Yesterday I did a price check on the PC I built two years ago. It went from $2300 to $3650. The bulk of that increase was that the ram went from $210 to $940. Its now more expensive than when DDR5 was new.
The memory in the PC I put together early last year is now worth about three times the total cost of all the parts I used to build the thing. It is absolutely crazy.
I regret not building the PC when I was looking at it. It's not a money thing, at the end of the day, but I can't bring myself to do it.
I had it all priced out, but a bunch of birthdays in my family were coming up and I felt like I shouldn't buy something for myself if it's really their time.
My old laptop will have to cut it for a while. :-)
GN did a documentary on the situation from the perspective of consumer-facing companies. Seems pretty dire for them, and it's hard to see the long-range consequences, but the idea of consumers being priced out isn't too far out, which to me is a little alarming.
"The costs are negligible and justified when compared to all the benefits. If you look at the performance gains, the overall cost has in fact been reduced."
- Altman, a.k.a. Dory from Finding Nemo and/or Dario, a.k.a. Carl from Jimmy Neutron.
My bet is that the prices will crash once OpenAI (and/or Antrophic) IPO's have happened.
Right now the biggest threat to their IPO's is that people realize that local models are good enough for whatever they're peddling, what's the most important factor to even running good enough models? RAM since you want the models in memory to not be total slogs.
They are actually quite a bit better than you might think. Qwen3.6 27B is pretty capable at coding.
For non-coding work, they are more than good enough. A lot of the ways my non-technical family members have interacted with AI would be perfectly served by using a local model.
After all, people were more than satisfied with the results from GPT 3. That has long since been surpassed by open weight models.
Honestly, that's the output I get from non-local models, anyway. If I'm going to get plausible nonsense either way, I may as well run it on my own hardware.
I looked at my eBay receipt in 2023 and I paid $84.98 for a "Kingston FURY Beast 64GB (2x32GB) 3200MHz DDR4" listing and now the equivalent on eBay "Buy It Now" is $374.99 for "Kingston FURY Renegade 64GB (2x32GB) DDR4 RAM 3200MHz (KF432C16RBK2/64)". What a timeline it has become for consumer computing three years later.
My anecdata: bought used DDR4 ECC 16gb ram sticks (i.e. serverpulled ram) for $11 off ebay last year (Jan 2025), now as of June 2026 the lowest listing I see is $42, most are around $50.
To be fair, DDR3 is still kinda cheap. It's really only DDR4 and DDR5 that are massively in demand. DDR3 is a bit too old for it to be in high demand by consumers.
While DDR3 is cheaper it's still tripled or quadrupled in price over the past two years. I just bought a pair or sticks for an old Mac Pro and it was 4x what i paid just a few years ago.
I recently booted up an old 4790k system and it was fine on Linux but on Windows it would nag me to update but apparently the CPU is too old for new Windows. I ended up giving it away on the Internet to whomever could pick up but afterwards it ended up with one of those reseller chaps. Ah well, I wish it had made to some kid somewhere.
I played 100+ hours of RDR2 on a 2060 (non-super), as was the style at the time. When the 30 series came out, I sold that card for more than I paid for it.
Pretty insane. I built a Framework Desktop PC back in November. The motherboard (with 128G DDR5 RAM) was $1800. Now it's $2859. Almost 60% increase in 6 months.
I wanted to upgrade my SSD but prices are more than at the end of 2025. I refuse to pay 500 euro for a 4TB SSD. I rather go outside and play with my bike like when I was 5.
AMD just brought the popular 5800X3D back out of retirement to give people maintaining the DDR4 based platforms something to buy. Last I checked used DDR4 was half the price of used DDR5 after the prices of both shot up.
Crazy, the other day I looked in my local store order history and say that I bought G.SKILL RipJaws V F4-3600C18D-32GVK, a DDR4 32 GB 3600 MHz kit of two sticks.
I bought it for 82 EUR, before the whole ongoing situation.
Now the same spec costs upwards of 290 EUR, about 3.5x the original price and even on Amazon the best prices I can find are upwards of like 210 EUR (2.5x).
This is insane. Didn’t know how bad it got. I bought a mini PC a few years ago with 64 gigs in it for a home VM server for like $600 total. Looks like I’m keeping it a while.
Usually these bottlenecks lead to a price crash later. Of course that’s also part of what fuels the bottleneck. Companies are afraid of over investing in production and being left with underwater capital later.
Price should send a signal to manufacturers to build more capacity. I wonder if they will though, it takes quite a bit of time, and it's not certain that the demand will continue to exist once built.
Several Chinese manufacturers are doing just that, and have already expanded production: https://techwireasia.com/2026/04/chinese-memory-chips-ymtc-c... But because of tech trade barriers their primary focus is on the domestic market and only secondarily global markets.
Commerce Secretary Howard Lutnick has warned Samsung and SK Hynix they could face 100% tariffs, framing it as a choice between paying a 100% tariff or building memory fabs in America...
The bottleneck isn't the sticks, it's the chips. The chips are the same for consumer and server applications. What's been happening is that big companies have bought nearly all the wafer capacity for the next year or so, and perhaps some of that capacity has also been redirected from DDR5 to LPDDR5. If a stick manufacturer drops out of the consumer market that kinda doesn't matter, because manufacturing sticks is relatively low tech compared to manufacturing the memory chips. You can compare it to manufacturing video cards vs. manufacturing GPUs (as in the actual processing elements).
Government needs to get out of the way. Micron announced a memory fab in Syracuse in 2023. It took 3 years, 20,000 pages of "environmental review", deals with the government on amount of union contracts during building, etc. for them to break ground in 2026 for a 2030 opening date. In any reasonable world, a 2023 announcement should have broke ground in 2023.
OTOH, a celulose factory near me, built in the 1950's, got their permits fast and with little regard to environment. FF three decades, and their entire surroundings are destoyed for everyone else. Trials go nowhere, because they have all authorizations needed (and a lot of political leverage because they are the main employer in the region). Careful fast-tracking business that have zero incentives to avoid externalization of costs.
And? The primary goal should be to catch and stop pollution, not make manufacturers spend years promising not to do something they're not allowed to do. If someone wants to build a factory that can't operate without illegal emissions, then so be it. It's their money lost. All that matters is that they don't actually pollute.
Using red tape as some kind of prophylactic is ridiculous. If the state doesn't have the monitoring in place, you have to just trust the company, which is naive if not negligent. If you do have the monitoring, why require the extremely expensive song & dance? To protect corporations from negligently wasting money?
Answer: because the song & dance is primarily about extracting concessions, like union labor or even cash (e.g. promises to pay to fix someone else's pollution, or contributions to various interest groups). The friction and expense involved in today's beaucratic development review process is many times more costly to all involved than the social benefit.
"back in my time", a dialup and 32 megs(!) of ram was enough for most stuff, including internet browsing.
I have no idea why a weather forecast site needs tens of megabytes of resources, and gig+ of ram for my browser, since i get no more info from it, than i did back then. Same for chat programs (how is discord different than irc? and why does it need so much ram to do so? same for slack), mail clients, etc.
Maybe it's time to kick developers to start optimizing stuff a bit, since neither they nor the users can't afford "unlimited" ram anymore.
edit: i'm not saying we need to get back to literally 32 megs of ram, just to make developers performance test their stuff on a laptop that was on sale 3 years ago in their local supermarket, i.e. stuff their users use at home.
It'll calm down once the Antrophic and/or OpenAI IPO's are done, no need to protect themselves from people running local models by buying everything once the bosses have gotten their money.
OpenAI and Anthropic are certainly strong drivers, but there's a large demand from many other players: cloud provider, accelerator vendors, and so on. I think there's no end in sight.
This is textbook negative externalities, of the AI buildout on everyone who isn't using RAM/GPUs for AI, of the use of electricity and water on anyone who isn't using it for AI. The cynic in me thinks this will go down in history alongside asbestos, leaded gasoline/paint, and the opioid crisis.
Like clockwork, people naturally want to have their cake and to eat it too, so there will be the incessant complaining about the externalities. Half the people lack the brainpower to see the good and bad are intrinsically linked, and the other half just like complaining.
But at least for now, both halves aren't pulling back (in fact it's increasing), and money, not complaining, steers the ship.
We can be cynics of AI without ignoring reality, if no one wanted this no one would be chatting with Claude or ChatGPT directly, but people obviously are.
The fact is there are people that do in fact want this, and it isn't just CEO's hoping to cut jobs.
Honestly that seems slightly down even if it’s still ridiculous. The ram I bought for $100 a year ago was $500 a couple of months ago. Could just be the particular sticks I got though
$375 is the cheapest kit, and the price is using a promo code
"Price tracking courtesy of PCPartPicker now reveals the cheapest 32GB DDR5 RAM you can buy is $375. Specifically, four XPOWER kits from Silicon Power will set you back $374.97 thanks to a promo code."
I had it all priced out, but a bunch of birthdays in my family were coming up and I felt like I shouldn't buy something for myself if it's really their time.
My old laptop will have to cut it for a while. :-)
https://m.youtube.com/watch?v=zyQwAhppWj8
- Altman, a.k.a. Dory from Finding Nemo and/or Dario, a.k.a. Carl from Jimmy Neutron.
Right now the biggest threat to their IPO's is that people realize that local models are good enough for whatever they're peddling, what's the most important factor to even running good enough models? RAM since you want the models in memory to not be total slogs.
Because what to do with power-consuming outdated hardware ? let's say 5 years from now ?
They will need new RAM.
I wonder.
they are not. Unless you are satisfied with plausible, but mostly garbage output.
For non-coding work, they are more than good enough. A lot of the ways my non-technical family members have interacted with AI would be perfectly served by using a local model.
After all, people were more than satisfied with the results from GPT 3. That has long since been surpassed by open weight models.
I wanted to upgrade my SSD but prices are more than at the end of 2025. I refuse to pay 500 euro for a 4TB SSD. I rather go outside and play with my bike like when I was 5.
Would be nice to be able to own property.
The 5700X3D has been the smarter pick back then, it fits to the current latent user hostility of AMD to focus on the more expensive processor.
[1] https://tweakers.net/pricewatch/1419292/corsair-vengeance-lp...
I bought it for 82 EUR, before the whole ongoing situation.
Now the same spec costs upwards of 290 EUR, about 3.5x the original price and even on Amazon the best prices I can find are upwards of like 210 EUR (2.5x).
Looking at it from that frame, it seems reasonable.
I thought 128MB of SDRAM was a good deal at $100.
I also thought $479 for 32GB of DDR4 was nuts back in 2016/2017.
Usually these bottlenecks lead to a price crash later. Of course that’s also part of what fuels the bottleneck. Companies are afraid of over investing in production and being left with underwater capital later.
https://www.bloomberg.com/news/articles/2026-06-02/sk-hynix-...
It will happen, but yeah it takes time and money
https://techwireasia.com/2026/04/chinese-memory-chips-ymtc-c...
https://wccftech.com/another-chinese-dram-maker-breaks-into-...
Punishing instead of supporting.
Using red tape as some kind of prophylactic is ridiculous. If the state doesn't have the monitoring in place, you have to just trust the company, which is naive if not negligent. If you do have the monitoring, why require the extremely expensive song & dance? To protect corporations from negligently wasting money?
Answer: because the song & dance is primarily about extracting concessions, like union labor or even cash (e.g. promises to pay to fix someone else's pollution, or contributions to various interest groups). The friction and expense involved in today's beaucratic development review process is many times more costly to all involved than the social benefit.
This is similar situation to housing market. Prices are going up and supply is being restricted by whatever means.
It will be a bit of Catch 22.
This is the stupidest freaking timeline...
I have no idea why a weather forecast site needs tens of megabytes of resources, and gig+ of ram for my browser, since i get no more info from it, than i did back then. Same for chat programs (how is discord different than irc? and why does it need so much ram to do so? same for slack), mail clients, etc.
Maybe it's time to kick developers to start optimizing stuff a bit, since neither they nor the users can't afford "unlimited" ram anymore.
edit: i'm not saying we need to get back to literally 32 megs of ram, just to make developers performance test their stuff on a laptop that was on sale 3 years ago in their local supermarket, i.e. stuff their users use at home.
They will even tell you "it's not a static website", thinking that there were no other ways to add dynamic behavior other than using SPAs.
And they are hiring MIT-bred Leetcode ninjas at 300k+ in most of these startups/big-techs.
Like clockwork, people naturally want to have their cake and to eat it too, so there will be the incessant complaining about the externalities. Half the people lack the brainpower to see the good and bad are intrinsically linked, and the other half just like complaining.
But at least for now, both halves aren't pulling back (in fact it's increasing), and money, not complaining, steers the ship.
It's impossible to avoid using AI multiple times a day, just because it's forced into every product under the sun.
That is NOT demand. None of those users WANT this.
The fact is there are people that do in fact want this, and it isn't just CEO's hoping to cut jobs.
There was a period in 2012-2016 when things were pretty nice.
"Price tracking courtesy of PCPartPicker now reveals the cheapest 32GB DDR5 RAM you can buy is $375. Specifically, four XPOWER kits from Silicon Power will set you back $374.97 thanks to a promo code."
its paired to a 5950x so im sure it will be fine for a few more years