DRAM Shortage to Constrain AI Infrastructure Through 2028 as Hyperscalers Hit Memory Wall

A structural shortage in DRAM and high-bandwidth memory threatens to constrain AI infrastructure scaling through 2028, with supply meeting only 60% of projected demand as hyperscalers commit $602 billion in capital expenditure this year. The crisis stems from a zero-sum reallocation of wafer capacity toward AI workloads, where each gigabyte of high-bandwidth memory consumes three times the manufacturing capacity of standard DDR5 chips. SK Hynix has sold out its DRAM, NAND, and HBM capacity through 2026, while global DRAM inventory sits at a historically low 2-3 weeks of supply.

The shortage is not cyclical but structural, according to IDC, which describes it as “a potentially permanent, strategic reallocation of the world’s silicon wafer capacity.” Three manufacturers—Samsung Electronics, SK Hynix, and Micron Technology—control 95% of global DRAM production. SK Hynix alone holds 57% of the high-bandwidth memory market and 32% of overall DRAM, with capacity fully booked into 2026, per Tom’s Hardware. In Q4 2025, SK Hynix overtook Samsung in DRAM revenue for the first time since 1992, capturing 36% market share versus Samsung’s 34%.

The Memory Wall Constrains AI Scaling

Compute performance has been scaling at 3.0x every two years, but DRAM bandwidth has advanced at only 1.6x over the same period, creating what researchers describe as an AI memory wall. This divergence makes memory—not processing power—the dominant bottleneck for training and inference workloads. AI systems now require 2-3x more DRAM per generation than traditional computing applications, while high-bandwidth memory modules sell for $60-$100 per unit versus $5-$10 for comparable DDR5 capacity.

Hyperscaler capital expenditure is approaching $602 billion in 2026, up 36% year-over-year, with OpenAI announcing a $122 billion investment in April 2026, according to BuySellRam. Yet wafer supply trails demand by 20%, according to SK Group’s chairman, who projects the shortage will persist until 2030. AI will consume approximately 20% of total DRAM production by year-end, up from negligible levels two years ago. Nvidia’s Vera Rubin GPU requires nearly triple the memory bandwidth of its predecessor, while OpenAI’s Stargate project has committed 900,000 DRAM wafers per month—equivalent to roughly 15% of SK Hynix’s total capacity.

Capacity Expansion Lags Demand Growth

Memory supply is projected to grow 16% year-over-year for DRAM and 17% for NAND in 2026, while demand growth ranges from 20-35%, creating a widening gap. Samsung plans to increase HBM capacity by 50%, from 170,000 to 250,000 wafers per month by year-end, according to TrendForce. SK Hynix is targeting an 8x increase in DRAM production. Yet these expansions occur within existing facilities—new fabs will not reach volume production until late 2027 or 2028.

SK Hynix’s Yongin M15X facility began early production in February 2026 but won’t hit full capacity until mid-2027. Its West Lafayette, Indiana plant and Micron’s Singapore facility are both slated for late 2027. Samsung’s P5 facility in South Korea will not be operational until 2028. According to IEEE Spectrum, Intel CEO Lip-Bu Tan told attendees at Cisco’s AI Summit in February 2026, “There’s no relief until 2028.”

Geopolitical Concentration and Wafer Tradeoffs

South Korean manufacturers account for over 70% of global DRAM output, creating acute geopolitical vulnerability. The concentration is even more extreme in high-bandwidth memory, where SK Hynix and Samsung together control 91% of the market. Export controls, energy constraints, and regional instability pose direct risks to global AI Infrastructure. According to Bloomsbury Intelligence & Security Institute, the shortage is a national security concern, noting that “three manufacturers, 95% market share” leaves little redundancy.

The wafer allocation problem is zero-sum. Each HBM stack for an Nvidia GPU consumes silicon that could otherwise produce LPDDR5X for smartphones or SSDs for consumer laptops. HBM production is three times more wafer-intensive than DDR5, forcing manufacturers to choose between lucrative enterprise contracts and mass-market components. Samsung acknowledged in a company statement that despite plans to “invest approximately 30% of our sales in facilities in 2026 and accelerate the transition to 1c DRAM, it will be difficult to resolve the supply shortage,” according to Yahoo Finance.

Pricing Pressure and Consumer Impact

DRAM contract prices surged 171% year-over-year through Q3 2025, with Samsung’s memory division posting a 250% year-over-year operating profit increase in Q4 2025, according to Global Semi Research. Q1 2026 contracts saw 90-95% price increases, followed by another 30% hike for Q2 2026. Operating margins at SK Hynix and Samsung now exceed 50%, levels not seen since the memory super-cycle of 2017-2018. Yet unlike that earlier boom, current tightness reflects permanent demand from AI infrastructure rather than cyclical inventory restocking.

Consumer electronics manufacturers face a bifurcated market: hyperscalers secure capacity years in advance at premium prices, while PC and smartphone makers compete for remaining supply. According to IDC analysts, “every wafer allocated to an HBM stack for an Nvidia GPU is a wafer denied to the LPDDR5X module of a mid-range smartphone or the SSD of a consumer laptop.” Smartphone shipments are expected to contract by 3-5% in 2026 due to component shortages, while PC OEMs are extending product lifecycles rather than launching new models.

What to Watch

Samsung’s HBM4 mass production timeline will be critical—delays beyond Q3 2026 would extend the shortage into 2029. SK Hynix’s West Lafayette facility represents the first major U.S.-based memory production in decades; any regulatory or construction delays could further concentrate supply in South Korea. Micron’s HBM roadmap and market share gains will determine whether the duopoly fractures. Watch for hyperscaler contract announcements locking in 2027-2028 capacity, which would signal expectations of prolonged tightness. Energy grid constraints in South Korea could throttle fab expansion even if capital is available—semiconductor manufacturing accounts for 8% of national electricity consumption. Finally, monitor U.S.-China export control updates; any restrictions on advanced memory technology would bifurcate the market and worsen shortages in Western markets.

The supply-demand gap is not closing. Demand growth of 20-35% annually outpaces even the most aggressive capacity expansion plans, which target 16-17% annual increases. Geopolitical concentration in South Korea, wafer reallocation toward high-margin AI components, and multi-year fab construction timelines mean relief is unlikely before 2028. This creates a strategic vulnerability for hyperscalers racing to deploy frontier AI models: you can secure GPUs, power infrastructure, and data center space, but if memory supply remains constrained, the buildout stalls. The industry is entering a prolonged period where memory bandwidth, not compute power, determines the pace of AI scaling—a fundamental inversion of the last two decades of semiconductor economics.