The Memory Crunch, by the Numbers

Where the prices went

TrendForce, a Taiwan-based market research firm, tracks memory contract prices. By its count, server DRAM contract prices rose a cumulative 64% in the second half of 2025, and it expected a further jump of about 270% across 2026 Estimate.¹ Enterprise SSDs followed the same path: up around 35% in 2H25 and projected to rise a cumulative 235% in 2026.¹

For the fourth quarter of 2026 TrendForce expects the pace to slow rather than reverse: conventional DRAM contract prices up 10–15% quarter on quarter and NAND flash up 15–20%. Enterprise SSDs are the only memory category where it expects price growth to accelerate in 4Q26.²

High-bandwidth memory (HBM), the stacked DRAM packaged next to AI accelerators, is moving the most. In August TrendForce said HBM contract prices could rise 70–140% in 2027, even though some long-term agreements signed from 2Q26 onward include price ceilings.¹ A month later it revised its 2027 outlook upward, projecting the blended average selling price of HBM to rise 121% year on year, citing persistent supply constraints, a growing share of higher-priced HBM4 and the ramp of HBM4E in the second half of 2027.³

Why AI servers squeeze everything else

The demand side is the cloud. TrendForce estimates that major cloud service providers' (CSPs') total capital spending will rise 98% year on year in 2026 and another 50% in 2027, and that DRAM and NAND flash together will account for 47% of that spending in 2026 and 68% in 2027 Estimate.¹ It also estimates that HBM and RDIMM, the registered memory modules used in servers, will together take 51% of DRAM bit supply in 2026.¹

Suppliers are steering their most advanced process capacity to server products, which TrendForce says keeps the overall DRAM market undersupplied in 4Q26. Mobile DRAM allocations stay tight as capacity shifts toward servers, and PC DRAM supply could decline in 2027 for the same reason. Even server DRAM can't catch up: component supply, packaging and testing capacity, and limited flexibility in front-end DRAM production keep it short.²

The makers say the same thing in their own words. Samsung's second-quarter presentation describes addressing AI demand "with primary focus on server despite limited capacity" and expects demand for server DRAM, enterprise SSDs and HBM to keep the market undersupplied in the second half of 2026, with supply constraints continuing "in spite of efforts to increase production" Vendor-stated.⁷ SK hynix says demand for both AI memory and conventional memory is expanding together as AI evolves into agentic forms.⁶

Why HBM eats wafer capacity

HBM is DRAM in a different shape. Micron's quarterly filing describes it as "a 3D stacked DRAM architecture that utilizes through-silicon via ('TSV') connections," which gives it higher bandwidth at lower power than other memory types.⁴ The cost is silicon. In the same filing Micron states: "Due to the higher performance and more complex manufacturing process, HBM requires a higher number of wafers and more cleanroom space to produce the same number of bits as conventional DRAM in the same technology node" Vendor-stated.⁴ Every wafer moved to HBM therefore removes more conventional DRAM bits than it adds in HBM bits. Micron also notes the reverse risk: if HBM demand weakened and suppliers shifted capacity back, conventional DRAM supply could rise significantly.⁴

TrendForce describes HBM and conventional DRAM as competing for limited advanced-process and wafer capacity, with yields and output for next-generation products taking time to ramp, which keeps memory supply tight through 2027.³

The stack height matters too. An 8-Hi HBM stack uses fewer DRAM dies than a 12-Hi one, so the same HBM bit output can serve more GPUs, and TrendForce reports that GPU and ASIC makers are considering cutting HBM capacity per device for that reason. But every stack needs one base die whatever its height, so the 8-Hi stack spreads that cost over fewer bits; TrendForce expects 8-Hi parts to carry a price premium of roughly 10–20% per gigabit over 12-Hi in 2027 Estimate.³

Who makes HBM

Counterpoint Research estimates that SK hynix held 50% of global HBM revenue in Q2 2026, still first but down from 64% a year earlier; Samsung rose from 21% in Q1 2026 to 33%; and Micron held 18%.⁵ Counterpoint says most HBM revenue still comes from HBM3E, with HBM4 shipments expected to become visible in earnest from the second half of 2026.⁵

Shares of global HBM revenue as published by Counterpoint Research (Memory Tracker and Forecast Insights Report, Q2 2026), rounded by Counterpoint; they are not shares of bits shipped.

The newest generation is where the makers compete. SK hynix says it began mass shipments of HBM4 in the second quarter of 2026, will ramp production in the second half, and completed HBM4E sample shipments in the first half Vendor-stated.⁶ Samsung says it scaled up HBM4 sales and shipped the industry's first HBM4E samples to major customers, a claim that is its own.⁷

What the makers reported

The shortage shows up in the makers' operating results, given here exactly as each company reported them Vendor-stated. SK hynix reported second-quarter 2026 revenue of 79.3187 trillion won and operating profit of 60.5426 trillion won, a 76% operating margin, against revenue of 22.232 trillion won a year earlier; it said both DRAM and NAND flash prices rose significantly quarter on quarter.⁶ Samsung's Memory business reported sales of KRW 120.8 trillion in Q2 2026, against 21.2 trillion in Q2 2025.⁷

Micron's fiscal fourth quarter, which ended September 3, 2026, brought revenue of $54.23 billion, against $11.32 billion in the same quarter a year earlier, with a GAAP gross margin of 86.8% of revenue.⁸ Its Cloud Memory Business Unit, which covers memory for large hyperscale cloud customers and HBM for all data center customers,⁴ reported revenue of $16,283 million, against $4,543 million a year earlier; its Core Data Center Business Unit reported $18,002 million, against $1,577 million.⁸

Supply is increasingly sold ahead under contract. SK hynix says it has finalized long-term agreements with around 10 customers and is in further discussions with others.⁶ TrendForce notes that pricing mechanisms in such agreements can make some suppliers' increases lag the market average.²

Where China's CXMT fits

ChangXin Memory Technologies (CXMT), which Reuters describes as China's leading maker of DRAM chips, was founded with state backing in 2016 and runs three 12-inch DRAM fabs in Beijing and Hefei.⁹ Omdia data cited in CXMT's prospectus, as reported by Reuters, put its share of the global DRAM market at 4% in the second quarter of 2025, while Micron, SK hynix and Samsung together controlled more than 90% Estimate.⁹ Reuters also reported that CXMT aims to begin production by the end of 2026 at an HBM back-end packaging facility it is building in Shanghai.⁹ We couldn't open a primary source for a newer CXMT share figure, so none is given here.

What could ease it

New fabs come slowly. SK hynix says it is accelerating the mass-production schedule of its M15X fab and investing to expand capacity quickly after the Yongin Phase 1 cleanroom opens in early 2027.⁶ TrendForce expects process migrations and new fabs ramping in the second half of 2027 to lift combined server DRAM and HBM bit supply by 27% in 2027 Estimate.¹

Buyers can also use less. TrendForce says high memory costs give server and AI chip suppliers such as Nvidia more reason to raise prices, and that CSPs may respond by cutting memory per system: changing RDIMM configurations, reducing the HBM on future AI chips or exploring AI ASICs with the model architecture hardwired into the silicon.¹ Even so, it expects HBM to stay supply-constrained in 2027, with capacity cuts per device easing some GPU cost pressure but not reversing the rise in HBM prices.³