New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
What do you think needs to be done to bring RAM prices back to sanity?
ReliableSiteHosting
Member, Patron Provider
in General
What do you think needs to be done to bring RAM prices back to sanity?
Comments
Yes, but for now check Ticket: 093-311EAA80-0054 to bring my sanity back!
my shameless comment
Delete openai, anthropic and gemini.
Releasing HBM4 then HBM4E, DDR6
It's scheduled to be released in 18~24 and 36 months. With HBM4E, there will be generational leap (+100% BW).Thus, reducing RAM demand (DDR5 by 2029)
Tagging @yoursunny to check whether this will turn out to be true or not
I suspect it won't be, unless you're counting on the mass offloading of the ram in current use. The fab capacity is limited and they wound down DDR4 early already so they could put more capacity into the higher margin HBM, I wouldn't be surprised if they also scale back DDR5 for the same reasons.
Sanity.
And: Sanity on the stock market (No, LLMs will not evolve much anymore, any billions/trillions spend/invested on LLM-based "AI" will be a big loss).
Buy Chinese RAM.
https://www.theregister.com/2026/01/20/micron_powerchip_fab_acquisition/
This may help, but not for more than a year...
Came here to say same!!
Yes, it depends on several factors such as:
I think the rate of AI market expansion (within the next 3 years) can't justify the amount of investment ($500-600B). Thus, the excess capacity will left mostly idling. Then, HBM4 & HB4E will come with 1.5x and 2x bandwidth and performance which greatly depreciate the installed capacity.
We are probably going to see AI crash in the next three years I would say, the valuations are unrealistic and LLMs are inherantly flawed and will never become reliable enough.
The only reason they're still running now is investors have fear of missing out, like they did with Blockchain, NFTs and looking back further, the dotcom bubble era. Significant value adjustments will come, and the market for AI will scale back significantly. They won't ever be able to charge enough users what it actually costs to run to just break even.
Patience. Keep on working with existing hardware, there is plenty of it.
@ReliableSiteHosting wanna invest in a fabrication plant?
Same!! China got like two companies engaging in ram production alone. Or maybe just maybe pop the ai bubble.
We are down to the last two 8GB DDR stick in our DELL 7040 Micro, sorry.
We can't give out more DDR.
It's possible to give 160KB RAM on ESP8266 for 40 squats each but the RAM is word addresses and you cannot access individual bytes.
It's SRAM faster than your DRAM because it doesn't need refreshes.
Shipping excluded but we take your UPS account.
Packaging is free because we have a room full of Prime boxes.
If you want more RAM we suggest Fipsy FPGA.
You can write BRAM in Verilog and it can have dual ports.
With dual ports you can do DDR from two processors without interrupts, or form a queue between PCIe DMA and host processor ultra fast.
DDR4 expensive Fipsy only $26.
On the other hand, we added 8GB model for Antarctica IPv9 VPS and upgraded network speeds.
Face covering and push-ups are no longer needed because we received many complaints regarding outdated COVID era protocol.
There is nothing to be done. Even if we load all AI projects with constant requests through bots, if corporations fail, those same corporations will be bailed out by governments using the money taken from (workers) slaves. This system is too corrupted, so it can't come up with a proper solution; it just can't think outside the box with greed and profits and influence.
The investor sees future forecast and thought, "Either we capture this opportunity, or someone else will", without doing research whether the market will expand at the prescribed pace. They will eventually realized they spent hundreds of million to billions of dollar on flawed market research and demand concrete result (reaching an impasse and profit).
To break even, they should stop speculating on how much capacity they need. To compete, they should be aware of upcoming hardware refresh which will determine the cost structure of their and their competitors token pricing
I'm going to have a HUGE homelab once this ai bubble pops. I keep hoping an llm will replace me but so far no luck...
unfortunately, I don't think it will turn out like this, this time, at least not for many many years.
Make the AI bubble burst + have large quantities of chinese memory to force the "big 3" to drastically lower prices.
Plus IMO from a political perspective harshly punish the big 3 and painfully remind them that focusing only on profit is not their main task. Which is to keep the industry alive and well by providing products at reasonable prices.
AFAIC I will try hard to avoid products from the big 3 and strongly prefer chinese RAM.
Personally I give it about a year. I think we'll see little improvement in the frontier models, and with those stagnating, people will realize that it's not going to replace them yet. Obviously I could be wrong, but these last model updates have been pretty crap.
That sounds amazing and horrible at the same time.
What if it keeps doubling down even if AI proves unsuccessful? What if this AI project keeps borrowing because the governments will step in to save it?
Even if there is a bubble and it does pop, I'm not sure that it will lead to a lot of available hardware for home labs. A lot of the stuff in datacenters is bespoke and requires exotic power connections, weird water cooling, etc... Don't get me wrong, I hope you're right and I'm wrong. I want a giant weird homelab too.
In two to three years (after the ai bubble pops), just search ebay for "44MW DC to sale" - their is your giant weird homelab at a bargain (sold my meta).
Unfortunately this is just how most investing ends up working, most people don't invest in things they truly understand, just what is performing well. It's just a system based on greed, not for investing in genuinely good technology. At this point I think the investors may have realised and they're just in the sunk cost fallacy.
After the crash, LLMs will only gain major popularity again after computing power is cheap enough, the problem right now is that computing power isn't really cheap enough to run them well and that's why it cannot be profitable unless run at smaller scale with worse models. Either way, it's mathematically impossible to avoid hallucination with LLMs, so I doubt they will ever be trusted again after all this. There's already quite a few scandals happening in relation to it being unreliable.
Bubble like any other, once it pops, situation will be resolved. Kinda normal economic happening.
I want Ali Express to have a flood of Epyc+motherboard+RAM deals like we used to get with annual refreshes. With companies going more proprietary, I'm hoping that makes even more consumer stuff to hit the market.
It might lead to a rise of SOC type packages that include the RAM with the CPU. Intel have said doing that on Lunar Lake was a mistake, but I'm less sure. They should be able to get decent discounts on the volume they'd be buying and customers would know it's certified to work.
damn, too many ppl want AI bubble to burst.
it might burst because valuations are very high,
but LLMs advancing at rapid rate so AI gonna stay here and it will remove jobs of many people.
it already started at surface level.