Eugene Mymrin | Second | Getty Photographs
All computing units require an element referred to as reminiscence, or RAM, for short-term knowledge storage, however this yr, there will not be sufficient of those important parts to fulfill worldwide demand.
That is as a result of firms like Nvidia, Superior Micro Gadgets and Google want a lot RAM for his or her synthetic intelligence chips, and people firms are the primary ones in line for the parts.
Three major reminiscence distributors — Micron, SK Hynix and Samsung Electronics — make up practically the complete RAM market, and their companies are benefitting from the surge in demand.
“We have now seen a really sharp, important surge in demand for reminiscence, and it has far outpaced our capability to produce that reminiscence and, in our estimation, the availability functionality of the entire reminiscence trade,” Micron enterprise chief Sumit Sadana informed CNBC this week on the CES commerce present in Las Vegas.
Micron’s inventory is up 247% over the previous yr, and the corporate reported that web earnings practically tripled in the latest quarter. Samsung this week mentioned that it expects its December quarter working revenue to just about triple as effectively. In the meantime, SK Hynix is contemplating a U.S. itemizing as its inventory value in South Korea surges, and in October, the corporate mentioned it had secured demand for its complete 2026 RAM manufacturing capability.
Now, costs for reminiscence are rising.
TrendForce, a Taipei-based researcher that intently covers the reminiscence market, this week mentioned it expects common DRAM reminiscence costs to rise between 50% and 55% this quarter versus the fourth quarter of 2025. TrendForce analyst Tom Hsu informed CNBC that sort of improve for reminiscence costs was “unprecedented.”
Three-to-one foundation
Chipmakers like Nvidia encompass the a part of the chip that does the computation — the graphics processing unit, or GPU — with a number of blocks of a quick, specialised part referred to as high-bandwidth reminiscence, or HBM, Sadana mentioned. HBM is usually seen when chipmakers maintain up their new chips. Micron provides reminiscence to each Nvidia and AMD, the 2 main GPU makers.
Nvidia’s Rubin GPU, which not too long ago entered manufacturing, comes with as much as 288 gigabytes of next-generation HBM4 reminiscence per chip. HBM is put in in eight seen blocks above and under the processor, and that GPU will probably be bought as a part of single server rack referred to as NVL72, which fittingly combines 72 of these GPUs right into a single system. By comparability, smartphones sometimes include 8 or 12GB of lower-powered DDR reminiscence.
Nvidia founder and CEO Jensen Huang introduces the Rubin GPU and the Vera CPU as he speaks throughout Nvidia Reside at CES 2026 forward of the annual Client Electronics Present in Las Vegas, Nevada, on Jan. 5, 2026.
Patrick T. Fallon | AFP | Getty Photographs
However the HBM reminiscence that AI chips want is way more demanding than the RAM used for customers’ laptops and smartphones. HBM is designed for high-bandwidth specs required by AI chips, and it is produced in an advanced course of the place Micron stacks 12 to 16 layers of reminiscence on a single chip, turning it right into a “dice.”
When Micron makes one little bit of HBM reminiscence, it has to forgo making three bits of extra standard reminiscence for different units.
“As we improve HBM provide, it leaves much less reminiscence left over for the non-HBM portion of the market, due to this three-to-one foundation,” Sadana mentioned.
Hsu, the TrendForce analyst, mentioned that reminiscence makers are favoring server and HBM purposes over different purchasers as a result of there’s larger potential for progress in demand in that enterprise and cloud service suppliers are much less price-sensitive.
In December, Micron mentioned it could discontinue part of its enterprise that aimed to supply reminiscence for client PC builders so the corporate might save provide for AI chips and servers.
Some contained in the tech trade are marveling at how a lot and the way shortly the worth of RAM for customers has elevated.
Dean Beeler, co-founder and tech chief at Juice Labs, mentioned that a couple of months in the past, he loaded up his laptop with 256GB of RAM, the utmost quantity that present client motherboards help. That value him about $300 on the time.
“Who knew that will find yourself being ~$3,000 of RAM just some months later,” he posted on Fb on Monday.
‘Reminiscence wall’
AI researchers began to see reminiscence as a bottleneck simply earlier than OpenAI’s ChatGPT hit the market in late 2022, mentioned Majestic Labs co-founder Sha Rabii, an entrepreneur who beforehand labored on silicon at Google and Meta.
Prior AI programs had been designed for fashions like convolutional neural networks, which require much less reminiscence than massive language fashions, or LLMs, which can be fashionable right this moment, Rabii mentioned.
Whereas AI chips themselves have been getting a lot quicker, reminiscence has not, he mentioned, which results in highly effective GPUs ready round to get the info wanted to run LLMs.
“Your efficiency is proscribed by the quantity of reminiscence and the pace of the reminiscence that you’ve got, and should you maintain including extra GPUs, it is not a win,” Rabii mentioned.
The AI trade refers to this because the “reminiscence wall.”
Erik Isakson | Digitalvision | Getty Photographs
“The processor spends extra time simply twiddling its thumbs, ready for knowledge,” Micron’s Sadana mentioned.
Extra and quicker reminiscence implies that AI programs can run greater fashions, serve extra clients concurrently and add “context home windows” that enable chatbots and different LLMs to recollect earlier conversations with customers, which provides a contact of personalization to the expertise.
Majestic Labs is designing an AI system for inference with 128 terabytes of reminiscence, or about 100 occasions extra reminiscence than some present AI programs, Rabii mentioned, including that the corporate plans to eschew HBM reminiscence for lower-cost choices. Rabii mentioned the extra RAM and structure help within the design will allow its computer systems to help considerably extra customers on the identical time than different AI servers whereas utilizing much less energy.
Offered out for 2026
Wall Road has been asking firms within the client electronics enterprise, like Apple and Dell Applied sciences, how they are going to deal with the reminiscence scarcity and in the event that they could be compelled to lift costs or reduce margins. As of late, reminiscence accounts for about 20% of the {hardware} prices of a laptop computer, Hsu mentioned. That is up from between 10% and 18% within the first half of 2025.
In October, Apple finance chief Kevan Parekh informed analysts that his firm was seeing a “slight tailwind” on reminiscence costs however he downplayed it as “nothing actually to notice there.”
However in November, Dell mentioned it anticipated its value foundation for all of its merchandise to go up on account of the reminiscence scarcity. COO Jefferey Clarke informed analysts that Dell deliberate to vary its mixture of configurations to attenuate the worth impacts, however he mentioned the scarcity will possible have an effect on retail costs for units.
“I do not see how this is not going to make its manner into the shopper base,” Clarke mentioned. “We’ll do every little thing we will to mitigate that.”
Even Nvidia, which has emerged as the most important buyer within the HBM market, is dealing with questions on its ravenous reminiscence wants — particularly, about its client merchandise.
At a press convention Tuesday at CES, Nvidia CEO Jensen Huang was requested if he was involved that the corporate’s gaming clients could be resentful of AI expertise due to rising recreation console and graphics playing cards costs which can be being pushed by the reminiscence scarcity.
Huang mentioned Nvidia is a really massive buyer of reminiscence and has lengthy relationships with the businesses within the house however that, finally, there would must be extra reminiscence factories as a result of the wants of AI are so excessive.
“As a result of our demand is so excessive, each manufacturing facility, each HBM provider, is gearing up, they usually’re all doing nice,” Huang mentioned.
At most, Micron can solely meet two-thirds of the medium-term reminiscence necessities for some clients, Sadana mentioned. However the firm is at present constructing two huge factories referred to as fabs in Boise, Idaho, that may begin producing reminiscence in 2027 and 2028, he mentioned. Micron can be going to interrupt floor on a fab within the city of Clay, New York, that he mentioned is anticipate to come back on-line in 2030.
However for now, “we’re bought out for 2026,” Sadana mentioned.
