The great memory squeeze: How AI is rewriting the rules of the chip industry

The world’s three largest memory manufacturers are sending the same message... and it is one the technology industry can no longer ignore.

US memory maker Micron Technology recently announced plans to exit its Crucial consumer business to prioritise supply for the fast-growing AI and data-centre market. Soon after, Samsung Electronics warned that tightening memory supply is likely to impact pricing across the electronics industry. SK Hynix, while rebutting rumours of exiting the consumer segment, has stepped up investments in advanced packaging to support surging demand for high-bandwidth memory (HBM).

Taken together, these signals point to a stark conclusion: the global memory shortage is structural, not cyclical, and it is likely to persist well beyond the near term. Crucially, its impact is no longer confined to AI infrastructure. It is now reshaping costs across enterprise IT and consumer electronics, with pricing pressure expected to extend into 2027–28.

How memory became the choke point

For years, memory, both Dynamic Random Access Memory (DRAM), which supports real-time computing tasks, and NAND, which stores data long-term, was treated as a commoditised component. Prices moved in familiar boom-and-bust cycles, largely driven by PCs and smartphones.

That dynamic has changed dramatically with the rise of AI.

Smartphones use memory primarily to support operating systems and app responsiveness under tight power and thermal limits, prioritising efficiency over raw capacity or bandwidth. Traditional enterprise servers rely on large DRAM pools for databases, virtualisation and caching, with steady throughput requirements but rarely extreme bandwidth demand.

AI servers, however, fundamentally alter the equation.

“Large models, activation data, and constant parameter updates require sustained, ultra-high bandwidth to keep massively parallel accelerators utilised,” says Manish Rawat, associate analyst at  TechInsights. “In AI systems, compute can process data far faster than conventional memory can supply it, making memory, not compute, the limiting factor.”

This gap has exposed traditional memory hierarchies as inadequate and driven the industry toward high-bandwidth memory and advanced architectures optimised for throughput rather than cost-per-bit. As a result, memory has moved from a supporting role to a primary performance bottleneck.

Sanchit Vir Gogia, CEO and chief analyst at Greyhound Research, notes that a single AI training server can carry up to 4 terabytes of memory across CPU DRAM and GPU-based HBM, compared with 16 GB of RAM in even high-end laptops and smartphones.

“This is not a marginal uptick, it is a structural leap,” he says.

The AI-led surge arrived just as memory makers were emerging from one of the deepest downturns in the sector. The oversupply of memory in 2022–23 led to a significant price drop, forcing manufacturers to reduce capital expenditures and cut wafer output. When AI demand accelerated faster than expected, the industry found itself underprepared.

Why supply won’t catch up quickly

Despite rising demand, memory manufacturers remain cautious about scaling production.

Memory is among the most capital-intensive segments of the semiconductor industry, and suppliers are keen to avoid repeating past over-expansion cycles that led to sharp price collapses.

“After repeated boom-and-bust cycles, producers are now prioritising disciplined capital expenditure, expanding only in line with long-term demand,” says Rawat.

At the same time, suppliers are prioritising profitability over volume. AI-focused memory products, particularly HBM and server-grade DRAM, command far higher margins than consumer-grade memory. That has incentivised manufacturers to allocate capacity toward data centres rather than smartphones or PCs.

“Even as demand surges across AI, cloud and enterprise, suppliers are not racing to open the taps,” adds Gogia. “Resources are being funnelled into higher-yield opportunities like HBM and premium DRAM for data centres. That leaves less capacity for general-purpose DRAM and consumer-grade NAND.”

Any new capacity that comes online is often locked into long-term contracts with the largest buyers, further tightening the supply for the broader market. As a result, industry analysts increasingly expect memory constraints to persist until at least 2027 or 2028.

Who bears the cost

Over the past two years, memory pricing has surged sharply, turning what was once a modest cost component into a major pressure point across devices and infrastructure.

In smartphones, memory historically accounted for around 10–15 percent of the bill of materials. By 2025, that share has climbed beyond 20 percent, and is even higher in lower-end models. For budget smartphones, memory has now emerged as the most expensive component after the system-on-chip.

DRAM prices have spiked by nearly 70 percent, while NAND prices have more than doubled over the past 18 months, according to Greyhound Research. This rise in component costs is expected to weigh on demand, with global smartphone shipments forecast to decline 2.1 percent in 2026, according to Counterpoint Research.

In laptops and notebooks, memory and storage are also becoming a significant cost burden. DRAM and SSDs can now account for up to a quarter of total system cost. Most machines ship with 16 GB of DDR5 memory and a 512 GB SSD, both of which have seen sharp price increases.

“As a result, enterprises refreshing fleets are seeing that cost pressure in real time, either pay more or dial down specs,” says Gogia.

The shift is even more pronounced in servers, where memory is no longer just an expensive component but increasingly the primary driver of system cost.

“A standard dual-socket server with 512 GB of DDR5 can now see memory represent 30–40 percent of the BOM (bill of materials),” Gogia says. “In AI workloads, GPU accelerators such as NVIDIA’s H100, which ship with 80 GB of HBM each, can see memory account for more than half the total cost of the accelerator once packaging and board-level design are factored in.”

For now, AI infrastructure is bearing the heaviest memory cost burden, followed by enterprise IT. Consumer markets have been relatively insulated through product tiering and OEMs absorbing costs in the short term. But as shortages persist, pricing pressure is expected to cascade steadily from memory suppliers to system manufacturers and ultimately to end users.