Thank you for this very informative article. I was trying to read through a conclusion with possible solutions or insights as to when it will improve, but you didn't include that. Eventually this bubble, like all bubbles, will pop. When is anyone's guess, but there will come a time when the lack of supply and inflated prices will kill off demand.
AI Is Eating All the DRAM. DDR5 Prices Just Doubled. GPUs Could Be Next.
- Thread starter Scorpus
- Start date
yRaz
Posts: 9,548 +16,843
Venture capital must love you. Oh, wait, banks and VC stopped funding AI months ago and that's why Sam Altman is publicly begging for money
When every industry on the planet is now using AI for something, you just haven't for the right tool.
AI by itself is not bad (some of the tools are legit awesome), it's just that people expect it to become so big and huge that the massive investments made now will pay off... and I doubt that the vast majority will ever manage to make a profit.
ChipBoundary
Posts: 156 +79
AI used for scientific research has been insanely successful. It has led to the discovery of a bunch of game-changing things. It has accelerated scientific advancements across all industries. A five second Internet search would tell you this.
umeng2002
Posts: 383 +567
Lol, no.
Did someone say tariffs? Not that I could find here. No doubt the AI stampeded is a major factor but outrageous Trumpy's tariffs (aka taxes) took effect in last 3 months, late summer. Doctor Google says tariffs on RAM range from 20% for some countries and up to 60% for China. How come no media articles are mentioning that? Am I wrong? Or are AI written articles blaming AI? Frankly AI in Windows 11 and Microsoft Office is just annoying. And HDD drives are still a under $200 for 8TB; not having increased more that $10. NVME 4TB is $300 or more. Do we really needed to pay for SSD or Cloud storage rather than using a one time purchase HDD? Buy that HDD before the AI rage catches up to that storage media.
Last edited:
TechSavy101
Posts: 93 +118
Once again this happens right before launch of new nvidia products, coincidence? I think not, especially since Nvidia is in a position that can sign a few contracts to create shortages and sell their GPUs at higher prices. It happened in 2020, then the launch of 5000 series and now. This AI bubble would have already bursted if it wasn't beneficial for government, e.g new spying tools. Patriot act of 2001, they're tracking everyone; AI can replace some agents in analytics.
Megalomaniac
Posts: 1,199 +1,257
Fortunately, prebuilts are mostly OK.
The power grid bottlenecks are real, a lot of these new facilities can’t actually turn on because utilities can’t deliver what was promised on the timelines investors assumed.
That said, I think the situation is less pure bubble and more infrastructure whiplash. Companies over ordered GPUs because supply was unpredictable, demand was surging, and nobody wanted to get caught empty handed.
Now everything is bottlenecked by power, permitting, and transmission....not by lack of hardware. It creates the illusion of “shortage” when the real scarcity is grid capacity and physical buildout.
Will some of this shake out? Absolutely.
Not every company buying compute at insane valuations is going to survive, and some correction is probably healthy. But I’m not convinced the underlying trend goes away. The demand for AI workloads is very real and the industry just scaled in the messiest, most financially reckless way possible.
PPP - Piss Poor Planning
And there it is...it had to be dropped..the Tariff angle...
The tariff angle is getting blown way out of proportion here. Acting like this is the primary driver behind GPU shortages, data center delays, or AI hardware hoarding is TDS dancing in your head, I swear he lives there rent free.
If tariffs were the core issue, we’d see broad price spikes across the entire storage and compute market. But we’re not. HDD prices are flat. SSDs ticked up a bit, but nothing outside normal supply/demand swings. RAM prices have been rising for over a year due to AI demand from hyperscalers, not because someone in Washington DC slapped a tariff on DIMMs last quarter. Tariffs don’t create multiyear global backlogs of HBM or push NVIDIA to entire year allocation models.
The real bottleneck isn’t tariffs.....it’s power and infrastructure.
Data centers are literally being built faster than utilities can supply megawatts. Companies are buying hardware because workloads are exploding, but they can’t actually turn the buildings on. That’s why you see GPUs sitting in warehouses....not because of taxes, but because grid capacity and permitting move on decade long timelines.
The “NVIDIA is creating artificial shortages before every launch” thing just doesn’t match reality.
They don’t need to engineer shortages...hyperscalers are buying entire global production runs the moment they exist. The bottleneck isn’t some secret contract, it’s supply, packaging capacity, and the fact that data centers can’t get enough power to actually run what they’re building.
What happened in 2020 wasn’t a scheme .... it was a pandemic, crypto miners, and supply chain collapse. No conspiracy required.
And the idea that the AI boom is only alive because governments want more surveillance tools oversimplifies everything. Agencies have had analytics tools for decades. The massive demand today is coming from cloud providers and enterprises, not the NSA.
You can criticize NVIDIA’s pricing or the hype cycle, but global shortages aren’t being orchestrated from a back room. The boring explanation, real demand and real bottlenecks fits the facts a lot better than the conspiracy version.
yRaz
Posts: 9,548 +16,843
What people don't understand about that is that you don't need hyperscale level equipment to do that. Go look on github and you'll see that the most interesting and cutting edge stuff going on with AI is, I'm over simplifying this, is being made on threadripper systems with 8 5090s. Also,LLMs are not AI.
Solution to the problem, never pay for any AI service, they'll eventually collapse. They spent much more than the money they make and shareholders are starting to notice the huge gap. Also don't buy Ram as long as the prices are high. It's a cartel that has done this for a very long time.
That’s not really a “solution,” it’s just wishful thinking. AI companies aren’t going to collapse because a handful of people decide not to pay for anything, there’s massive enterprise, cloud, and licensing money behind the scenes that keeps them afloat regardless of whether you or I buy.
As for RAM, prices fluctuate because of supply cycles, production cuts, and demand spikes. Calling it a “cartel” every time prices go up ignores how the semiconductor industry actually works. If anything, holding off during peak pricing makes sense, but pretending there’s some grand unified conspiracy doesn’t.
unoficialoficial
Posts: 370 +325
So, given the special circumstances of the AI era, it’s clear that manufacturers like Intel and AMD should integrate a substantial amount of L4 RAM—such as 64+ GB—as chiplets directly on the CPU die or in close proximity, delivering ultra-fast memory access as a standard minimum configuration for all CPUs. Additionally, they should expand available lanes and support for extra RAM, scaling up to 1 TB for high-performance AI driven commercial applications.
Anton Longshot
Posts: 551 +831
OortCloud
Posts: 2,331 +3,364
Absolutely. No doubt about it.
Cartel is a historical reality. It's been a while there's no competition anymore. And I could say the same about cars, washing machines or fridges. About you're first point, I agree. I dropped this as a bottle in the ocean, as you said wishful thinking.
ChipBoundary
Posts: 156 +79
What are you smoking? LLM's are absolutely AI and have enabled just as many advancements in technology and science as any other type. Please stop commenting nonsense. And GitHub is such a small, small space of where any of this stuff is happening. Essentially insignificant. The most interesting and effective stuff is happening at hyper scale in labs from major companies. 90% of the stuff that is happening you've never seen or touched and you likely never will.
ChipBoundary
Posts: 156 +79
Yes, finely stated and backed by research. Classic Internet...
To keep with the vernacular and poor grammar: Lol, yes.
Do some actual research and then put down the Internet, because you're under qualified to use it.
yRaz
Posts: 9,548 +16,843
It's not AI. Once a model is trained that is the only level it can ever function at. It can't modify, upgrade or correct itself. As ChatGPTs latest model showed, it was actually a regression instead of an improvement dispite having access to vastly more GPUs. It's fine for stuff where they train it to do a singular application very well, like AlpaFold. But there isn't more Data to train AI anymore. AI needs lots of human intervention to improve. This isn't AI, it's just a program that uses inferencing and they're using AI as a marketing term.
Instead of the Turing test I'm going to start calling it the Altman test. Can an LLM pretend that it's indistinguishable from an AI?
Similar threads
Latest posts
-
France's court just blocked Macron's plan to ban kids from social media- Megalomaniac replied
