.

This actually points the opposite direction, to doubling down on commercial GPUs.

NVIDIA recently told their board partners that they will need to source their own RAM and will not be bundling it with chips anymore.

If there is a supply crunch on DRAM, commercial GPU production lines will start having idle downtime. That is literally the worst possible thing that can happen to a company that has invested heavily in tooling and they will negotiate at or below cost production runs to fill the gaps if a customer can bring their own DRAM to the table.

What do you mean “might be” trying their hand? It’s been widely reported that they are doing exactly that. It’s even known that they are doing it with Broadcom.

https://www.theregister.com/2025/10/13/openai_broadcom_deal/

> How does Apple do it?

Apple uses off the shelf LPDDR modules. They have nothing whatsoever special about them.

Apple gets high bandwidth out of these modules not with high bus speeds, but with a very, very wide bus. This is expensive on the SoC side (requires a large die, which is why others don't necessarily do this), but allows for commodity memory modules.

That's what HBM is actually. The memory dies are directly next to the GPU die, on the same substrate. The main difference between Apple SoC and GPUs is the former use regular LPDDR while GPUs use HBM.

One of the key points of HBM is that dies are stacked up with many, MANY, more signals and channels. That's how NVIDIA has a memory bandwidth an order of magnitude higher than M4: 550GB/s for the M4 Max, 4.6TB/s for H200. And yes, that's bytes per second, not bits per second.

shorter traces than soldered DIMM allow higher MT/s, this is fixed by CUDIMM/CAMM2, the other part of this is # of memory channels on the board, not sure why, but most consumer DDR5 boards have been 2 memory channels, you need to go to threadripper to get 4 or 8, It's unclear to me if this will still be an issue with future platforms.

That's not blogspam. There's a lot more in the mlid post than the linked news item above. Whether it's commentary or speculation I urge everyone to read and decide for themselves.

As someone currently dealing with pricing changes for DDR5 volume contracts (admittedly, only ~30k DIMMs, but still), those contracts are a lot less rigidly priced than we (engineering) realized at the time - “agreed price” is a lot more... flexible... until the pallet is delivered. Especially because our contract is with our manufacturer who has their own contract with a DIMM manufacturer who has their own contract with the DRAM manufacturer. DRAM is substantially more like a spot market than the average consumer would expect (at least at my volumes). The same is true for my HDDs (~100k unit/year contract) and my CPUs (10k unit/year contract). At least for HDDs our contract quantities are getting honoured (and we actually still have Seagate and WD fighting for our business), which I’ve heard hasn’t been the case for smaller orders.

For a while, yes. But I'm sure they will want to sell it for longer than a year or so.

Even Samsung is running into this issue now: their own internal foundry is refusing to give them a long term contract now so the S26 series will become more expensive.

If this happens to Samsung, what leverage will a player like Valve have?

That's their costs, not our costs. Regular PCs will cost more, which means Valve can price their consoles higher and collect margin, because there's no other competition/substitutes.

It's just surge pricing to manage demand, not "price gouging". Being able to buy RAM at high prices is a lot better than being unable to source it at all because everyone else is panic-hoarding.

> It's 100% retailers price gouging with their current stock that they got for cheap

The retail price of a product is a function of the market rate at the intersection of supply and demand. The price paid for inventory on the shelf doesn’t matter.

It works both ways. If retailers bought a lot of RAM at high prices and then the market suddenly dropped, they could have to sell it at a loss.

Some people get irrationally angry at this, but you do it too. If you bought a house for $500K and the market went up such that it was worth $700K, you wouldn’t think it was “price gouging” to list it at market rate. You’re just trading an asset for cash at the current price. The price you bought it at is irrelevant to the price you’re going to sell it for.

> Apple's RAM is faster than PC RAM

Are you sure about that? The M5 memory has a max bandwidth of ~150 GB/s, meanwhile there is PC memory that reaches 200 GB/s

Why did European prices have the same increase then?

The tariffs are bad, but you're simply wrong here - US tariffs do not affect global prices, it's OpenAI buying up production in advance and cornering the market.

Afraid not. This has not happened across the board to all components.

More specifically, it's OpenAI quietly contracting away 40% of global DRAM capacity, only to build wafers of memory and shove it in a warehouse.

https://www.mooreslawisdead.com/post/sam-altman-s-dirty-dram...

12-24 months is my best guess. I bought some at inflated prices today at a retailer that failed to update to the latest market price by 15%. I'm not waiting that long only to be faced with a dilemma about waiting even longer for DDR6 to become reasonably priced.

Someone managed to boot a machine from a file system on Google Drive.

https://ersei.net/en/blog/fuse-root

Increasing swap is more like adding more HDD. It won't help memory pressure issues.

Yeah, not exactly just 1 million times slower.

Hopefully this blows over in a few years and you’re right and I’m just catastrophizing.

Yh, my moderately capable machine is from 8 years ago and had already waited for an upgrade. Was going to be this Black Friday. Crabcakes.

I can afford it. Other people cannot, and the hobby is driven by a market existing for games. If newer people don’t enter the hobby as others die out, it fades away.

Yeah. Thanks for trying to keep me positive.

Hopefully this all calms down eventually. But it's hard not to feel like shit in this situation.

It’s frustrating that it’s been one thing after the other, seemingly aimed directly at the thing I enjoy most.

> In 2015 there were 3,000 games released to Steam, last year there were 18,000. In 2015 Steam's peak concurrent user count was 8.6 million. This year it's 41 million.

This is like saying "Spotify's subscriber count grew by 800% over the last 10 years. Music is doing better than ever!"

I'm finding that getting an LGA socket to give me two reliable memory channels is far more valuable than the sticks that go into them. I've got at least two motherboards in use at my home right now with only one working channel of memory. There are sticks just sitting around.

My pile of worthless 512MB DDR1 sticks has bad news for you

Nothing stops you building an EPYC based workstation. For the last socket you could even get mATX boards from Asrock Rack which were kind of ridiculous - I was very tempted to build my PC around one.

AliExpress has been selling 2010 or whatever Intel xeons in dual socket desktop board kits for gaming. they are fairly affordable and hold up to almost modern games so people that can't afford new gen systems.

You can buy old rackmount servers on ebay for relatively cheap and used them as desktop PCs AFAIK (though they can be on the noisier side).

At least with ECC memory it's very obvious when its failing as you'll see reported correctable errors.

Why? Doesn’t that suggest that demand justifies the “bubble”?

Memory chips don't contain rare earths, and e-waste recyclers pay a tiny fraction of what used-hardware buyers do.

> If we measure the value of the dollar against megabytes of RAM

We should be measuring it by the amount of RAM in a typical household PC in 1987 and today.

Even though a "meg" of RAM costs less than 1 cent today, I can't do anything useful with it.

Even if we are generous and buy a whole $1 of RAM today, it only gets us 150 MB of RAM, which, while infinitely more useful than 1MB, is still completely useless for running a modern OS/Browser.

What does your math say about that?

A dozen eggs is up over 350%, but a 6-pack of Budweiser is only up 60ish% since 2000. So you know, it all balances out. Maybe drink a few extra cans of Bud with your next meal.

This explores the ideas behind your post: important things, like education and healthcare, have disproportionately risen in price while not-so-imortant thing have gotten less expensive.

https://www.yesigiveafig.com/p/part-1-my-life-is-a-lie

I don't exactly agree with the numbers, but I think the basic ideas are true

The inflation basket only represent a hypothetical average person, who doesn't even exist.

It's more useful to construct multiple separate inflation measures that represent different types of people. Like a "typical renter" inflation figure vs a "typical homeowner" inflation figure. It wouldn't be hard to do and would shine a light on inequality and help explain the rise of populism in certain segments of society.

An even better measure would somehow appropriately normalize the figure by the average disposable income in each of the segments to come up with a figure that measures the felt impact better.

The figure would be negative for wealthy people (who actually benefit from inflation because of asset price inflation) and positive for poor people (whose disposable income mostly goes to rent).

Making America Great Again, Beer and TV.