It is a seller’s market for DRAM going into 2026 with AI driving up demand, even as the performance gap between processor and memory is now a critical challenge.

NAND flash for AI storage will ride the coattails of DRAM demand, while emerging memories like MRAM, ReRAM and FRAM will continue to pursue more niche, lower volume opportunities and grapple with economies of scale.

TechInsights’ “Memory Outlook Report 2026” sees AI reaching a crossroads—traditional memory technologies are no longer sufficient to meet power and performance demands, which means memory technology is becoming the bottleneck of the data economy. The coming year will be a critical one for memory scaling to meet the demands of AI head on.

In a briefing with EE Times, Mike Howard, director of DRAM and memory markets at TechInsights, said the performance gap between processor and memory will drive the adoption of high-bandwidth memory (HBM) and the CXL protocol to address the “memory wall.”

But this wall is not the only challenge in the year ahead, Howard said. The “drunken spending” by hyperscalers that began in the tail end of 2023 to build the capacity for 2025 has left an industry short on wafer capacity, with production shifting from DDR and LPDDR to HBM to meet AI demand, he said. “The margins on HBM are just so delicious that these guys are falling all over themselves to add HBM capacity.”

It is also not clear where the ceiling is for AI demand, Howard added. “We just can’t keep up from a capacity perspective.”

Consumer devices will take brunt of DRAM undersupply

AI data centers do not just need HBM, Howard noted. Many agentic AI workloads only need traditional servers, which are also seeing strong demand. “There’s not enough DRAM to go around.”

TechInsights is forecasting 40,000 wafers out of SK Hynix in 2026, most of which will be allocated to HBM, Howard said. Samsung has some capacity in Pyeongtaek, but Micron Technology’s second Idaho fab will not be online until 2027, he noted.

This shortage makes for higher prices and better profit margins for DRAM makers due to significant undersupply and robust demand, Howard said.

This boom is not unprecedented—the DRAM and flash markets have historically followed a boom-bust pattern. This latest uptick in demand for memory follows a post-pandemic lull, but even the lows are higher than they used to be, Howard said. “We’re not shedding companies anymore. We’re done consolidating.”

Whatever supply is available, either for HBM or DDR5, is going to get allocated to big tech companies who have the cash, Howard said, which means the DRAM undersupply will lead to shortages for consumer devices, such as smartphones.

TechInsight’s tight memory forecast is echoed by Counterpoint Research’s “Memory Solutions for GenAI” bi-weekly report released in late November. It predicts memory prices are likely to rise 50% from current levels through the second quarter of 2026 due to critical chip shortages.

Howard said the forecasts for high DRAM demand need to be tempered with the uncertainty that comes with AI—its rapid emergence in 2023 could be the beginning of a bubble that could pop in the next year. “For every story you read about demand going gangbusters, there’s a corresponding story about the AI bubble.”

Howard added that DRAM makers have not forgotten the 2018 downturn, and many went through the 2008 downturn. “They’ve seen the cyclicality. They are realists.”

This means they are going to increase capex by 20% next year, but with contingency plans in case they must slam on the brakes. “They’re sipping the Kool-Aid maybe but not wholeheartedly chugging it.”

Meanwhile, the focus on DRAM means NAND will get a little less attention, Howard said. “NAND obviously is going to feel a tailwind from this, but that’s further exacerbated by the shortage in HDDs.”

He said hard drive makers are looking at an 18-month window to ramp up capacity, but they do not want to find themselves in a position where they are oversupplied. “The HDD shortage is driving more demand to SSDs.”

NAND underinvestment puts pressure on AI storage

Since AI’s inception, throughput has become more important, Howard said, which is a boon for SSDs. But NAND has seen a dearth of investment in the last five years—profitability has been poor and there is a reluctance to invest.

SSD maker Phison USA is keeping a watchful eye on NAND supply, Sebastien Jean, the company’s CTO, told EE Times. “The NAND industry is cyclical, and we’ve come through a period where production cuts and AI-driven demand have tightened supply.”

He said Phison’s view is a bit different from a pure NAND manufacturer because the company sits in the middle of the ecosystem. “We work closely with all the major NAND vendors and with a broad base of customers across servers, automotive, industrial and gaming, so we have good visibility into both sides of the supply–demand equation.”

From an operational standpoint, Phison has been planning for tighter supply, Jean added. “We qualify our controllers and SSDs across multiple NAND generations and suppliers. We mitigate NAND shortages through long-term partnerships and close coordination with our suppliers and customers.”

He said Phison also has the option to enable multiple NAND sources in parallel. “While the industry will need continued investment to keep pace with AI and high-density SSD roadmaps, we’re confident in our ability to support our enterprise and AI customers through upcoming cycles.”

Jean explained that AI has turned storage from a background service into a front-line performance component that is critical to how AI pipelines work. They need SSDs that can sustain high bandwidth and IOPS while keeping tail latencies under control. “That’s a big reason we see enterprises shifting critical AI datasets off disk and onto flash,” he said.

Phison is seeing a need for new classes of SSD specifically optimized for AI in ways that go beyond basic speed and latency. “Model sizes and context windows keep growing.”

That means SSDs must meet two hard requirements simultaneously: more memory close to the GPU and more capacity in the storage tier. “We are seeing more coordination between GPU and SSD to greatly improve the user experience without adding the risk of putting storage directly on the GPU,” Jean said.

Edge and automotive are also driving memory demand

Beyond hyperscale and enterprise data centers, Jean said Phison sees strong SSD demand in edge and embedded systems, as well as client devices and vertical markets like automotive and government.

“On the edge, robotics, industrial IoT and telco nodes are starting to run local AI inference, which makes low-latency, rugged flash storage essential,” he said. “Automotive platforms are adopting more flash for ADAS, in-vehicle infotainment and data logging, and we’re also seeing secure, on-premises AI deployments in government and defense.”

Jean added that the common thread across all these segments is that once AI becomes part of the workflow, spinning disks cannot keep up with the latency, power and durability requirements, so flash becomes the default.

TechInsights’ Howard affirms these trends—there are edge applications with trillions of devices that require more memory and storage to run AI. The automotive segment is also a driver of demand, albeit more slowly, he said. “Their semiconductor content steadily increases as we go through time.”

Automotive memory buyers are seeing price increases, and Howard said smaller customers for DRAM will have to fight for allocation. He said the DRAM undersupply, coupled with shifts in the hardware stack, might create opportunities for emerging memories. “We’ve gone from the CPU-dominated to more of a GPU-dominated compute paradigm.”

And with “crazy” price run-ups for DRAM, emerging memory that looked uneconomical six months ago starts look better. “If you couple that with some sort of performance augmentation or differentiation, there’s an opening there.”

Emerging memories opportunities are not dependent on AI

In a recent webinar jointly hosted by Objective Analysis’ principal analyst Jim Handy and Coughlin Associates’ president Tom Coughlin, the theme was that the adoption of emerging memories, such as MRAM, PCM, ReRAM and FRAM, are not being driven by AI.

All these memories have common attributes, Handy said. Among them are a single transistor cell and their higher speeds relative to flash memories. They are all non-volatile, unlike DRAM and SRAM. “That’s a big plus.”

Intel’s 3D Xpoint PCM technology, Optane, was an attempt to move that persistence closer to the process, which has its advantages, Handy said, but getting the infrastructure in place to support it was taking a long time.

All the emerging memories have write-in-place capabilities that are not available in SRAM, NAND and NOR flash, which means they have better scaling limits, Handy said. They also offer significantly better endurance than NAND flash. “The wear mechanisms are nowhere near as difficult as dealing with flash.”

Emerging memories can be both embedded and discrete. Coughlin said there is a lot of embedded activity around MRAM, especially as NOR flash is being deprecated in embedded applications because it does not scale below 28 nm. “It also has endurance issues if you’re updating code storage,” he said.

There are several types of MRAM in production by multiple companies, including Everspin, TSMC, Samsung and GlobalFoundries. “There’s broad support. It’s in the leading position right now in terms of actual applications.”

Coughlin said the other major contender is ReRAM, of which there are several types for discrete and embedded applications and has the potential to rise above other emerging memories in the long run. “It’s using materials that are already common in the manufacturing, and it can be radiation tolerant as well,” he said. “There are many different players out there and many approaches to it.”

But arguably the emerging memory that gets the most use is FRAM, which has a high production volume in some forms, but low wafer run rates, Coughlin said.

FRAM is already used in financial smartcard applications, transit payment and set-top boxes. Compared to existing EEPROM technologies, FRAM is more resistant to data corruption via electric fields and radiation.

Coughlin said the appeal of FRAM is its low energy, high speed and high endurance. The drawbacks of this memory are that it is not compatible with traditional processes and some of the materials are not friendly in a semiconductor fab environments.

The fourth emerging memory that has potential is PCM, which Coughlin said has encountered challenges due it’s high energy and thermal sensitivity. “The materials are somewhat challenging.”

The most notable form of PCM is 3D Xpoint, co-developed by Micron and Intel, which the latter brought to market as Optane. Coughlin said there is noise about companies developing another PCM-based technology, but there is no major commercialization efforts right now.

Emerging memories still struggle with economies of scale

Handy said market factors affect how well emerging memories are adopted compared with entrenched memories which are big standards, such as DRAM and NAND flash.

Whether the memory is embedded or discrete is s a major factor, and they are very different markets; Handy said the opportunities overall for emerging memories are for niche applications, which is where both discreet and standalone memory chips are selling right now.

Military and aerospace are two markets where emerging memories will find demand, due in part to their radiation tolerance, Handy said, whether it is for weapon systems or satellites. Earth orbit has a lot of radiation. “That market may have low unit volumes, but they do pay extremely high prices,” he said. “That’s been something of a saving grace for a lot of these companies.”

The industrial market is also adopting emerging memories, as is the medical wearables segment because the low power consumption allows for longer battery life in devices in hearing aids and cardiac pacemakers. Handy said endurance and reliability across wide temperature ranges are also appealing characteristics of emerging memories, specifically in automotive.

As NOR flash and SRAM are hitting scaling limits, it should present a “windfall” for emerging memories, he said.

Handy and Coughlin’s overall forecast for emerging memories is that embedded forms will drive early volumes and that wafer volume continues to drive economies of scale that will open new markets for stand-alone chips.

But achieving economies of scale remains the biggest factor affecting the adoption of emerging memories, Handy said. “You might get a smaller die size with these new technologies, but that doesn’t guarantee that the part’s going to be cheaper.”

Wafer volume greatly affects the cost of production of an emerging memory and low price always wins out over better features. “You’ve got a real chicken and egg problem. Until you can get your volumes up, you can’t really compete against established technologies,” Handy said.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved