Source: Compiled from Yole

As hyperscale data centers rewrite the rules of the memory market, the memory shortage is likely to persist until 2027.

Fearing potential shortages in the future, customers are scrambling to secure memory supplies for production. Robust demand for server DRAM from AI data centers has driven up memory prices across the entire market.

The DRAM market is in an AI-driven upcycle. Since the third quarter of 2025, hyperscale data centers have been consuming a large portion of the supply, pushing prices higher. Because AI servers require far more DDR5 (and HBM) memory per system than traditional servers, the memory supply for PCs, smartphones, and other end markets is tightening.

Against this backdrop, John Lorenz, Director of Memory and Computing at Yole Group, highlighted a key driver of current price dynamics: concerns over future supply shortages. As DRAM manufacturers prioritize the production of higher-margin HBM and server-grade DDR5, other market segments are taking defensive measures, typically by procuring memory in advance, thereby exacerbating shortages and driving up spot prices.

The latest round of price increases began in the third quarter of 2025, with DRAM prices rising 13.5% quarter-on-quarter. Although the DRAM market is highly volatile, with historical price fluctuations reaching 15% to 20%, this increase comes on the back of a strong rebound from 2023 to the end of 2024 and early 2025. This indicates that the market has reached a cyclical peak and a correction is imminent. However, early signals from corporate earnings reports suggest that prices could rise by an additional 30% in the fourth quarter.

John Lorenz, Director of Memory and Computing at Yole Group, stated: "If you are a company looking to produce new smartphones or personal computers, you need DRAM to populate your devices. If you see that a large amount of production capacity is being diverted to meet server demand, you will worry about the availability of your own supply and try to buy more memory than you actually need… This is a psychological effect. The fear of future supply shortages is driving today's price volatility."

Spot prices for DDR5 memory used in servers have soared by as much as 100% in some cases. PC manufacturers are already feeling the impact: both HP and Dell have warned that they may remove certain laptop models from their product lines next year, either due to high DRAM prices or concerns about being unable to procure sufficient memory.

AI Infrastructure Is Reshaping the DRAM Demand Curve

At the core of this imbalance is the construction of AI infrastructure. Data center operators are procuring AI accelerators on a large scale, along with the general-purpose servers required to run these accelerators. AI accelerators rely on High Bandwidth Memory (HBM), while host servers consume massive amounts of standard DDR5 memory.

An AI server configured with eight accelerators (each equipped with 200GB of HBM memory) contains a total of approximately 1.6TB of HBM memory and about 3TB of DDR5 memory. In contrast, a typical non-AI server built in 2025 uses less than 1TB of DRAM memory in total. This rapid growth in memory capacity per system has outpaced supply capabilities.

HBM is further distorting the market. With prices and profit margins far higher than those of DDR5, manufacturers have a strong incentive to prioritize HBM production. Producing HBM may require up to four times as many wafers per gigabyte as DDR5, meaning that moves to increase HBM output reduce the available capacity for traditional server memory.

These impacts are spreading to other end markets. Automotive applications typically use LPDDR4 and LPDDR5 memory, which are also used in smartphones, tablets, and laptops. However, since the automotive industry remains a strategic priority for memory suppliers—especially as the development of autonomous vehicles requires more memory—they are unlikely to cut ties with this sector. Nevertheless, they do have the leverage to charge automotive customers higher prices to continue securing supplies.

John Lorenz from Yole Group explained: "Since approximately 80% of the manufacturing process for LPDDR is the same as that for DDR, production capacity is somewhat interchangeable. In other words, if a DRAM company has an incentive to produce DDR5 and HBM for its AI server customers, the supply of LPDDR will also be affected."

This dynamic helps explain strategic moves such as Micron's decision to phase out its Crucial consumer business, reflecting the company's greater focus on higher-margin AI-driven demand rather than direct-to-consumer products.

Outside of data centers, smartphones account for approximately 25% of global DRAM bit demand, while personal computers account for around 10-11%. Consumer electronics other than mobile phones and PCs, including gaming devices and wearables, make up another 6%. The automotive industry accounts for about 5%, and industrial, medical, and military applications combined account for approximately 4%.

Data centers dominate the market, accounting for around 50% of total DRAM bit demand. AI workloads alone account for about 30% of total demand (including both HBM and non-HBM), giving them significant influence over pricing.

Demand from hyperscale data centers is increasingly shaping DRAM pricing.

History shows how quickly DRAM market cycles can shift. Between 2014 and 2016, weak demand led to a drop in DRAM prices, prompting Android smartphone manufacturers—particularly those in China—to compete by increasing memory capacity. This new demand absorbed excess supply and drove prices higher, until cost pressures squeezed profit margins and manufacturers stopped increasing memory capacity or shifted to lower-spec models.

This time around, the usual self-regulating mechanism where high prices trigger a demand pullback has not yet materialized. Hyperscale data centers and server manufacturers are far less price-sensitive than consumer electronics manufacturers. They are willing to pay premium prices to secure DRAM supplies and maintain their competitiveness in the AI space, thereby driving up prices for all other market players.

On the supply side, measures to ease supply constraints are structurally limited by long production cycles. Building or expanding a DRAM fab typically takes 2-3 years to reach mass production. Some new supply is expected to come online in 2026, but the volume remains limited.

China's CXMT is expanding production capacity, but it primarily serves domestic customers and has not yet met the requirements of leading global buyers. Samsung is adding equipment at its P4 fab but is prioritizing HBM production over broader DRAM products. SK Hynix's M15X fab is expected to start production in the second half of 2026, with more significant output in 2027, while Micron's new fab in Boise is also projected to ramp up capacity in 2027.

Until then, smartphone and PC manufacturers will need to slow the pace of memory capacity growth, or AI infrastructure spending will need to moderate, to alleviate price pressures before large-scale capacity additions come online.

As AI infrastructure continues to reshape memory demand, DRAM pricing will remain a focal point across the entire electronics ecosystem—far beyond the data center sector. Understanding the interplay between technological transitions, supply allocation, and hyperscale data center procurement strategies is critical for forecasting risks and opportunities in various markets.

Disclaimer: This article was created by the original author. The content represents the author's personal views. Our reprint is solely to convey a different perspective and does not imply our endorsement or support of such views. If you have any objections, please contact the backend.