[−]SubiculumCode · 2026-11-15 Sun 06:11 UTC ·
link
If they are price fixing, eff them.
Finally secure a research grant and had budgeted money for a new compute workstation and the RAM prices are killing my budget, because I'd love to get at least 256 GB of RAM and that's 4-5 grand now. Ridiculous, especially since I'd like to have a GPU in it with at least 32Gb of VRAM for medical image model training, so that is also like 4-5 grand too.
Do I wait two years and hope, or pay, isthe question I keep asking myself.
At least in Europe we are building a common super-computer infrastructure exactly for this. Tier 0 (EuroHPC) to tier 3 (University clusters) now have considerable amount of GPU (both due to demand, but also since classical CPU-heavy FLOPs have becime even more expensive due to RAM prices). Our system will be in a ISO 27001 / TISAX level 3 scope (rather relevant for engineering). The system work up to a PUE of 1.1 and will also solve part of heating decarbonization problem. We need souvereignity in research. But we at least in research we do not need each lab buying their own workstations anymore. The systems are actually also now also offered to start-ups and small enterprises that don't have the capex/opex for efficient and secure compute themselves.
[−]SubiculumCode · 2026-11-15 Sun 08:07 UTC ·
link
That is a third possibility, because there are Federal supportrd resources for compute, see [1]. And I can and will probably use at some point, but there is something about not having to deal with other people's systems and having that bare metal right there.
More like building shared infra to get at a better and more efficent end of scaling laws, than having many small clusters all over. I don't think it really is a privacy question but just a way more economical solution.
Basically make your own hpc provider for academic purposes, and rent out extra capacity to smaller commercial interests to utilise that spare capacity and also prop up the local economy with usable research compute.
Here it is not the privacy concern of the user but that of third parties (e.g. those in medical images). And yes I am a strong advocate of using shared infrastructure for this rather than letting some PhD student handle this a part of their research (which is the reality with lab clusters)
Information sources vary per country, but there should be lists like this one [0] for Germany. Access condition are a bit of a mess. EuroHPC was started to harmonize things a bit [1] but also here the reality is complicated at least for the 50% funded by national money.
[−]SubiculumCode · 2026-11-15 Sun 07:52 UTC ·
link
It's possibility that I've considered, am still considering. Mostly because I am used to working locally, but also because I am not sure which works out cheapest in the end. For my use cases, if I need anything with more than 24 or 32 GB, it definitely makes sense to rent I thing. Those cards are extra pricy. Also, I keep the GPU and analysis computer for the lab after the grant ends, but I don't keep my budget for online conpute..that would go back to the Federal government, so it makes sense for some level of hardware to be around for lean times
I mean, Runpod is relatively cheap at that VRAM tier, it might be a good interim solution so you don't have to completely sit on your hands while you wait.
[−]jareklupinski · 2026-11-15 Sun 11:20 UTC ·
link
> On 5 April 2006, Sun Woo Lee, Senior Manager of DRAM at Samsung Electronics, entered into a plea bargain with the US Government for his involvement in the price fixing conspiracy. Following the plea agreement he was sentenced to 8 months in prison and fined US$250,000. Lee was subsequently promoted to President of Samsung Germany in 2009, and then President of Samsung Europe in 2014.
how many successful careers include 8 months in prison
The allegations haven't been proven to be true. Thats what the courts will decide on when the class action continues.
[−]woadwarrior01 · 2026-11-15 Sun 12:06 UTC ·
link
You might want to read the court docket attached at the bottom of the page. The plaintiffs present substantial circumstantial evidence, consistent with tacit collusion (which is not illegal). It's for the courts to investigate and decide if there was explicit collusion, which is illegal.
Tacit collusion and just working within the same market can also look the same. I have no skin in the game (other than wanting prices to go down again), but there is also way too much emotion involved in the discussions about this because it is hurting our wallets.
- Cutting DRAM production in 2022 as a result of low prices/demand made sense if they didn't expect the market to recover as quickly as it did. A lot of the recovery is also likely caused by the start of the AI boom.
- The deals OpenAI made with SK Hynix and Samsung was a smart move by OpenAI where they drove up prices for everybody else while securing "cheap" supply themselves. I dont think its clear whether SK Hynix and Samsung came out ahead in that deal as the demand likely would have risen either way and it wouldn't have needed to go up by as much as it has now.
- Moving production to HBM also makes perfect sense when that is what the market demands and is willing to pay a premium for.
Every billionaire has incentive to make public statements influencing their competitors while not doing what they say. If everyone else reduces production, you get to take the excess.
See the repeated prisoner's dilemma, where the fact that you keep repeating the game completely changes what it means to have an optimal strategy. A company doesn't make a choice to collude or not once, they make it over and over again.
Billionaires and corporations collude every day via Signal, Telegram, WhatsApp, etc. We've just made it almost impossible for them to leave a paper trail, and for us to prove it. Thus we get lots of people defending their actions as "just business" or something: plausible deniability is all most people need.
[−]matheusmoreira · 2026-11-15 Sun 15:40 UTC ·
link
> but there is also way too much emotion involved in the discussions about this because it is hurting our wallets
That's a matter of course. This goes way beyond our wallets. They are actively pricing us out of computing. They are deliberately pushing computers back to the big iron mainframe era.
Computers are subversive: they democratized cryptography, copyright infringement, all sorts of things. Now that humanity has developed AI, they will not cannot tolerate mere citizens having access to powerful computers any longer.
If we don't fight this somehow, it's over for us. Everything the word "hacker" ever stood for will be destroyed. We need some kind of second amendment but for computers.
I was going to say the same, so I'll just add to your comment: it's possible too that there's a brave new world of local LLMs just around the corner that this kind of thing is retarding.
(Speaking of which—I've seen people suggest a seti@home-like approach to creating a frontier-level, fully open LLM. C'mon, guys, let's vibe code this thing!)
> The judge found that long-term HBM lock-ups, refusals of long-term conventional DRAM contracts, Micron's Crucial closure, fab delays, the failure to expand despite record prices, near-identical customer vetting and the companies' prior cartel history supported a plausible inference of conspiracy.
[−]philipallstar · 2026-11-15 Sun 18:21 UTC ·
link
The only conspiracy is that they can make more money selling HBM and fabs are slow to spin up.
Tacit collusion is only legal in specific circumstances. If the companies involved seem to be acting against their own economic self interests, are communicating indirectly, or are collectively using some external mechanism (like realpage or whatever it was called) to set prices, it can be illegal.
And they were proven to be true. From the Wikipedia article:
“To date, five manufacturers have pleaded guilty to their involvement in an international price-fixing conspiracy between July 1, 1998, and June 15, 2002, including Hynix, Infineon, Micron Technology, Samsung, and Elpida.[3]“
I have to disagree with you there, the parent comment is claiming that this has happened before and they cite that Wikipedia article. ;-)
In my opinion, the fact that this has happened before, and that the companies involved eventually plead guilty, is an important piece of information for people following this story. Especially in light of the statements from the companies involved today.
The comment said "this isn't the first time", present tense, meaning they are currently being found to have done it again. That's not true (the finding has not happened yet), and what the person you're going back forth with is arguing.
The situations are quite different though. The DRAM industry price fixing scandal came in the aftermath of them heavily scaling up production. The manufacturers had all scaled up the production so much that they couldn't sell it all. They would all have had to sell at a loss if they were to compete.
It is the complete opposite of the current situation.
Collusion is a viable strategy, regardless of the market situation. They've been shown to apply price fixing in the past; presumably they still know each other's phone numbers.
Having done all three over my career (local machines as student, University's HPC cluster as grad/postgrad and eventually cloud in the private sector), I'd 100% rent out compute. Buying a custom rig for local ML training makes zero sense in 2026. With renting you'll be able to train bigger models/datasets faster on much better hardware while paying far less in total. Especially if you're doing project-based work.
If you use a custom disk image too, you can have it boot, apply security updates, reboot, then ssh in within minutes.
Work for however long, shut down. I do this with linode and their GPU VMs. If you shutdown nightly and on weekends, the costs drop dramatically on a monthly spend.
I only need it a few hours per day, so very cheap for custom compute.
Oh, you're just talking about deleting it when you're not using it in general, rather than during peak times. Makes sense.
[−]newComputerWhoT · 2026-11-15 Sun 19:12 UTC ·
link
The usage pattern i've seen on flexible compute is more aligned with 'I need this process to execute, run it when it's cheap' not 'I'm going to kick off this process, stop it if it get's expensive.' for example AWS Spot Instances
I mean, there are lots of reasons why you'd want something local, but I agree that the cost makes it unreasonable right now.
[−]probably_wrong · 2026-11-15 Sun 12:13 UTC ·
link
Friends of mine wanted to buy a GPU server last year for their new project. Management told them no, that there was a deal with cloud providers and that investing the budget in cloud compute was better.
The project ran out of budget roughly three months in. From what I heard, status meetings consist of everyone getting together once a month, asking "have we gotten any more budget yet? no?", and then wait until the project runs out.
Cloud providers need to make a profit too, and they've gotten very good at ensuring that their clients forget this point. The whole advantage of cloud compute is that you don't pay for the time you're not using it, but once you reach a certain point you're the one covering the overhead for everybody else.
As a scientist with a big workstation and access to several HPCs I can say that there is nothing better than having a powerful workstation. Yes, doing some sort of long runs is better on a cluster, but being able to do things locally without queue delays and various limitations imposed by HPCs is extremely valuable for initial development/experiments.
HPC ops teams are either really good or they're a dumping ground for grumpy DBAs/Operations people who have no skin in the game of getting anything done and it's a complete nightmare working with them.
As former HPC ops albeit in public service, we also tend to have funding for everything but ops personnel. This means you also often cannot afford the personnel you might be wishing for as they tend to go more towards better paying fields like devops etc.
It might (I say might) be worth your looking at xc [1] as a way to get more performance out of local/LAN hardware - assuming your program is data-parallel anyway.
[−]SubiculumCode · 2026-11-15 Sun 17:50 UTC ·
link
Thanks!!
[−]SubiculumCode · 2026-11-15 Sun 17:49 UTC ·
link
Exactly. The number of times you go, "oops, it would be better if I changed X first" matters less if you can restart right away without waiting to restarts (and also using up your credits on the aborted run).
[−]thatsabadlook · 2026-11-15 Sun 11:34 UTC ·
link
Simultaneously everyone needs to be doing AI research but no one can afford to do it except 5 companies. It's like no one remembers how major discoveries happened in the past.
I would love to believe the price will be lower in two years, but I doubt it. Why would it be lower? There’s still a shortage and most of the stuff it goes into hasn’t even been built yet.
Most of the stuff it goes into hasn’t even been built y̶e̶t̶.
Kevin O'Leary's 20 GW data centre is never getting built. Maybe he'll do a "phase 1" at a few hundred MW, but he's not going to build the largest power station in the world.
The only consumer GPU with 32GB of RAM in the current generation is the RTX 5090. NVidia claims it hasn't been discontinued but it's also impossible to buy and the resale prices have spiked to $5k+. So for a PC with 256GB of RAM and a 32GB GPU you're over $10k just for those two parts. Many believe, myself included, NVidia's new "flagship" will be announced in early 2026 and will be the RTX 5080 Super with 24GB of RAM, also probably costing $2k+ anyway.
What you probably want instead is a Mac Studio with an M5 Ultra. The 256GB model is ~$11k. The problem? Demand is so high that the lead time last I checked was March. But this will give you 256GB of shared memory, well above the 32GB you'll get on a 5090.
If you can live with 96GB (or less) on a Mac Studio it's quite a lot cheaper and the lead times are far quicker.
> The only consumer GPU with 32GB of RAM in the current generation is the RTX 5090.
Well not really, there is the AMD Radeon AI PRO 9700, which is a RX 9070 (or possibly the XT, dunno) with 32GB VRAM. It does use slower GDDR6, but it has "held its price better" (for some definition of better) and is still available for like 2.3k€ (this includes taxes). Though I remember when I was thinking of buying one when it was 1.4k not too long ago and didn't :( now I just don't bother looking anymore.
Slower in what specifically? It wouldn't surprise me if the card is up to 3x slower (it has roughly 3x less bandwidth), but do note that we are comparing a 2.3k€ card vs a +9k€ card here (plus on top of that the power usage on the RTX 5090 can be up to 2x under load).
The RX 7900 XTX has slightly higher bandwidth at almost 1kGBps compared to AI PROs almost 650GBps but has only 24GB instead of 32GB, though I can't find the card on sale anywhere rn.
In the end I see it mostly as a "is the massive jump in price worth it, or do you risk it and maybe get multiple AI 9700s instead and deal with ROCm instead". I can't tell how good ROCm has gotten in regards to training and such, but for inference Vulkan on my 9070 XT is pretty solid.
[−]SubiculumCode · 2026-11-15 Sun 17:56 UTC ·
link
Someone should try vibecoding a massive backplane machine that takes e-waste cards like Tesla K80 as RAM sticks. Doesn't work opex wise but eBay sellers would love a bulk deal at this point.
You might want to consider DDR4, it's much cheaper, and make up for it with a lot of memory channels (Epyc or Threadripper Pro). Those older platforms can be had for a relative song on eBay.
Depends on what you're doing with it, though, obviously.
Finally secure a research grant and had budgeted money for a new compute workstation and the RAM prices are killing my budget, because I'd love to get at least 256 GB of RAM and that's 4-5 grand now. Ridiculous, especially since I'd like to have a GPU in it with at least 32Gb of VRAM for medical image model training, so that is also like 4-5 grand too.
Do I wait two years and hope, or pay, isthe question I keep asking myself.
[1] https://allocations.access-ci.org/resources
You're building a common infrastructure to address privacy concerns about the use of shared infrastructure??
Basically make your own hpc provider for academic purposes, and rent out extra capacity to smaller commercial interests to utilise that spare capacity and also prop up the local economy with usable research compute.
[0] https://gauss-allianz.de/en/hpc-ecosystem
[1] https://access.eurohpc-ju.europa.eu/
https://en.wikipedia.org/wiki/DRAM_industry_price_fixing
how many successful careers include 8 months in prison
The allegations haven't been proven to be true. Thats what the courts will decide on when the class action continues.
- Cutting DRAM production in 2022 as a result of low prices/demand made sense if they didn't expect the market to recover as quickly as it did. A lot of the recovery is also likely caused by the start of the AI boom.
- The deals OpenAI made with SK Hynix and Samsung was a smart move by OpenAI where they drove up prices for everybody else while securing "cheap" supply themselves. I dont think its clear whether SK Hynix and Samsung came out ahead in that deal as the demand likely would have risen either way and it wouldn't have needed to go up by as much as it has now.
- Moving production to HBM also makes perfect sense when that is what the market demands and is willing to pay a premium for.
It made less sense back when billionaires needed to get on trains and spend hours to collude. Now they just make podcast appearances.
See the prisoner's dilemma.
That's a matter of course. This goes way beyond our wallets. They are actively pricing us out of computing. They are deliberately pushing computers back to the big iron mainframe era.
Computers are subversive: they democratized cryptography, copyright infringement, all sorts of things. Now that humanity has developed AI, they will not cannot tolerate mere citizens having access to powerful computers any longer.
If we don't fight this somehow, it's over for us. Everything the word "hacker" ever stood for will be destroyed. We need some kind of second amendment but for computers.
(Speaking of which—I've seen people suggest a seti@home-like approach to creating a frontier-level, fully open LLM. C'mon, guys, let's vibe code this thing!)
> The judge found that long-term HBM lock-ups, refusals of long-term conventional DRAM contracts, Micron's Crucial closure, fab delays, the failure to expand despite record prices, near-identical customer vetting and the companies' prior cartel history supported a plausible inference of conspiracy.
They didn't even announce plans to expand production until after the news of this class action lawsuit broke, despite massive DRAM price increases.
“To date, five manufacturers have pleaded guilty to their involvement in an international price-fixing conspiracy between July 1, 1998, and June 15, 2002, including Hynix, Infineon, Micron Technology, Samsung, and Elpida.[3]“
What hasn't happened is that they have been found guilty in the latest class action lawsuit. Which is what this discussion is about.
So far the court has only decided that the lawsuit can proceed, which indicates there is some substance to it.
In my opinion, the fact that this has happened before, and that the companies involved eventually plead guilty, is an important piece of information for people following this story. Especially in light of the statements from the companies involved today.
It is the complete opposite of the current situation.
Work for however long, shut down. I do this with linode and their GPU VMs. If you shutdown nightly and on weekends, the costs drop dramatically on a monthly spend.
I only need it a few hours per day, so very cheap for custom compute.
The custom volume keeps your persistent install, configs, work in progress, etc.
You can programmatically kick off a linode, attach the volume, then boot, as well as delete, all via curl or python.
So not variable rates, but no rates when the beast is deleted.
The project ran out of budget roughly three months in. From what I heard, status meetings consist of everyone getting together once a month, asking "have we gotten any more budget yet? no?", and then wait until the project runs out.
Cloud providers need to make a profit too, and they've gotten very good at ensuring that their clients forget this point. The whole advantage of cloud compute is that you don't pay for the time you're not using it, but once you reach a certain point you're the one covering the overhead for everybody else.
1: https://github.com/ThrudTheBarbarian/xc
The glut of AI chips will lower prices meaning consumer GPUs start looking high margin. And if they pile into it again, prices will come down.
Kevin O'Leary's 20 GW data centre is never getting built. Maybe he'll do a "phase 1" at a few hundred MW, but he's not going to build the largest power station in the world.
What you probably want instead is a Mac Studio with an M5 Ultra. The 256GB model is ~$11k. The problem? Demand is so high that the lead time last I checked was March. But this will give you 256GB of shared memory, well above the 32GB you'll get on a 5090.
If you can live with 96GB (or less) on a Mac Studio it's quite a lot cheaper and the lead times are far quicker.
Well not really, there is the AMD Radeon AI PRO 9700, which is a RX 9070 (or possibly the XT, dunno) with 32GB VRAM. It does use slower GDDR6, but it has "held its price better" (for some definition of better) and is still available for like 2.3k€ (this includes taxes). Though I remember when I was thinking of buying one when it was 1.4k not too long ago and didn't :( now I just don't bother looking anymore.
The RX 7900 XTX has slightly higher bandwidth at almost 1kGBps compared to AI PROs almost 650GBps but has only 24GB instead of 32GB, though I can't find the card on sale anywhere rn.
In the end I see it mostly as a "is the massive jump in price worth it, or do you risk it and maybe get multiple AI 9700s instead and deal with ROCm instead". I can't tell how good ROCm has gotten in regards to training and such, but for inference Vulkan on my 9070 XT is pretty solid.
Depends on what you're doing with it, though, obviously.