An unprecedented supercycle is taking hold of the global memory market, fueled almost entirely by the AI industry’s insatiable demand for High Bandwidth Memory (HBM). The market, valued at $2.93 billion in 2024, is on a trajectory to hit $16.72 billion by 2033, reflecting a 21.35% compound annual growth rate. Goldman Sachs paints an even more bullish picture, projecting the HBM market could reach $116 billion by 2027 and a staggering $168 billion by 2028—a massive upward revision. This isn’t just growth; it’s a fundamental reshaping of semiconductor economics and the entire technology landscape.
For memory manufacturers, the situation is extraordinary. The entire 2026 production capacity for all three major HBM suppliers—SK Hynix, Samsung, and Micron—is already sold out. Demand is projected to outstrip supply through at least 2027, with SK Group Chairman Chey Tae-won bluntly warning that wafer shortages could persist until 2030. This isn’t a cyclical fluctuation; it’s a structural crisis. The root cause lies in HBM’s manufacturing complexity, which requires three to four times the production capacity of conventional memory and still suffers from low yields, particularly for the highest-performance chips. Advanced packaging methods like Mass Reflow Molded Underfill (MR-MUF) and Thermal Compression Non-Conductive Film (TC-NCF) further stretch manufacturing timelines, compounding the supply bottleneck.
Driving this supply crunch is the AI industry’s voracious appetite for HBM. By 2026, AI and machine learning workloads will consume over 55% of total HBM supply. Each successive generation of AI accelerators, from NVIDIA’s H100 to the forthcoming Blackwell and Rubin platforms, demands a quantum leap in HBM bandwidth and capacity that pushes manufacturing to its limits. To put it in perspective, a single NVIDIA NVL72 rack consumes the memory equivalent of a thousand high-end smartphones. The knock-on effect is severe: memory makers are reallocating significant wafer capacity away from traditional DRAM, leading to tighter supply and surging prices for PCs, smartphones, and servers.
Pricing power has swung decisively to HBM vendors. These chips command a five to six-fold premium over equivalent DDR5 capacity, thanks to their intricate 3D stacking and Through-Silicon Via (TSV) processes. Prices have already jumped 20-40% year-over-year, with another 44% increase expected in 2027. The shift has also sent conventional DRAM prices soaring by 80-90% in a single quarter for certain products, with some common types rising 50% quarter-over-quarter. In a clear sign of desperation, customers are reportedly paying double or triple the standard price to secure supply. Consequently, major hyperscalers and AI firms are now locking in long-term agreements with prepayments, volume commitments, and penalty clauses, giving HBM producers unprecedented revenue visibility and stability.
Among the “Big Three” HBM manufacturers—SK Hynix, Samsung Electronics, and Micron Technology—the competitive landscape is anything but static. SK Hynix maintains a commanding lead with roughly 50-55% market share, a direct result of its early HBM3E mass production and its tight relationship with NVIDIA. However, Micron, once a latecomer, is rapidly closing the gap, with its market share hitting 21% in Q2 2025, temporarily eclipsing Samsung. Samsung, the overall DRAM leader, has stumbled with HBM3E qualification delays but is now aggressively ramping up its 12-high HBM3E production. All three are now in a sprint to develop HBM4 and HBM4E, expected in 2025-2026 and 2027-2028 respectively, to meet the relentless demands of future AI systems.
While massive capacity expansions are underway, they offer no short-term relief. SK Hynix is pouring capital into its HBM lines, targeting 170,000 wafers per year by 2025 and a new $60 billion mega-fab by 2027 designed to reach 350,000 wafers per month. Samsung is countering with its own significant fab investments, including the P5 Fab 1 expected in 2027-2028 and P5 Fab 2 in 2029-2030, each with a capacity of 300,000 wafers per month. The hard reality, however, is that building these advanced facilities takes years. Supply growth will simply not keep pace with the exponential demand curve.
