During a recent virtual roundtable of TechInsights analysts, the consensus was that memory makers have learned from previous boom-and-bust cycles and are showing greater discipline in ramping up to meet AI-driven demand. Forecasted HBM and DRAM shortages stem not from supply chain disruptions, but from unprecedented and largely unanticipated adoption.

AI growth will continue to be driven primarily by the cloud, with the focus gradually shifting from training to inference.

Cameron McKnight-MacNeil, process analyst at TechInsights, said Nvidia’s Rubin platform, announced in September 2025 and touted as a new class of GPU, appears to be positioned as a platform for inference in the cloud.

He said the massive scale of HBM required for AI platforms will continue in 2026, noting that each of Nvidia’s Blackwell accelerator packages contains eight HBM modules, with each module comprising eight DRAM dies plus a controller die. With hyperscalers deploying super pods of racks full of GPUs, “HBM4 is going to be the memory flavor as it were for AI in ’26.”

Cameron McKnight-MacNeil Cameron McKnight-MacNeil (Source: TechInsights) McKnight-MacNeil said the hunger for HBM and the corresponding packaging raises yield concerns. “TechInsights has highlighted the potential concerns around the poor yields from stacking [DRAM dies] and the impact of sustainability,” he said.

JEDEC’s recent update increased the allowable package height for HBM, enabling stack heights up to 16 dies, made possible in part by thinning DRAM dies and hybrid bonding.

There is extensive R&D aimed at improving system efficiency, alongside increased adoption of optics in data centers, McKnight-MacNeil added. “They’re going to make the transistors more efficient.”

But any improvements are immediately snapped up by the request for more compute, he said. “It’s a growth situation.”

Dan Kim, chief strategy officer at TechInsights, said that despite discussions about a potential AI bubble, the current memory market is not speculative. “This is purely a function of demand and supply.”

He said the market began recovering in 2023 after a cyclical downturn, with growth expected to remain healthy through 2026. “What we’re seeing is just a robust demand for accelerated compute that is being manifested in the growth of GPUs and ASICs.”

This “extraordinary” growth is happening primarily in the data center and the cloud, Kim added, contributing to rising prices for both DRAM and NAND.

“There is also an insatiable demand to lower the power consumption of AI compute,” he added.

That means innovation in more efficient power systems is becoming just as critical as scaling GPU and memory bandwidth.

Growth outside the data center space is not as aggressive, Kim said, but all segments are competing for the same foundry space. “This will shape up to be a very interesting growth market for 2026.”

But if CES is any indication, edge AI is seeing remarkable and even transformative growth, Jack Narcotta, head of consumer electronics at TechInsights, said. “In the consumer market, there’s always that phrase ‘bigger, better, faster, more’. A handful of things were just outright vaporware.”

He said the smart home industry is rapidly approaching a point where brands must consider some fundamental questions about their relationship with AI and where it will be located.

Putting AI directly on the device raises power and thermal issues, Narcotta said, but this would provide a smoother, faster, and significantly more reliable experience by reducing dependence on cloud communication.

He said the consumer segment for AI is the “wild west” right now and that the Nvidia Rubin announcement seemed of place at CES given the platform is for the data center. But it will have a ripple effect, as consumer brands will need to understand what its capabilities mean for their products, Narcotta said. “Consumer electronics tends to lag some of the macro trends that are happening.”

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