430
32GB of DDR5 now costs $375 minimum — AI shortage continues to squeeze PC building
(www.tomshardware.com)
This is a most excellent place for technology news and articles.
So, two points.
First, new memory fabs start coming online in 2027, and there are more being constructed that will be coming online in subsequent years.
But, second...I think that some perspective is in order. Set new production aside. Let's imagine a world where that didn't happen. In fact, let's imagine that not a single additional memory chip was going to be produced. Video games were around when I played games on an Atari 2600, to pick an early video game console. I had fun with it. It didn't have the latest, real-time rendered photorealistic graphics. But...the Atari 2600 had 128 bytes of memory. Not gigabytes, not megabytes, not kilobytes. Bytes.
There are people building microcontrollers right now that have onboard memory, and those aren't impacted by this. It's just the high-density dedicated memory chips that go on DIMMs that are seeing all that demand.
According to Wikipedia, there were 30 million Atari 2600s made. The CPU I currently have in my desktop has a little over 145MB of onboard cache. Twenty-six of those CPUs, looking just at their onboard cache, no external memory from Micron/Samsung/SK Hynix, have more memory than all of the 30 million Atari 2600s ever manufactured, combined.
Like, don't get me wrong. I enjoy using all this memory that we have had available in recent years. But...video games are here to stay and would be even if no dedicated memory chips were around.
Where is this optimism coming from? China? Korea?
There are three major DRAM chip manufacturers: Micron, in the US, and Samsung and SK Hynix, both in South Korea.
Micron has two new fabs coming online in Boise, Idaho. The earliest one is scheduled to start operation in the first half of 2027 (they recently announced that they'd moved that timeline up from the second half of 2027) though it'll take time to ramp up; it will not be doing output at full capacity immediately when it first starts up.
https://www.micron.com/us-expansion/id
They announced late last year that they were going to do a second one as well for more capacity.
They also have New York fabs that they're doing:
https://www.micron.com/us-expansion/ny
For the South Korean manufacturers:
https://koreajoongangdaily.joins.com/news/2026-03-12/business/industry/Samsung-and-SK-are-expanding-fast-but-why-is-memory-still-in-short-supply/2540153
Samsung
SK Hynix
That's detailed! That's good and there are also Chinese ramping up memory production (I wish EU did something as well), but sadly it'll still take at least a year if not more. The interesting situation would arise if AI bubble bursts in a year leaving us with huge memory surplus. One can hope, right.
The memory surplus wouldn't be immediate after the bubble pops; at least not for regular people. What they're currently producing isn't one-to-one compatible with desktop PCs - most of the secondhand stuff from decommissioned AI datacenters wouldn't be usable outside of servers, and it'd take a while for the newly freed fabs to start churning out consumer-grade memory again, factories to install it on consumer chips, and for it to make its way to the market (mass shipping is much slower than people think). That delay would hit producers hard, possibly gumming up the works even further. Modern economics is not at all equipped for supply chain failures.
There's a reason people are panicking about this bubble, and that's not even going into the far more devastating stock market crash likely to happen when it pops. It's a nightmare in both economic and technological terms, but a small group of people stand to make a ton of money from it so they've gutted the regulatory agencies that would have prevented things from getting this bad, or at least softened the blow.