Why AI Is Driving a Global DRAM Shortage—and What It Means for PCs, Phones, and Data Centers

The current memory crunch is best understood as a capacity-allocation shock, not simply a sudden increase in demand. AI servers require large quantities of high-bandwidth memory (HBM), which is built from stacked DRAM dies and depends on advanced packaging. As memory manufacturers prioritize this higher-value segment, less capacity is available for conventional memory used in servers, switches, PCs, and smartphones.

The price signal is striking. South Korea’s DRAM export unit value reached $92,183 per kilogram during the first 20 days of August, up 401% from a year earlier and roughly 12.5 times its January 2023 trough. 11 That number should be read carefully: it is an aggregate customs unit value calculated from export value and weight, not a quoted price for a standardized DRAM chip. 13

Why HBM is squeezing ordinary DRAM

HBM combines multiple DRAM dies in a vertically stacked package and is designed to provide the high bandwidth required by AI accelerators. That makes it more dependent on scarce wafer, packaging, and qualification capacity than a conventional memory product.

The commercial incentive is equally important. HBM and high-capacity server memory serve the fastest-growing and highest-value part of the market, so manufacturers have strong reasons to shift production and investment toward them. The resulting trade-off is tighter availability of ordinary DDR4, DDR5, LPDDR, and server DRAM—even when demand for those products has not disappeared.

Goldman Sachs expects the overall DRAM supply shortfall to widen from 5.0% in 2026 to 5.9% in 2027, citing HBM for AI servers and high-capacity server DRAM as consumers of limited production capacity. 4 That is a forecast, not a guaranteed outcome, but it illustrates why the shortage is being treated as structural rather than a short-lived spot-market disruption.

Data centers face a two-part bottleneck

AI data centers need both accelerators and memory capable of feeding them. Scarce HBM can therefore constrain server deployment even when an operator has secured the compute processors themselves.

Higher prices for HBM and conventional server DRAM also increase the cost of each server. The largest cloud providers are better positioned to secure supply through scale and long-term purchasing relationships, while smaller operators and enterprise buyers may face delayed deployments, higher system prices, or pressure to prioritize only the most valuable workloads.

This does not mean every data-center project will stop. It means memory becomes a more important limit on deployment speed and system economics, alongside accelerators, power, networking, and advanced packaging.

Enterprise switches and WLAN equipment are unusually exposed

The shortage is also reaching enterprise networking equipment. Campus switches and wireless-LAN products use DDR memory, so their manufacturers are exposed to the same component-price escalation affecting servers and PCs.

Dell’Oro reports that the spot price of DDR 8Gb memory chips increased 788% over the preceding year. Vendors typically purchase memory through contracts rather than at spot prices, but the spot market is an indicator of the pressure feeding into those negotiations. Dell’Oro estimates that memory, previously a low- to mid-single-digit share of the bill of materials for campus switches and WLAN equipment, could rise to roughly 30% in affected products. 54

The likely business consequences are straightforward: higher equipment prices, reduced memory configurations, postponed refresh cycles, and greater reliance on existing designs. Dell’Oro says meaningful relief for this market may not arrive before 2028. 54 That timing remains an outlook rather than a fixed deadline, because supply, demand, and product mix can change.

PCs and smartphones are likely to become more expensive

Consumer devices have fewer ways to absorb a major memory-cost shock. Manufacturers can raise prices, reduce base memory or storage, trim other components, or accept lower margins. Each option creates a trade-off for buyers or vendors.

Gartner forecasts that 2026 PC shipments will decline 10.4% and smartphone shipments 8.4% from 2025 levels. The firm also estimates that combined DRAM and SSD prices could rise 130% by the end of 2026, pushing average PC prices up 17% and smartphone prices up 13%. 18 21

Lower-priced devices are particularly vulnerable because memory represents a larger share of their bill of materials and buyers are more sensitive to price increases. The market response may therefore include fewer entry-level configurations, higher prices for comparable specifications, and longer replacement cycles as households and businesses keep existing devices in service.

The semiconductor market is splitting by memory type

The shortage creates very different outcomes within the semiconductor industry. Companies with strong HBM capabilities, advanced packaging, and established AI-accelerator relationships are positioned to benefit from the shift toward premium memory. Suppliers focused on PCs, phones, networking equipment, and commodity electronics face a more difficult combination of higher input costs, constrained supply, and weaker unit demand.

That changes which capabilities matter strategically. Competition is no longer defined only by leading-edge logic fabrication. DRAM process technology, HBM stacking, packaging capacity, yields, and customer qualification are all increasingly important parts of the AI hardware supply chain.

The result is a more segmented market: premium AI memory can remain highly profitable while conventional memory becomes scarce and expensive because manufacturers are reallocating capacity away from it.

Processing-in-memory is promising—but not a quick fix

One potential response is processing-in-memory (PIM), which places limited computing logic near or alongside the DRAM banks. By performing some calculations where the data is stored, PIM can reduce the movement of data between memory and an AI processor.

At Hot Chips 2026, Samsung presented LPDDR5X-PIM and reported a test using Meta’s Llama 3.1 8B model on an edge-AI accelerator. The system produced 81.3 tokens per second, compared with 27 tokens per second using conventional LPDDR5X in the reported comparison. 34

That result is significant for the specific test, but it does not establish that PIM can replace HBM across the AI market. PIM is most relevant to selected inference and edge-AI workloads; large-scale training and other applications that require very high memory bandwidth still depend on different system architectures. Nor does a successful demonstration prove broad commercial deployment in PCs or smartphones.

What could end the shortage?

Forecasts extending into 2027 or 2028 should be treated as scenarios rather than certainties. Several developments could reduce the pressure:

• New DRAM and advanced-packaging capacity could arrive faster than expected.
• AI infrastructure spending could slow, reducing HBM demand.
• Device demand could weaken enough to free conventional-memory supply.
• PIM and other architectural changes could reduce memory traffic for specific workloads.
• Buyers and manufacturers could redesign products around different memory densities or configurations.

The opposite is also possible. Stronger AI accelerator demand, slower packaging expansion, or manufacturing and qualification problems could prolong the squeeze.

The central takeaway is that AI is reshaping the memory market from the inside out. HBM demand is not merely adding another product category; it is changing how scarce DRAM capacity is allocated. That is why a shortage originating in AI servers can raise the cost of a campus switch, extend a PC replacement cycle, and make a smartphone bill of materials harder to manage at the same time.