Is NVIDIA's AI Memory Spec Cut a False Signal? JPMorgan: AI-Driven Super Cycle Could Extend to 2028

Reports of specification cuts in NVIDIA's (NVDA-US) new platform initially led markets to interpret this as a sign of cooling demand. However, a recent report by JPMorgan (JPM-US) refutes this view, arguing that the changes are merely a temporary measure due to supply constraints, and that memory manufacturers are set to enter the longest upcycle in history.

Recently, global memory stocks have seen a significant pullback. Coupled with reports that NVIDIA's next-generation AI platforms are scaling down specifications for High Bandwidth Memory (HBM) and SOCAMM modules, many investors have begun to question whether the current memory bull market has peaked.

Nonetheless, JPMorgan's latest report presents a starkly different outlook. It argues that supply-demand imbalances are not easing—in fact, they could reach their most severe point in 2027. The price-increase cycle is expected to continue through 2028 before showing signs of moderation.

The bank has significantly raised its future market size forecasts, revising up its global DRAM and NAND flash memory market projections for 2026 to 2028 by 4%, 6%, and 8% respectively. This translates to market sizes of $970.9 billion, $1.442 trillion, and $1.826 trillion.

The report emphasizes that this cycle will far outlast any typical memory cycle seen over the past two decades.

Spec Reductions Reflect Supply Constraints, Not Cooling Demand

One trigger for market concern was the reported reduction in memory capacity per chip in AI servers.

The report notes that the SOCAMM module capacity on NVIDIA's Vera processor has been reduced from the originally planned 1.5TB to 768GB. The number of HBM4E stacks on Rubin Ultra has also been revised from the planned 16 layers to either 8 or 12 layers. As for the Rubin GPU, two versions of HBM4—288GB and 192GB—may be launched simultaneously.

JPMorgan believes these adjustments do not indicate shrinking AI appetite for memory, but rather reflect customer adaptations to limited supply.

As AI application focus shifts from model training to large-scale inference, data throughput from long-text processing, KV caching, and AI agents continues to rise. Rather than clinging to high memory capacity per chip, companies are increasingly adopting a 'scale-out' strategy—deploying more chips to meet overall computing and storage needs.

Notably, memory demand is expanding from GPUs to CPUs. JPMorgan forecasts that AI-focused CPU shipments will achieve a compound annual growth rate of up to 155%.

In other words, even if per-chip memory configurations shrink, as long as chip shipment volumes grow fast enough, overall memory demand can remain strong.

The bank interprets this 'downsizing' as a transitional phase in AI infrastructure development—from single-device stacking to large-scale deployment—rather than a precursor to an economic reversal.

DRAM Supply-Demand Tightest in 2027; Price Hikes Expected for Five Years

Among memory products, DRAM supply pressure is particularly evident.

JPMorgan estimates that global DRAM bit demand will grow by approximately 30%, 31%, and 25% from 2026 to 2028, while supply growth will be only around 26.5%, 24.4%, and 24.8% respectively. The gap between demand and supply growth will widen to about 7 percentage points in 2027, marking the most imbalanced year.

The primary driver is servers rapidly replacing PCs and smartphones as the core source of DRAM demand.

The report projects that servers' share of global DRAM bit demand will rise from 39% in 2025 to 56% in 2026, 67% in 2027, and further to 74% in 2028. In contrast, the shares of PCs and mobile devices will fall to around 7% and 13% respectively.

This indicates a clear decline in the memory industry's reliance on traditional consumer electronics cycles.

Driven by supply shortages, JPMorgan forecasts DRAM average selling prices will surge 236% in 2026, rise another 29% in 2027, and still increase 7% in 2028. Corresponding market sizes are projected at $634.1 billion, $982.9 billion, and $1.31 trillion.

In terms of price trends, DRAM has been on a year-on-year upward trajectory since Q1 2024. JPMorgan estimates this uptrend could continue through the end of 2028—approximately 20 consecutive quarters of price increases—far exceeding the typical 6–7 quarter cycles of the past two decades.

HBM Remains in Short Supply Even with Downgrades; Prices Expected to Rise Over 40% in 2027

HBM is the product with the most significant projected changes in the report. Considering specification downgrades in the Rubin series, delays in 12-layer HBM mass production, and postponed timelines for 16-layer products, JPMorgan has revised down its HBM bit demand forecasts for 2026–2028 by 4%, 10%, and 19% respectively.

Even so, supply gaps persist. The report estimates HBM supply-demand gaps of approximately 15%, 14%, and 22% from 2026 to 2028. Even assuming about 70% of Rubin and Rubin Ultra products switch to downgraded HBM specs, shortages cannot be fully resolved.

Against this backdrop, the HBM market size is expected to expand from $33.5 billion in 2025 to $64.3 billion in 2026, $145.6 billion in 2027, and peak at $252.1 billion in 2028.

On pricing, JPMorgan forecasts the blended average selling price of HBM will increase 42% in 2027 to about $2.80 per Gb, and rise another 22% to $3.40 in 2028. Price increases for same-spec products in 2027 are estimated between 30% and 40%. New-generation products typically command a 15%–30% premium over previous generations—for example, HBM4E is expected to carry about a 20% premium over HBM4.

As HBM's share of total revenue grows, Samsung Electronics' (005930KS) HBM market share is projected to rise to 37%–38% in 2027–2028, approaching its overall DRAM market share. Samsung, SK Hynix (000660KS), and Micron Technology (MU-US) are currently jointly investing in R&D for HBM5-generation thermal and packaging technologies.

Capacity Expansion Can't Keep Up with Demand; Massive Capex Still Insufficient

Memory giants are not ignoring the shortage, but JPMorgan points out that massive capital expenditures in this cycle may not signal oversupply as they have in the past.

First, building a new wafer fab and reaching full capacity typically takes 2 to 2.5 years. Second, EUV lithography equipment intensity and construction costs continue to rise.

More critically, HBM production displaces significant DRAM wafer capacity and carries a 'die penalty' of about three to four times—meaning producing the same number of bits in HBM consumes 3–4x the wafer resources of traditional DRAM. As a result, each dollar of capital expenditure yields fewer DRAM bit outputs year after year.

JPMorgan estimates global DRAM capex will rise from $59.2 billion in 2025 to $99.8 billion in 2026 and $144.3 billion in 2027, reaching $154.6 billion in 2028. Annual increases in 2026 and 2027 will be as high as 68% and 45% respectively.

Even so, to achieve DRAM supply-demand balance by 2028, an additional 300,000 wafers per month would be needed—requiring about $58 billion in additional investment, equivalent to a 41% increase over the existing 2027 capex forecast. For NAND, an additional 45,000 wafers per month and about $7 billion in investment would be needed.

JPMorgan states bluntly that filling this capacity gap in such a short time is 'physically extremely difficult'.

AI Inference Drives 'Memory Wall' Effect; NAND Gains New Growth Momentum

Compared to DRAM, NAND flash memory historically faced higher risks of long-term oversupply, as manufacturers could rapidly increase bit output on the same wafer area by stacking layers toward 300 or 400 layers.

However, JPMorgan...