MICHEL MARTIN, HOST:
Demand from AI data centers has caused computer memory to get a lot more expensive. It's subverting a theory that tech tends to get faster and cheaper over time. Our colleagues from The Indicator, Darian Woods and Wailin Wong, explain.
DARIAN WOODS, BYLINE: Steve Burke got into building computers when he was a teenager.
STEVE BURKE: I built my first system when I was 15. So my dad took me to CompUSA, which is no longer in business.
WAILIN WONG, BYLINE: RIP. I remember going to CompUSA with my dad.
(LAUGHTER)
WOODS: Today, Steve runs a website called GamersNexus that covers the computer industry. Over the decades, he's seen the cost of processors and memory and storage come way down.
WONG: But in the last year, the AI industry's demand for RAM - or memory - has sent prices skyrocketing. Some products have doubled in price or more.
WOODS: That's not really supposed to happen because of something called Moore's law. It's named for the late Gordon Moore, the cofounder of Intel and a tech industry pioneer. Here he is back in 2006 when his law was going strong.
(SOUNDBITE OF ARCHIVED RECORDING)
GORDON MOORE: You know that next year, you'll be able to buy something cheaper and more powerful, providing it's electronic.
WOODS: That claim was born out of a prediction he made that the number of transistors - those microscopic switches that make computer chips work - that could fit on a chip would double every year, making electronics faster, smaller, cheaper. This idea became known as Moore's law.
IRIS BAHAR: I love talking about Moore's law.
WONG: Iris Bahar is a computer science professor and head of the Computer Science Department at the Colorado School of Mines.
WOODS: Iris says the computing industry saw tremendous growth in performance through the '90s and early 2000s. But as chips and transistors shrunk in size, achieving these exponential gains got harder.
BAHAR: When you're operating in the single nanometer scale, you really are looking at these, you know, single electrons. This just is a real physical limit.
WONG: To put a nanometer in perspective, a single human red blood cell is 7,000 nanometers wide.
WOODS: Iris says, occasionally, some people in the industry would start drafting the obituary for Moore's law, but then there would be some breakthrough.
WONG: Just last month, IBM announced it had created the world's smallest transistor, measuring under one nanometer.
WOODS: Long live Moore's law.
WONG: But Iris says it is requiring more creativity and money to achieve these advancements. And we can no longer take for granted that improvements in technology are associated with cheaper prices, and that's because of our old friend, supply and demand. Right now, demand is huge for computing power and memory.
WOODS: The RAM-pocalypse (ph).
WONG: Memory prices are going through the roof thanks to the AI boom and the rapid build-out of data centers. These facilities require a lot of computing power and a lot of memory.
WOODS: Take OpenAI. Last year, it announced a partnership where Samsung would be one of the main suppliers of memory for OpenAI's Stargate initiative. This is a $500 billion AI infrastructure project.
WONG: Steve Burke of GamersNexus says the amount of memory OpenAI needs for Stargate amounts to 40% of the world's supply.
BURKE: That was really, for me, the one that stands out as the peak of, OK, this is going to be a problem.
WONG: AI companies have a seemingly unending appetite for computing power and the deep pockets to match. The billion-dollar question is, can technological breakthroughs keep up with the incredible demand for more computing power?
WOODS: Who knows? Maybe in a few years, it will once again be time to dust off the Moore's law death notice.
Darian Woods.
WONG: Wailin Wong, NPR News.
(SOUNDBITE OF OLDSTAR'S "NEWSTAR") Transcript provided by NPR, Copyright NPR.
NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.
