Three companies manufacture RAM at scale, after triple-digit gains, is there still room to run?
First the AI buildout needed graphics processing units (GPUs), then it needed power generation, now it needs memory. While those first two components, among others, are still in high demand from all the AI data centers being built worldwide, the supply of dynamic random access memory (DRAM) is the latest in a string of rolling bottlenecks associated with the multi-trillion dollar AI infrastructure project now underway. What makes memory such an interesting component, too, is that it’s only produced at scale by three companies: SK Hynix and Samsung in South Korea and Micron Technology in the United States. Year to date, SK Hynix and Micron have seen the stock prices more than double, while Samsung isn’t far behind with 80 per cent growth in 2026 so far.
For investors, the question now is whether those memory names still have room to run. Two CIOs offered a consensus view on that question, which each detailing why memory has become such a crucial component of the AI buildout. Elliot Johnson, CIO at Evolve ETFs, stressed the ongoing macro need for memory emerging from the AI trend. James Learmonth, Co-CIO and Portfolio Manager at Harvest ETFs, highlighted the scale that the three companies building DRAM memory can command and the discipline they’ve shown in capturing upside in this new gold rush.
“We’re starting to see something that seemed unthinkable just a couple of years ago. We’re seeing long term supply commitments and agreements with hyperscalers, trying to ensure they’ve got supply,” Learmonth says. “That’s a really unique and interesting change for the industry. It’s very early on as to how it’s going to play out but, but it’s quite possible that over time investors start to look at these companies more through a longer term growth lens with a much more predictable cycle as opposed to the boom and bust cycles that we’ve seen historically.”
Learmoth explains that the memory market became a three-player system after the initial internet boom of the late ‘90s and early 2000s. As demand started to peter out, memory hardware became cyclical and started to behave like a commodity. Smaller companies couldn’t survive price fluctuations and these three companies emerged as the top players in the space. AI changed all of that.
AI requires what’s called a “context window” to function. When you ask an AI to do a task or answer a question, there’s a certain amount it needs to remember in order to answer that question or perform that task effectively. The more it can remember, the better it is. Memory is also essential in the training of AI, as it recalls more and more of the feedback it was given to get smarter and more effective at completing its tasks. Elliot Johnson believes that this core use of memory in AI has made demand skyrocket today, and should push demand even higher in future.
“The appetite is going to be huge,” Johnson says. “If I think about our business, we’ve only been around for less than nine years, we’ve only got a small number of products, 41 products. But I still have a lot of data: trading data for every position in my company, prospectuses for all our funds, marketing materials etc… I would like to take all that data, give it to an AI, and have it react. But right now there’s not enough memory on my AI tool.”
Johnson explains that as AI agents become more commonplace, the need for memory will continue to grow. AI agents are designed to be autonomous, to complete functions for people and businesses in the background without prompting. That will require huge amounts of memory to execute effectively. Johnson describes the demand we now see for DRAM as essentially “infinite.”
Markets have already reacted to this massive demand spike, powering huge stock growth for Micron, SK Hynix, and Samsing. Both Johnson and Learmonth, however, argue that there’s still plenty of room for those stocks to keep running. They explain that hardware manufacturing bottlenecks are much harder to overcome and, as we saw with GPUs, demand for hardware can result in long, sustained performance from manufacturers. Learmonth adds that there are no new players looking to enter the place with scale. While some AI hyperscalers had success building their own XPUs as specialized replacements for GPUs, Learmonth explains that memory is more challenging to vertically integrate. Instead, these AI hyperscaler companies are electing to secure long-term deals with the existing memory manufacturers.
Johnson highlights another knock-on effect of this memory squeeze: consumer electronics. Phones, laptops, and even some appliances all use DRAM. Tight supply and high costs might mean that some major device makers may have to change pricing or rollout schedules of their new products. He notes the example of Apple’s Mac Mini, which is essentially out of stock worldwide, in part due to these supply bottlenecks and demand from data centers.
For advisors looking to play the AI memory trend, access can be a bit of an issue. Both Johnson and Learmonth say that broad-based tech exposure may still be the best means for a retail investor. They note that while this memory trend is significant, there are many other similar trends in technology now that investors may want to be exposed to. Direct access is limited, too, as SK Hynix and Samsung are listed in South Korea. Johnson adds that a broad South Korean equity ETF would be too diversified to offer meaningful beta to those two names. They argue that a diversified tech play gives investors a means of access to this theme, and the next one when it comes.
“We’ve continued to see that AI narrative broaden out, even within the technology sector. You’re starting to see other companies perform that weren’t initially seen as prime beneficiaries,” Learmonth says. “Memory was very much in that situation until late summer last year. Then, all of a sudden, storage and memory stocks just were off to the races when, when these shortages started to show up. So I think a diversified technology approach is a great way to play it.”
