In a way this feels a bit absurd for these AI centers to hog HDDs.
As pointed by others neither training nor inference require HDDs and storing raw data should not require that much.
So my hypothesis is that it is a double whammy of overall declining consumer sided HDD demand, leaving data centers as main source of demand and additional demand from the new AI centers.
I feel like the AI centers are just buying HDDs because why not throw a HDD in each server blade even if there is no need? The money is there to be spent and it must be spent.
As someone who has been building computers since 1989 it feels like end of personal hobby casual building.
I will end with an imperfect analogy with multiplayer gaming. It is quite common in multiplayer games for higher level players to wish to acquire some tradeskill they neglected to acquire earlier. maybe a new quest appears, or new "must have" item that requires such skill.
They (past me included) have too much game money and no wish to acquire tradeskill items slowly. So the "rich" will overpay by 2x or 10x or even 100x the usual price.
That is free market at work right?
In the process whole low level economy is destroyed due to 2nd order effects. Meaning a new player starting out can only be a farmer.
So if a student comes to me wishing to start building computers what advice do I give them? Farm something?
https://www.reddit.com/r/hardware/comments/1q9zd0f/power_sup...
We have a long way to go before the average PC costs even half as much as it did in 1989 (adjusted for inflation). And of course the performance for typical consumer use is orders of magnitude better than it was back then.
What's next, the great CPU shortage of 2026?
On the other hand the total storage capacity shipped each year has risen, as a combination of HDDs getting larger and larger, and demand shifting from smaller consumer HDDs to larger data center, enterprise and NAS HDDs. I'm not sure how flexible those production lines are, but maybe the reaction will be shifting even more capacity to higher-capacity drives with cutting-edge technology
https://twitter.com/sama/status/2023233085509410833
Funny how they all meet at the "You'll own nothing" Swiss club.
This is assuming most of what we stored are either images or video.
Which goes into why I think this might be good. Developers have kind of treated disks as "oh well" with binaries ballooning in size, even when it can easily solved, and there is little care to make things lightweight. Just like I now figure out a different solution to recover space, I'm hoping with a shortage this kind of thing will be more widespread, and we'll end up with smaller things until the shortage is over. "Necessity is the mother of all invention" or however it goes.
And for the same reason - to avoid the dominant players going "oh shiny" on short term lucrative adventures or outright trying to manipulate the market - causing people to starve and making society grind to a halt.
Still, at least in the short to medium term, there are many companies, institutions and individuals that are used to getting new hardware that is under warranty - developer hardware refreshes, store hardware updates and deployments, some schools and academic institutions that provide hardware to students or even just people getting a laptop for schools use.
I think people would be used to getting new hardware for all of these & might be pretty much left out cold without a machine or having to opt for a reused or resold machine without the necessary technical knowledge to support it themselves.
China is now the only solution to fix broken western controlled markets.
That's when I sell of my current hardware and house, buy a cow and some land somewhere in the boondocks and become a hermit.
"You will use AI, because that will be the only way you will have a relaxed life. You will pay for it, own nothing and be content. Nobody cares if you are happy or not."
You also have to look at the current status of the market. The level of investment in data centers spurred by AI are unlikely to last unless massive gains materialize. It's pretty clear some manufacturers are betting things will cool down and don't want to overcommit.
Could we though? Even if gerrymandering and voter suppression weren't already out of control, and getting worse, there are very few politicians who could or would do anything about all this.
Elections aren't going to save us.
Then they came for the RAM, but I did not speak out, for I had already closed Firefox.
Then they came for the hard drives, but I did not speak out, for I had the cloud.
Then my NAS died, and there was no drive left to restore from backup.
Source: I've been using Chrome's built in translation heavily since around 2018-19, and not before that only because that's the time I've realized it'd became god damn good.
Maybe to clarify my argument a little bit more, "AI" was already good and was being used for useful/productive purposes before it became the hype. It'd be better had it stay the same way instead of it being the latest money generation machine and sucking the air out of the room.
Thanks a lot, AI: Hard drives are sold out for the year, says WD
Now that crypto has failed to paperclip maximize the world, maybe LLMs are attempt number two.
> They largely come from hyperscalers who want hard drives for their AI data centers, for example to store training data on them.
What type of training data? LLMs need relatively little of that. For example, DeepSeek-V3 [1], still a relatively large model:
> We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens
At 2 bytes per token, that's 29.6 terabytes. That's basically nothing compared to the amount of 4K content that is uploaded to YouTube every day.
There are many new data-centers they are being filled with servers. Most servers have at least 2 HDD (mirror) for the OS. I would not be surprised if on a huge scale even 2 HDD per server could cause HDD shortage.
There are likely models which are trained on 4k video and it should be stored somewhere too.
Even things like logs and metrics can consume petabytes for a large (and complex) cluster. And the less mature the software the more logs you need to debug it in production.
In the AI race investments if not unlimited at least abundant. In such conditions optimization of hardware usage is the waste of time and velocity is the only things which matters.
So hopefully once they all go bust, there should be a lot of cheap enterprise HDDs on the marked as creditors pick through the wreckage. :)
Gpu I understand but hard drive looks excessive. It's like if tomorrow there is a shortage of computer cabling because ai datacenter needs some.
You have 31/999 credits remaining. What activity would you like to productively participate in today?
The real advantage of SSDs in this use case is storage density and power efficiency. On the other hand, your compute resources might be packed in so tight with power-intensive stuff that you appreciate the spinning rust “wasting” space.
Why is this allowed on HN?
Identifying generated comments is not always easy, as others have pointed out, plus we don't come close to seeing everything that gets posted. If you see a post that ought to have been moderated but hasn't been, the likeliest explanation is that we didn't see it. You can help by flagging it or emailing us at hn@ycombinator.com.
With LLM comments, there's an important distinction between legit users (who may have no idea that they're breaking a rule, especially because it isn't explicit in the guidelines yet) and accounts that appear to be posting nothing but gen-AI text. If you (anyone and everyone!) see a case of the latter, definitely please email us because we've been banning those accounts.
Even in more borderline cases, though, it's still helpful to email us because sometimes we contact the user, if we can, to let them know that we've been getting such reports. Or we might tell them we've suspended their account until we hear from them that they won't post LLM-generated or processed comments.
1) The comment you replied to is 1 minute old, that is fast for any system to detect weird comments
2) There's no easy and sure-fire way to detect LLM content. Here's wikipedias list of tells https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
How do you know that ? Genuine question.
The "isn't just .., but .." construction is so overused by LLMs.
In this case “it’s not x, it’s y” pattern and its placement is a dead giveaway.
It's not ironic, but bitterly funny, if you ask me.
Note: I'm not an AI, I'm an actual human without a Claude account.
It seems personal computing is being screwed so people can create memes, ask questions that take 30 seconds to find the answer to with Google or Wikipedia, and sound clever on social media?
If we are talking generative AI, again from my experience, things get a bit blurry. You can use smaller models to dig data you own.
I personally used LLMs, twice up to this day. In each case it was after very long research sessions without any answers. In one, it gave me exactly one reference, and I followed that reference and learnt what I was looking for. In the second case, it gave me a couple of pointers, which I'm going to follow myself again.
So, generative AI is not that useful for me, uses way too much resources, and industry leading models are well, unethical to begin with.
ChatGPT does this just as much, maybe even more, across every model they've ever released to the public.
How did both Claude and GPT end up with such a similar stylistic quirk?
I'd add that Kimi does it sometimes, but much less frequently. (Kimi, in general, is a better writer with a more neutral voice.) I don't have enough experience with Gemini or Deepseek to say.
LLMs do these things because they are in the training data, which means that people do these things too.
It is sometimes difficult to not sound like an LLM-written or LLM-reworded comment… I've been called a bot a few times despite never using LLMs for writing English⁴.
--------
[1] particularly vapid space-filler articles/comments or those using whataboutism style redirection, which might be a significant chunk of model training data because of how many of them are out there.
[2] I overuse footnotes as well, which is apparently a smell in the output of some generative tools.
[3] A lot of pre-LLM style-checking tools would recommend this in place of hyphens, and some automated reformatters would make the change without access, so there are going to be many examples in training data.
[4] I think there is one at work in VS which I use in DayJob, when it is suggesting code completion options to save typing (literally Glorified Predictive Text) and I sometimes accept its suggestion, and some of the tools I use to check my Spanish⁵ may be LLM based, so I can't claim that I don't use them at all.
[5] I'm just learning, so automatic translators are useful to check what I'm written isn't gibberish. For anyone else doing the same: make sure you research any suggested changes preferably using pre-2023 sources, because the output of these tools can be quite wrong as you can see when translating into a language you are fluent in.
[6] Another common “LLM tell” because they often have weighting functions especially designed to avoid token repetition, largely to avoid getting stuck in loops, but many pre-LLM grammar checking tools will pick people up on repeated word use too, and people tend to fix the direct symptom with a thesaurus rather than improving the sentence structure overall.
I've find myself doing it, a time or two.
We'd much rather hear you in your own voice: https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...
Using LLMs to learn, of course, is great. HN posts benefit from things people learn however they learn them. Just please write them yourself.
(Also, I'm sorry for the harsh reactions people have been showering on you for saying this. The community feels really strongly about it, and of course the norms around all this are still in flux, not just here but in society.)
This is the game plan of course, why have customers pay one time for hardware when they can have you constantly feed them money over the long term. Shareholders want this model.
It started with planned obsolescence, now this new model is the natural progression.. There is no obsolescence even in discussion when you're only option is to rent a service, that the provider has no incentive to even make competitive.
I really feel this will be China's moment to flood the market with hardware and improve their quality over time.
Yep. My take is that, ironically, it's going to be because of government funding the circular tech economy, pushing consumers out of the tech space.
It's no coincidence that Microsoft decided to take such a massive stake in OpenAI - leveraging the opportunity to get in on a new front for vendor locking by force-multiplying their own market share by inserting it into everything they provide is an obvious choice, but also leveraging the insane amount of capital being thrown into the cesspit that is AI to make consumer hardware unaffordable (and eventually unusable due to remote attestation schemes) further enforces their position. OEM computers that meet the hardware requirements of their locked OS and software suite being the only computers that are a) affordable and b) "trusted" is the end goal.
I don't want to throw around buzzwords or be doomeristic, but this is digital corporatism in its endgame. Playing markets to price out every consumer globally for essential hardware is evil and something that a just world would punish relentlessly and swiftly, yet there aren't even crickets. This is happening unopposed.
It's so hard to grasp as a problem for the lay person until it's too late.
Fortunately we won't ever see a shortage of monitors and input devices, because then how would we consume the rent-a-remote-desktop services.
Things are bad and I don't know what can be done about it because the balance of power and influence is so lopsided in favor of parties who want to do bad.
The only silver lining is that newer devices will have to scale down memory, so developers will have to ditch memory-sucking frameworks and start to optimize things again.
If it’s temporary I can live with it.
I guess this was inevitable with the absolute insane money being poured into AI.
Hyperscalars refresh hardware and firesale old stock.
~2028 is going to see a lot of high power refurb supply hit the market.
Forget the details, as it's been a decade+, but everything gets junked at some point.
They're almost always stripped down, then sold as board+chassis only.
How many useless living humans do you know? They go somewhere. Something happens to them. Whatever it is it’s about to happen to 30% of the population.
What’s the opposite of survivor bias?
Oh, I can think of about 77 million right off the top of my head.
Given this has been going on for years at this point, the high prices of graphics cards through crypto and now AI, it feels like this is the new normal, forever propped up by the next grift.
US is medium risk and increasing rapidly. Run away quickly.
Sure you have to isolate certain rogue states - North Korea, Russia, USA. Always the way.
Big tech will be deemed "too big to fail" and will get a bail out. The tax payers will suffer.
Just like the price of labour. Your salary went up and doesn't come down
In the UK weekly earnings increased 34% from December 2019 to December 2025.
CPI went up 30% in the same period.
Obviously that CPI covers things which went up more, and things which went up less, and your personal inflation will be different to everyone elses. Petrol prices end of Jan 2020 were 128p a litre, end of Jan 2025 they are 132p a litre [0]. Indeed petrol prices were 132p in January 2013. If you drive 40,000 miles a year you will thus see far lower inflation than someone who doesn't drive.
[0] https://www.rac.co.uk/drive/advice/fuel-watch/
That's how inflation works. In this case it seems more narrow though, there's hope the prices will go down. Especially if the AI hype finds a reason to flounder.
Work is fucked. 23TB of RAM online. Microservices FTW. Not. Each node has OS overhead. Each pod has language VM overhead. And the architecture can only cost more over time. On top of that "storage is cheap so we won't bother to delete anything". Stupid mentality across the board.
It could even become a kind of renaissance of efficient code… if there is any need for code at all.
The five guys left online might even get efficient and fast loading websites.
Honorable mention of the NO-TECH and LOW-TECH magazine site because I liked the effort at exploring efficient use of technology, e.g., their ~500KB solar powered site.
https://solar.lowtechmagazine.com/about/
We went from using technology to solve problems to the diametric opposite of creating new problems to solve with technology. The latter will have to contract considerably. As you say, many problems can be solved without code. If they even need to be solved in the first place.
On the efficiency front, most of what we built is for developer efficiency rather than runtime efficiency. Also needs to stop.
I'm a big fan of low tech. I still write notes on paper and use a film camera. Thanks for the link - right up my street!
Call me cynical if you like, but I don’t see this optimism that assumes the banal idea that somehow good always wins, when that’s simply not possible and in fact bad-guys have won many times before, it’s just that “dead men tell no tales” and the winners control what you think is reality.
Same way the price of groceries going up means people buy only what they need and ditch the superfluous.
I can understand that incumbents may not want to overinvest in capacity, which could be financially precarious, but they are also putting themselves in danger by opening up avenues for competition. One more thing ruined by AI mania, I suppose.
I wasn't saying it'll be good and that the good guys win, but a lot of insane creativity to circumvent restrictions will pop up.
No. Prices will just go up, less innovation in general.
/s