AI Data Centers Are Squeezing the Memory Market—And Raising the Stakes for PCs and Phones

AI infrastructure is creating a memory bottleneck because the same industry must supply both premium AI memory and the DRAM and flash used in everyday electronics. Memory makers can earn more from high-bandwidth memory (HBM), high-capacity server DRAM and enterprise storage, so capacity is being directed toward data centers rather than lower-margin PC and handset components. The result is a supply squeeze that reaches far beyond AI accelerators. 4152

The core issue: finite capacity is being redirected

HBM is a premium form of DRAM used alongside AI accelerators. Its production requires DRAM capacity as well as sophisticated stacking, testing and packaging. At the same time, AI servers require substantial conventional server memory and enterprise NAND flash storage.

That combination matters because memory capacity is not instantly interchangeable or expandable. As Samsung, SK hynix and Micron prioritize AI-oriented products, the supply available for conventional DRAM tightens. S&P Global describes this shift toward HBM as a direct driver of conventional-DRAM shortages and higher prices. 41

The market is also concentrated. Samsung, SK hynix and Micron account for roughly 90% to 95% of global DRAM production capacity, so a change in their product mix has global effects. 47

Are DRAM prices really up more than 200%?

The headline needs context. TechInsights has forecast that DRAM prices could rise more than 200% year over year, but that is a forecast—not a confirmed, market-wide price print. 36

Still, published market data points to extraordinary price pressure. TrendForce forecast conventional-DRAM contract prices would rise another 13% to 18% quarter over quarter in the third quarter of 2026, even after earlier sharp increases, while noting that the market remained extremely tight. 51 Separately, Korean export-unit-price data reported a 401% year-over-year increase for DRAM exports in part of August; that is a trade measure, not a universal retail or contract-price benchmark. 33

The practical takeaway is clear: price movements differ by product, contract, geography and timing, but the direction of travel is sharply upward.

Why “sold out through 2027” should be read carefully

Reports of sold-out memory should not be interpreted as every type of RAM or flash chip being unavailable until 2027. The stronger evidence is that leading-edge HBM and high-density server memory are under exceptionally tight allocation, and that customer demand is being booked far ahead.

For example, Micron’s entire 2026 HBM output was reported as sold out, while reports have also described 2027 DRAM and HBM capacity at the three major suppliers as heavily allocated. 3440 This is best understood as an allocation market: buyers with long-term commitments and strategic importance gain access first, while other customers face higher prices, less flexibility or alternative specifications.

What the shortage means for PC and smartphone buyers

Device makers have limited options when memory costs jump:

• Absorb higher component costs, which compresses margins.
• Increase retail prices, particularly for models with larger RAM and storage configurations.
• Reduce specifications or promotions, including lower memory configurations, fewer discounts or revised product mixes.
• Adjust production plans if preferred components are unavailable or uneconomic.

IDC expects the pressure to affect both pricing and volumes. Its current forecast calls for worldwide PC shipments to decline 11.3% in 2026 while revenue rises 1.6% because of higher average selling prices. It forecasts smartphone shipments to decline 12.9% in 2026, with memory supply problems persisting through 2026 and likely well into 2027. These are forecasts, not settled outcomes, and consumer demand remains an important constraint on how much cost OEMs can pass on. 56

The impact will not be uniform. Memory-heavy gaming PCs, workstations, premium phones and devices with large SSD capacities are generally more exposed to component inflation than entry-level products, though manufacturers can change configurations to manage the hit.

Why new fabs will not fix the problem overnight

Building capacity is a multi-stage process: construction, tool installation, process qualification, yield improvement and volume ramp all take time. HBM adds another layer of difficulty because advanced packaging and stacking must scale alongside DRAM wafer output.

Some additions are coming sooner. SK hynix’s M15X facility is slated for utilization by mid-2027, while Samsung’s P5 facility is expected to be operational in 2028. 21 But other major SK hynix projects have first-cleanroom schedules in late 2028 and 2029, underscoring why announced investment does not equal immediate sellable supply. 17

This is why 2027 may bring incremental improvement without producing an immediate glut. TrendForce expects several suppliers to bring capacity online in 2027, but says construction, equipment installation and materials preparation will delay meaningful production ramp-up. 52

Will the market ease by 2028?

Relief by 2028 is plausible, but not guaranteed—and the outlook differs by memory type.

TrendForce expects NAND flash to move into a looser supply environment in the second half of 2027 as new capacity arrives and consumer-electronics demand stays weak. For DRAM, it expects continuing HBM allocation, AI-server demand and procurement of server memory to keep supply constrained and prices rising. 52

Samsung has also warned that it does not expect a significant increase in incremental supply through 2028. 53 The variables to watch are AI-server buildout, HBM yields, advanced-packaging capacity, the pace of new-fab ramps and whether producers shift more capacity back toward conventional DRAM.

China can help some segments, but not the near-term HBM bottleneck

China’s two leading memory names have different roles:

• YMTC is primarily a NAND flash producer. More YMTC output could eventually matter for client SSDs and other flash products, but it is not a direct substitute for the advanced DRAM and HBM most constrained by AI demand.
• CXMT is the more relevant Chinese supplier for DRAM. It has reportedly begun small-scale HBM3E production for domestic AI-chip developers, but its capability at large-scale production remains unproven. 4

Export controls add uncertainty. Reuters reports that YMTC is on the U.S. Entity List, restricting its access to U.S.-origin suppliers, software and production tools; further restrictions affecting both YMTC and CXMT have been under discussion. 1 Those limits can make capacity expansion and access to leading-edge manufacturing capabilities less predictable.

China may therefore become increasingly important in conventional DRAM and NAND, especially for domestic demand, but available evidence does not support treating it as a near-term solution to the frontier-HBM shortage.

Bottom line

The AI memory crunch is not simply a surge in demand for one premium chip. It is a supply-allocation problem across wafers, advanced packaging and long-term customer commitments. That is why a rush for HBM can raise costs for ordinary DRAM and flash used in PCs, smartphones and enterprise systems.

The most defensible outlook is continued tightness through 2027, with a possible but uneven easing beginning as new capacity ramps. NAND may loosen first; DRAM and HBM are more likely to remain constrained as long as AI data-center demand continues to outpace qualified supply. 5253