Why AI Data Centers Are Driving a Global Memory-Chip Shortage

AI data-center construction is turning memory into a system-wide constraint. The same build-out that increases demand for AI accelerators also requires high-bandwidth memory (HBM), conventional server DRAM and enterprise SSDs. With suppliers directing scarce capacity toward those products, PC and smartphone makers face higher costs and less flexible supply. The result is a potentially prolonged memory cycle—not a short-lived component shortage. 3334

The bottleneck is broader than HBM

HBM is the specialized stacked memory used close to AI accelerators, but it is only one part of an AI server’s memory bill. Large systems also require conventional DRAM for CPUs and system memory, while data-center storage demand supports enterprise SSDs. TrendForce projects combined server DRAM and HBM bit demand to grow at a 37.4% compound annual rate through 2028. 33

This matters because HBM requires scarce leading-edge DRAM capacity and advanced packaging. When a manufacturer prioritizes HBM and high-capacity server memory, fewer resources are available for the commodity DRAM used in PCs and phones. TrendForce says AI servers are driving demand for both conventional DRAM and HBM, while capacity prioritization is adding cost pressure to PC and smartphone markets. 3439

Why higher prices may last through 2027

The supply response is slow. Building capacity is not simply a matter of announcing a fab: clean rooms, equipment installation, process qualification, yield improvement and advanced packaging all have to ramp before meaningful output reaches customers.

A TechInsights forecast reported in September projected DRAM prices could rise by more than 200% year on year—an unusually aggressive forecast, not a confirmed market result. Separately, reporting on TrendForce estimates put 2026 server-DRAM component-price growth at roughly 270% year on year and enterprise SSD price growth at 235%. 4935

The market is also being tightened by forward purchasing. DigiTimes-based reports say Samsung, SK hynix and Micron have allocated their 2027 DRAM and HBM capacity early. That claim should be read cautiously: it is industry reporting rather than a formal assertion by the manufacturers that every bit of output is unavailable. But the underlying tightness is supported by other signals. Reuters reported that SK hynix CEO Kwak Noh-jung expects 2027 to be the industry’s worst-ever year from a supply perspective. 31

What it means for PC and smartphone buyers

Memory inflation does not automatically translate into the same price increase on every device. Manufacturers can absorb some cost, change configurations, shift production toward higher-margin models or raise retail prices. Still, the direction of pressure is clear: TrendForce expects PC and smartphone demand to slow as AI servers receive supply priority and component costs rise. 34

The most likely effects are:

• Higher average selling prices. Brands with pricing power can pass part of the higher memory bill to buyers.
• Fewer low-cost configurations. Entry products may carry less memory or storage, or disappear where their margins no longer work.
• Tighter supply for lower-margin devices. PC makers, Chromebook vendors and midrange phone brands have less room to absorb higher component costs than premium vendors.
• Potential shipment pressure. A vendor may reduce output of unprofitable models even if consumer demand itself remains intact.

There is no single reliable forecast for global PC or smartphone shipments attributable to memory alone. Replacement cycles, exchange rates, tariffs and broader economic demand will all influence the outcome. The stronger conclusion is that expensive memory makes affordable-device strategies harder to sustain.

DRAM and NAND are likely to diverge

Treating all memory as one market obscures an important distinction. DRAM—especially HBM and server DRAM—appears set to remain structurally tight, while NAND has a clearer route to balance as bit output expands.

Segment	2027 outlook	What could happen next
DRAM and HBM	AI demand, continued HBM capacity allocation and increased CPU-memory procurement are expected to keep supply constrained and pricing under upward pressure. 39	New capacity and better yields may eventually relieve the imbalance, but the timing and scale of price normalization remain uncertain.
NAND and enterprise SSDs	AI-related storage demand has kept NAND undersupplied in 2026. 42	TrendForce expects process migrations, bit-output growth and weak consumer-electronics demand to ease NAND tightness in the second half of 2027, potentially putting downward pressure on prices. 4239

That divergence is central to the 2027–28 outlook. HBM conversion and AI-server DRAM demand constrain the DRAM market in ways that do not map neatly onto NAND. NAND can still be volatile, but available evidence points to a looser market sooner than in DRAM rather than a guaranteed continuation of HBM-style scarcity. 3942

China’s CXMT and YMTC: more supply, but not an immediate global release valve

China’s memory expansion could improve domestic supply resilience, but it is not yet a straightforward substitute for leading-edge HBM and server DRAM from the established leaders.

CXMT, China’s leading DRAM challenger, has reportedly started small-scale HBM3E production for domestic AI-chip developers. Commercial integration is contingent on successful trials, with a targeted capacity expansion in 2027. 18

YMTC is China’s major NAND producer. Together, CXMT and YMTC can add strategically important DRAM and NAND capacity for domestic devices, SSDs and data-center deployments. But export restrictions on advanced HBM and chipmaking tools complicate China’s ability to scale frontier memory technology and the surrounding packaging ecosystem. The United States added advanced HBM to export restrictions in December 2024. 18

China’s build-out may not quickly ease shortages elsewhere. Reporting indicates CXMT’s additional output is likely to be absorbed by Chinese cloud-service providers, while restrictions make direct supply of advanced memory from Korean manufacturers to Chinese cloud providers more difficult. 17

The practical takeaway

The AI memory crunch is not simply a story of sold-out HBM. AI clusters raise demand for several memory categories at once, while the specialized production and packaging behind HBM constrain how quickly suppliers can expand. That combination gives memory manufacturers stronger leverage in contract negotiations and shifts cost pressure toward consumer electronics. 4434

The clearest base case from the available evidence is sustained DRAM tightness through 2027, with the biggest effects concentrated in HBM and server DRAM. NAND may remain pressured in the near term but has a more plausible path toward easing by late 2027. For device buyers, the impact is likely to appear less as a universal price shock than as pricier mainstream products, fewer bargain configurations and tighter availability in lower-margin segments. 3942