> South Korean SK Hynix has exhausted all of its chip
production for next year and plans to significantly increase investment, anticipating a prolonged "super cycle" of chips, spurred by the boom of artificial intelligence, it said on Wednesday after reporting a record quarterly profit.
> Adata chairman says AI datacenters are gobbling up hard drives, SSDs, and DRAM alike — insatiable upstream demand could soon lead to consumer shortages
Manufacturers learned a valuable lesson a few years ago: overproduction leads to lower prices. Samsung was the first to address this issue by scaling back, and other manufacturers soon followed suit (collusion, cough cough). The past couple of years have been extremely profitable for the entire industry, and they’re not about to increase production and risk hurting their profits.
I suspect they would rather face shortages then satisfy market demand.
Yeah, it's easy to jump to the collusion theory, especially with this industry's, let's say, history. But honestly, I think it's less of an evil conspiracy and more just good old-fashioned fear mixed with inertia. These guys remember getting burned hard by oversupply cycles where they were left with mountains of useless chips. Nobody's gonna drop tens of billions on a new fab that could become a pumpkin in three years if the AI hype train just slows down a little
And on top of that fear, you have the pure technical reality: you can't just flip a switch and start pumping out wildly complex HBM instead of mass-market DDR5. That's like trying to retool a Toyota factory to build Bugattis overnight. So you get this perfect storm: a massive, near-vertical demand spike hits an industry that's both terrified of risk and physically incapable of moving fast. So yeah, they're absolutely milking the situation for all it's worth. But it's happening less because they're master villains and more because they're both scared and incredibly slow
I hope this AI craze will crash soon enough. Maybe then various things normalize in price again. And consumers get cheaper products with less limitations.
Even if the hype around LLMs dies down, the demand for AI compute won't disappear. It will just shift from giant language models to more specialized areas: computer vision, scientific computing (like AlphaFold), drug discovery. All of that requires massive amounts of hardware
Beware of what you wish for. Without the so called AI craze you wont get enough money to fund the current 2-3 years cadence of leading edge Fab development.
> spot prices of DRAM, used in various applications, nearly tripled in September from a year earlier.. improving profitability of non-HBM chips has helped fuel memory chipmakers' share price rally this year, with Samsung's stock up more than 80%, while SK Hynix and Micron shares have soared 170% and 140% respectively... industry is going through a classic shortage that usually lasts a year or two, and TechInsights is forecasting a chip industry downturn in 2027.
Micron has US memory semiconductor fab capacity coming online in 2027 through 2040s, based on $150B new construction.
> Microsoft CEO Satya Nadella has said the company has AI GPUs sitting idle because it doesn’t have enough power to install them.
If the PC supply chain will be impacted by memory shortages until 2027, could Windows 10 security support be extended for 24 months to extend the life of millions of business PCs that cannot run Windows 11?
Hopefully this will put pressure on the market to produce much more efficient AI models. As opposed to bigger, then bigger, and then even BIGGER models (which is the current trend).
FYI: gpt-oss:120b is better at coding (in benchmarks and my own anecdotal testing) than gpt5-mini. More importantly, it's so much faster too. We need more of this kind of optimization. Note that gpt5-mini is estimated to be around ~150 billion parameters.
You're right. And it's not just about parameter count. Efficiency is a full-stack problem: from the architecture (like MoE instead of dense transformers), to the inference techniques (speculative decoding), to the data formats (quantization down to INT4/FP4). The shortage will force everyone to optimize every step of the process, not just "add more layers"
Not just server memory, desktop memory has gone up for the same reason... it's all going to AI. Forget building a new gaming pc, or buying a laptop, or even an arm SBC, because the supply is just gone.
This shortage is the best thing that could have happened for R&D in model efficiency. The "who has more parameters" race is about to hit the physical wall of hardware availability. Now the real race for efficiency begins: quantization, distillation, Mixture-of-Experts, new architectures
Hardware constraints are the single biggest driver of software innovation
23 comments
[ 5.8 ms ] story [ 55.2 ms ] threadEither way, without competition expect it to increase further.
https://www.tomshardware.com/pc-components/dram/openais-star...
> South Korean SK Hynix has exhausted all of its chip production for next year and plans to significantly increase investment, anticipating a prolonged "super cycle" of chips, spurred by the boom of artificial intelligence, it said on Wednesday after reporting a record quarterly profit.
https://en.ilsole24ore.com/art/korean-chip-race-sk-hynix-has...
> Adata chairman says AI datacenters are gobbling up hard drives, SSDs, and DRAM alike — insatiable upstream demand could soon lead to consumer shortages
https://www.tomshardware.com/tech-industry/big-tech/adata-ch...
I suspect they would rather face shortages then satisfy market demand.
And on top of that fear, you have the pure technical reality: you can't just flip a switch and start pumping out wildly complex HBM instead of mass-market DDR5. That's like trying to retool a Toyota factory to build Bugattis overnight. So you get this perfect storm: a massive, near-vertical demand spike hits an industry that's both terrified of risk and physically incapable of moving fast. So yeah, they're absolutely milking the situation for all it's worth. But it's happening less because they're master villains and more because they're both scared and incredibly slow
> spot prices of DRAM, used in various applications, nearly tripled in September from a year earlier.. improving profitability of non-HBM chips has helped fuel memory chipmakers' share price rally this year, with Samsung's stock up more than 80%, while SK Hynix and Micron shares have soared 170% and 140% respectively... industry is going through a classic shortage that usually lasts a year or two, and TechInsights is forecasting a chip industry downturn in 2027.
Micron has US memory semiconductor fab capacity coming online in 2027 through 2040s, based on $150B new construction.
Are some HBM chips idle due to lack of electrical power? https://www.datacenterdynamics.com/en/news/microsoft-has-ai-...
> Microsoft CEO Satya Nadella has said the company has AI GPUs sitting idle because it doesn’t have enough power to install them.
If the PC supply chain will be impacted by memory shortages until 2027, could Windows 10 security support be extended for 24 months to extend the life of millions of business PCs that cannot run Windows 11?
FYI: gpt-oss:120b is better at coding (in benchmarks and my own anecdotal testing) than gpt5-mini. More importantly, it's so much faster too. We need more of this kind of optimization. Note that gpt5-mini is estimated to be around ~150 billion parameters.
Hardware constraints are the single biggest driver of software innovation