This is the real risk after the slow death of personal computing. Even internet resources like servers will be hoarded by the hyperscalers that are the only ones who can afford to order years of compute hardware in advance.
I’ve already started buying cheap old business PCs just in case I’ll ever need to have simple barebones machines to run things on.
> This is the real risk after the slow death of personal computing.
By the time you can have a slow death of personal computing, capacity will improve and prices will improve.
In the shorter term sitting on an old computer or regressing a couple years on specs or paying an extra $100/$200 for 8GB/16GB works.
> Even internet resources like servers will be hoarded by the hyperscalers that are the only ones who can afford to order years of compute hardware in advance.
I don't see why hyperscalers would be so much better at handling price increases.
For some average business paying a week's wages for the computer you use, they can afford that doubling to two weeks just fine.
For all the normal server rental companies, okay the guy on the $10 plan either pays $16 now or cuts their resource allocation and keeps paying $10. That's not going to cause a sea change. And higher end hosting isn't that much different.
> By the time you can have a slow death of personal computing, capacity will improve and prices will improve.
I love this baseless optimism. Reminds me of the economic theory (forgot who put it forward): Everything will be fine in the long run.
...and it's rebuked by Maynard Keynes: "We're all dead in the long run".
Yes, by the time capacity & prices will improve, we'll be all dead.
> I don't see why hyperscalers would be so much better at handling price increases.
I'm pretty sure that you don't know how much discount they can get when they order "I'll buy the whole production in Y2027". When first generation EPYC was launched, it didn't reach academic or local datacenters, because AMD sold all production to Hyperscalers and Dropbox back in the day.
Money is always more valuable today than tomorrow, so if you can pay today, you'll get massive discounts.
> I love this baseless optimism. Reminds me of the economic theory (forgot who put it forward): Everything will be fine in the long run.
Seems accurate though, I've already noticed no-name Chinese manufacturers stepping it up, throwing caution and capital to the wind and leaning in as hard they can. Following the typical Chinese model of how they get involved in markets, we're just a couple years away from memory being a market with skyrocketing prices and limited availability to a market where various nations are considering bailouts and tariffs to protect their local manufacturers from dirt-cheap Chinese exports to keep their memory producers from going through what domestic solar panel producers went through (a near extinction level event).
I specifically didn't say disk mfg as well only because I haven't yet noticed a new big spinning HDD Chinese brand. But they're definitely active in the lower end of the DDR5 market.
I don't see that possibility from a data retention duration perspective. SSDs are glorified and addressable capacitors at the end of the day, and they leak.
> ...and it's rebuked by Maynard Keynes: "We're all dead in the long run".
Keynes said that when somebody complained that his theories didn't work in the long run. And the true rebuke is, we're now in the long run and Keynes is dead
> ...and it's rebuked by Maynard Keynes: "We're all dead in the long run".
The trouble with Keynes is that it's only fully true on the time scale of the heat death of the universe, and in that context it's fully nihilistic. Whereas most economic theories do operate on timescales where the finding out comes within the lifetime of the people fucking around. And to the extent that it doesn't, it generally comes within the lifetime of their kids. Meanwhile that quote is used to justify every piece of short-term thinking that screws the next generation to juice this year's numbers.
> When first generation EPYC was launched, it didn't reach academic or local datacenters, because AMD sold all production to Hyperscalers and Dropbox back in the day.
When first generation EPYC was launched, it broke Intel's effective monopoly on performant servers that everyone was eager to get out from under, but the first generation was being fabbed by Global Foundries using the decaying infrastructure being sustained only by the few uncompetitive Opterons nobody had really wanted in years.
It's the example of the thing you're saying doesn't happen. The following generations were fabbed by TSMC who has dramatically more capacity than GF and expanded it even more since EPYC launched, to the point that AMD's share in servers this year is almost 50%, up from ~0% the year before EPYC launched.
> Money is always more valuable today than tomorrow, so if you can pay today, you'll get massive discounts.
The real issue here is capacity planning. It costs billions of dollars to build more fabs so they only do it if they're confident the demand isn't going to crash. But cash-rich customers willing to pay in advance are a good way to do that. You give them a contract that says they pay you now and agree not to dump the hardware into the market if the bubble pops (e.g. customer agrees to maintain possession of the hardware for 3 years after delivery and use only for AI) and then the AI companies are the ones taking the risk instead of the hardware companies, which makes the hardware companies willing to build more fabs. Which in turn is what gets the price back to something ordinary people can afford.
> The trouble with Keynes is that it's only fully true on the time scale of the heat death of the universe, and in that context it's fully nihilistic. Whereas most economic theories do operate on timescales where the finding out comes within the lifetime of the people fucking around
You got this completely backwards: Keynes argument is that economic policies cannot just rely on the fact that things are going to be fine “in the long run”, because the “long run” may be something we never see. He was in fact arguing against economic theories that are “only fully true on the time scale of the heat death of the universe”, not the other way around.
No, the quote is regularly brought out as a justification to ignore long-term and even medium-term effects whenever the result is anything less than instantaneous. Even right now we're in a discussion about manufacturing where the lead time is a low single digit number of years. The expectation that we'll all be dead before it shakes out is rather implausible but there is the quote.
And there are no valid theories that are only true if you wait until the heat death of the universe, because by then everyone is dead and there is no one to constitute an economy. There are, however, many theories that could be valid even though they take decades or more to shake out, and that quote is used especially against those to rationalize exactly the short-term thinking that lets people ignore that even when they are dead, we, i.e. humanity and its future generations, will be holding the long-term consequences of whatever we choose to do right now.
Late reply. This is all assuming that the corps producing the hardware will live in the same world as before this squeeze. The reality is that the combination of a highly concetrated manufacturing landscape + the global political environment means that the same games that were played with ASML, Nvidia and TSMC will likely be repeated with the relatively lightly affected generic manufacturers of components. Given the close connections the hyperscaler CEOs have with the USGov, and the leverage the US has over South Korea and Japan, one executive order is enough to tip over the entire manufacturing sector and supply chain to the hyperscaler’s advantage.
As others noted, only the Chinese sector has a chance to fill the space for consumers in this nightmare scenario.
So far its just metastasized not smoothed out. GPUs then RAM then electricity then disk… long way to go till land then sunlight, hope we don’t get there!
Hyperscalers buy so many CPUs and so much RAM they can dictate prices (to a point) or at least make an agreement to buy at a much lower price but buy X amount for the next 5 years.
When the market's so hot I don't see why anyone would give them a particularly big discount. And they would have been getting a discount before, so they probably end up seeing a larger percentage spike in what they have to pay.
Maybe if they're locking in long enough to fund new fab construction? But in that case after a few years a ton of capacity will come online so they're actually helping solve the problem.
I work on a desktop, but as a backup I have bought a refurbished Dell Latitude, there are a lot of decent ones for €250 on eBay. Put Linux on it, it’s good enough for most workloads.
I just hope my top shelf 2020-era desktop doesn’t die on me because it would get very expensive to get a new build these days.
> I just hope my top shelf 2020-era desktop doesn’t die on me because it would get very expensive to get a new build these days.
I could probably sell my gaming rig (12900K, 64GB of DDR5, 4TB NVME, RTX 3090) for more today than what I built it for about 4 years ago, it's absurd. I won't, of course, because it's still glorious for 4K gaming even today. In retrospect, $5000 very well spent.
I have a similar rig, but a 10th gen i7, 128GB of DDR4, 2TB NVMe, and dual RTX 3090s. Built it in 2021 for crypto mining, it ended up paying itself off just before Eth went PoW. I kept it mining even after profitability, because it ran warm enough to heat my apartment leaving my HVAC on fan-only to circulate the heat around.
After winter, I started playing with various other GPU loads until LLMs and SD became easy enough to use. Now it's my experimentation machine.
It's already paid for itself, so anything I sell it for would be profit, but it is still super nice for running local LLMs that power various projects "for free".
I won't sell my 3090. It's still one of the best cards for a local AI and gaming mixed use. I put a waterblock on it and it just purrs along at 50C no matter what I throw at it.
I did exactly this and suggest it. Used Dell Latitude (the one I got is one from 2019 - model 5300). I put Linux Mint Debian Edition (https://linuxmint.com/download_lmde.php) on it. Works absolutely great.
Small Dell Optiplexes are good for desktop computers.
This is a market shock. The reason it doesn't alleviate sooner is the chip manufacturers are very wary of investing billions on capacity that won't come online until after the shock is over and subsequently going bankrupt. Because that's happened before.
IF IT LASTS, capacity will increase.
But it won't last. The AI boom is in exponential growth but it's based on heavy speculation about future value and the bubble will absolutely pop, how agressively depends on how dumb people are about now. The current growth may or may not be entirely justified but it's not sustainable, the free investor money does run out. These back and forth self-dealing deals where companies that own big pieces of each other announce "partnerships" where companies are selling resources essentially to themselves and counting the revenue several times... those are a sign of the approaching peak.
If the hyperscalers demand is persistent, more hardware players will notice and will want a piece of the pie. You already see non traditional players entering.
But what you see is a cautious strategy from the existing players. They are hedging against a bubble. They don’t want to pour today tens of billions of dollars in capacity that they will have to sell it to a deflated market
> This is the real risk after the slow death of personal computing.
This.
People tend to take things for granted, but the world we've grown up with is not guaranteed to live forever, in twenty years personal computing could look like the personal CD or video tapes collection of the 90s-00s: a thing of the past for most people.
It's quite crazy to imagine personal computing to die. It always felt as a staple, that I could always buy components, that someone would manufacturer them and I could cobble together a machine. But now the prices going so high that it kills the market is... just sad.
I don't think component prices will kill personal computing. Prices used to be far higher in the heyday of personal computing. Millions of people buy phones every two years that cost twice as much as a perfectly capable PC.
Personal computing in the sense of actually controlling our hardware and software will be killed (or rather degraded) by government regulation and platform oligopolies.
> It always felt as a staple, that I could always buy components, that someone would manufacturer them and I could cobble together a machine.
I feel the same. But at the same time the generalization of laptops (with custom motherboard, and soldered parts) and smartphones (custom SoC) means this is already a niche thing. Gamers were the last mass market for upgradeable modular deskops, but a sufficiently long hardware drought would most likely make game editors way more resource conscious, which may kill this niche entirely.
Personal computing is already dead. We use locked down mobile endpoints that run pre-approved thin client software connecting to giant mainframes over monopolized radio networks.
> I’ve already started buying cheap old business PCs just in case I’ll ever need to have simple barebones machines to run things on.
Hmmmm...
That's a bit dramatic. I did buy a few SSDs when RAM prices started going up for I didn't want to be facing a shortage of SSDs but China is already very busy at producing cheap server motherboards: gone are the days where he only option for a server motherboard was a $1000 one. There are ultra cheap, and fully functional, chinese motherboards now.
I'm totally convinced China is not going to sit doing nothing: if RAM prices goes up and there's a business opportunity, China is going to seize it and start producing lots of RAM.
We did see "RAM price quadruple in no time so nobody had time to adapt", that's for sure.
But I'm really not sure it's "RAM prices goes 4x and then the world stays as it is and nobody adapts for decades to come".
Also as to used servers: people ordering years of compute hardware in advance aren't hoarding five years old servers.
Heck, a ten years old Xeon machine is plenty capable and the usual people are buying stocks of those (from companies updating their fleet), refurbishing them and reselling them one by one on the usual marketplace, at the same prices (OK maybe instead of 150 EUR it's now 200 EUR if there's 64 GB of ECC RAM).
My usual seller is doing its business as usual although we're well into the memory craze.
Just to put things into perspective: the "PC sales crash" from 2021 to 2024 saw, what, a 20% drop, 10% of which already recovered. In 4 years more than a billion PCs were produced.
A PC is not a rare thing. We won't run out of PCs, just like we won't run out of cheap used ten years old Xeon servers.
But OK you got me: I may contact my seller and hoard two or three more just in case.
Sheesh is HN full of paranoid people and, darn, is it contagious.
I can't wait for a scrappy startup to spend a couple billion dollars and several years setting up a new fab for chips so we can see the prices go down!
Plenty of startups are doing exactly that. Perhaps not for your use case, but they do exist. Add to that, that china will eventually crack the ASML process and then there were three. Add some leaks and industrial espionage and then there were four. The market is a ninja, and we should be thankful for it!
That's the rub. There is no such thing as "anyone that can supply hardware cheaply"
Either it's an established vendor with designs and fabs or it's a newcomer that needs to invest a massive pile of cash in designs and fabs. Neither are cheap.
It is supply and demand but chip supply isn't very elastic and the producers are conservative about adding capacity that won't be needed by the time it's online.
I mean, I want to double the the number of billions of dollars I have by 2031, doesn't mean I will. (actually I have 0 billion and double 0 billion is still 0 billion, so unlike them I'll accomplish that goal).
No. It doesn't have to be a bubble at all. It just has to be a cycle of rapid capital equipment build-out that returns to more normal levels in a few years.
Hard to believe that an industry making money hand over fist are reluctant to spend tens of billions expanding capacity that won't come online for years?
There have been SEVERAL crashes that have wiped out the market and it's the reason there are so few players, the rest of them went bankrupt after periods of over-expansion. (in the 80s caused by Japan, in 1997, in 2001ish after the dotcom bust)
You're even calling it a bubble so it's not exactly "hard to believe" it will pop.
Idealistically, actual consumer driven capitalism would be much closer to supply/ & demand than this (imo) almost completely artificial, government sponsored bubble. I think pricing before this current bubble reflects that.
Because, oddly enough, there are a lot more things going on. Government incentives, companies trying to stranglehold markets, FOMO-fueled massive investiment into the current fad, etc. Just to name a few.
I was also told this. Why isn't it happening? Can some capitalists explain why capitalism isn't happening the way it's supposed to? Is it because of government regulations, do we need to deregulate?
“Is it because of government regulations, do we need to deregulate?”
Insufficient law enforcement. The same memory manufacturers already broke antimonopoly laws in the past, pleaded guilty. Apparently the fines were too small for these companies to care, and the people responsible were promoted instead of being punished. More information: https://en.wikipedia.org/wiki/DRAM_price_fixing_scandal
In this context, it takes spending enormous piles of money over the course of at least several years to spin up new semiconductor production.
We do need more capacity to keep up with AI datacenters' usage, yes.
But adding long-term capacity years down the road for a thing that some folks seem to confidently think is a bubble that can pop at any time is risky. And (because capitalism), we have to manage carefully balance our risks and rewards in order to maximize our odds of success.
If there is no bubble and demand stays high long-term, then the payoff for that risk is potentially enormous.
If there is a bubble and it bursts, then the cost of that risk is potentially devastating.
(Capitalism works most-predictably when cheating is possible, such as with Biff's use of the time machine in Back to the Future II. But without cheats, it's always a gamble.)
Sadly this is not a joke nor some doomsday thinking. Google recently repurposed thousands of phones for server needs. Why? Well guess why… to teach everyone tis coming.
To clarify: UC San Diego is planning to build a cluster of 2000 Pixel phones for computer science class use and Google, in support, helped with a test with 20 phones.
To show they’re working on reducing the impact of data centres on the environment, and that they’re taking action on e-waste, all while saying their pixel phones are so powerful they can be clustered into servers.
And their announced test with 2000 phones, where one server is 25-50 phones, is only 40-80 servers. Interesting, but hardly hyper scale.
It’s not just the hyperscalers, it’s also the boom in personal projects being launched. AI has massively decreased the bar to entry, and a lot more people need servers compared to before
I believe that the explosion of software _production_ in the last 6-12 months in all areas of society is beginning to run headlong into a global bottleneck in the capacity to _deliver_ and _distribute_ that much more software.
Unlike previous boom/bust cycles, I do not believe that demand that's driving the capacity bring-up that's underway right now is going to level off in the near future.
I am personally witnessing teams at my employer that were previously unable to produce software - non-Eng teams - that are now empowered to do so, and some of them are building some cool stuff. I think this is happening at every company everywhere, and I think when we and everyone else solve the "how do we deliver and govern this stuff?" problem there's going to be an even bigger unlock.
True but this is probably because now they have much more demand as other competitors got to expensive and now people are going for the smaller ones even with low service levels
Right, instead of using a managed serverless platform where you pay a premium for proprietary cloud provided services to take away the overhead of managing and patching servers for you and let you just deploy containers, you can… pay a premium for proprietary cloud-hosted AI engines to take away the overhead of managing and patching servers for you and let you just deploy containers.
That's what I'm doing for my personal stuff. It's all running on a pair of Raspberry Pi's. There's an automation in Codex to, once a week, run a ton of health checks, updates, make sure backups are running, etc., and let me know the outcome. Tailscale acting a little wonky? One quick prompt (from my phone! don't even need to sit down at the keyboard any more) and it's sorted out.
It's lowered the threshold for homelab style projects dramatically. It's not doing anything I couldn't have, but the juice wasn't worth the squeeze before compared to, say, Pikapods. Now there basically is no squeeze, just juice.
Using AI to manage self-hosted servers is completely different from using AI to manage VPS servers.
VPSes are usually compared to full-service cloud hosting and especially to serverless offerings, with proponents usually claiming that the costs of self-management of a VPS are vastly less than the premium AWS or Google charge for providing managed compute.
I’m mostly pointing to the irony of eliminating the ‘advantage’ of a VPS (that you can self-manage it like a grown-up, you don’t need a cloud provider to babysit you) by outsourcing management to a stochastic sysadmin who charges by the keypress.
I don't think any VPS provider has come close to raising their prices this much. There's absolutely not enough people flocking to Hetzner to justify this.
That's all true, but this seems disproportionate even to the hardware market. They already raised their prices recently as hardware got more expensive, and now they are quadrupled. Maybe I am just misreading the hardware market right now.
If you see photos/videos of Hetzner datacenters, their servers are essentially plain, low-end motherboards (1G network, few slots) sitting on shelves (don't mean that as derogatory, it's an efficient design). What it does mean is that their per-server costs are absolutely dominated by the very components that are exploding in cost right now: RAM, SSDs and (to a lesser extent) CPUs.
OVH still seems cheaper here for smallest VPS (what's needed to host a personal website), especially when you consider that you have to pay your IPv4 1.7 € with Hetzner.
For the exact same specs (CX23 vs VPS-1 2027), price is 6 vs 4.5 € and you can get a 15% discount on OVH if you order a full year.
A ton of cloud workloads are still running on old Haswell-era CPUs with RAM that was bought a decade+ ago. Probably the costs will be made up with new VM shapes.
I doubt this has to do with the hardware discussion. This is just them increasing their lock-in and trying to curb businesses running to other CDNs (whole point of the peering).
Inching closer to Vultr prices. There are some Rails projects I might have later this year, and I had already been thinking of putting them onto Vultr via Hatchbox since Vultr offered a managed db. Maybe for some stuff that I can run a Rails 8 Solid Stack app with just sqlite, I'd use Hetzner. I tested both with Hatchbox but have nothing in production on either yet and generally use Heroku and Render still.
Has anyone here used Vultr much? I'm curious how they felt about bang for buck. At least with Hatchbox it's easy to run multiple domains on one box.
We use Vultr, DigitalOcean and Hetzner for global coverage. Vultr is by far the worst - some DC like Australia are pretty bad, lots of connectivity issues, some are OK. Their forte is that they offer a lot of DCs. We are migrating some workloads back to DO, where things are usually way smoother. Hetzner is our core, but does not offer DCs in Asia, Africa or Latam.
Oh, didn't realize that. Any recommendation for mitigating that if deployed there? Had thought about putting something on Vultr in a few months. Also open to any other good recommendations for providers that have a managed Postgres.
I use it as a hosting target for automated deployment tooling we wrote. Tools were originally Digitalocean-only and I wondered if LLMs could add support for a second provider, so asked Claude to add Vultr which it did very nicely. But other than run the automated tests (which create VMs, DNS entries, check validity and tear down) and pay the monthly bill, we haven't yet used them for production deployments.
The hyperscalers are the ones buying up all the memory and storage hardware in the first place. They aren't affected by the problem, they are the source of the problem.
They're not the only ones, Anthropic and Google recently announced they'll be using xAI data centers, and I wouldn't be surprised if AWS also started to have capacity problems in the near future.
I’ve already started buying cheap old business PCs just in case I’ll ever need to have simple barebones machines to run things on.