TL;DR
· JPMorgan has raised its global memory TAM forecast for 2026 to 2028 by 4% to 8%, expecting the shortage to worsen in 2027 and see marginal relief in 2028 without fully disappearing.
· According to the reporting model, memory's share of CSP hardware capital expenditures may rise from under 10% before the AI era to 31% in 2026 and 49% in 2027.
The reduction in SOCAMM and HBM specifications can alleviate delivery pressure, but it is insufficient to reverse the shortage; additional DRAM capacity remains the greatest constraint on the supply side.

JPMorgan raised its forecast for the global memory market size in its global storage report released on August 9, concluding that AI servers continue to drive up the value of memory, and global memory shortages may persist until 2028.

What stands out most in this report is not just the bullish storage prices, but the internal reallocation of AI capital expenditures: memory is rising from a previously under 10% component to a core part potentially accounting for nearly half of CSP hardware spending.

For storage vendors, this means continued growth in revenue, profits, and bargaining power; for cloud providers, it means AI capital expenditures are no longer just about purchasing GPUs—HBM, server DRAM, and related memory solutions are rapidly increasing the overall system cost.

JPMorgan raised its global memory TAM forecast for fiscal years 2026 to 2028 by 4% to 8%. The report anticipates that the shortage will intensify in 2027 and slightly improve in 2028, but the accumulated absolute gap from prior years will not be fully closed. These figures are derived from sell-side models and should not be construed as established industry facts.

In 2027, nearly half of CSP hardware spending may go toward memory.

Over the past two years, discussions around AI capital expenditures have focused more on GPUs, advanced packaging, and data center power. However, the increasing value of memory in AI servers is becoming equally critical.

After the Q2 earnings reports, consensus estimates for capital expenditures by the three major cloud providers and Meta have continued to rise. Meanwhile, according to JPMorgan’s calculations based on Omdia’s hardware capex data, the proportion of the memory market size targeting CSPs relative to their total hardware spending has increased from under 10% before the AI era to 31% in 2026, and is projected to reach 49% in 2027.

That means, by 2027, approximately $49 of every $100 in CSP hardware capital expenditures could be allocated to memory-related products. Figure 5 further shows that, under this model scenario, this proportion could rise to as high as 60% by 2028.

JPMorgan estimates that the proportion of CSP-level memory TAM to CSP hardware capital expenditures could rise from under 10% before the AI era to 31% in 2026, 49% in 2027, and 60% in 2028; the report also notes that a share exceeding 50% may be difficult to sustain long-term.

AI servers require not only more HBM, but also more server DRAM, SOCAMM, and other memory solutions around AI CPUs and accelerators. As more GPUs and AI CPUs are shipped, it becomes increasingly difficult to maintain the traditional memory configuration ratios from the era of conventional servers.

But JPMorgan also acknowledged that memory value占比 exceeding 50% may cause investors to question whether AI capital expenditures are sustainable. If cloud providers cannot further increase hardware budgets in the future, they may continue to reduce per-server memory specifications or shift toward lower-cost tiered memory and storage solutions.

Therefore, the increased memory usage is not only beneficial for memory manufacturers but also means that CSPs are facing rising costs and investment return pressures for their AI servers.

Specifications reduced, yet still unable to reverse the supply-demand gap

A key recent factor in the volatility of memory stocks is market concern that a reduction in AI server memory configurations signals weakening demand.

Reports indicate that the SOCAMM capacity of NVIDIA's Vera CPU has been reduced from 1.5TB per chip to 768GB. Rubin Ultra may also reduce the number of compute dies, and for some SKUs, lower the HBM4E configuration from 16-Hi or 12-Hi to 12-Hi or 8-Hi; Rubin may offer configurations of 288GB and 192GB.

Under a more conservative single-GPU memory assumption, JPMorgan has reduced its HBM bit demand forecast for 2026 to 2028 by 4% to 19%.

However, within the report’s analytical framework, such adjustments primarily reflect how customers are responding to supply shortages: reducing memory configurations per chip or per system to enable more GPU, CPU, and server deliveries, rather than indicating that AI training and inference have suddenly stopped requiring memory.

JPMorgan Chase also acknowledges that the specification downgrade does indeed reduce per-unit memory demand and narrows the gap between supply and demand growth. However, the increased shipment volume of new AI chips and servers may still offset the impact of reduced memory capacity per unit.

According to the updated model, the annual supply-demand growth gap for DRAM is approximately -3% in 2026, widening to -7% in 2027; by 2028, supply growth may exceed demand by about 3 percentage points. The annual supply-demand growth gap for NAND is approximately -3%, -5%, and -1%, respectively.

This means the marginal supply and demand situation for DRAM may improve by 2028, while NAND will still have a slight shortfall. However, due to the accumulated shortages from before, J.P. Morgan concludes that neither product will truly regain supply-demand balance by 2028.

JPMorgan's August model shows the annual supply-demand gap for DRAM declining to -7% in 2027 and improving to +3% in 2028; for NAND, the gap improves from -5% in 2027 to -1% in 2028, but the previously accumulated absolute deficit has not yet been fully closed.

HBM follows the same logic. JPMorgan estimates that the updated HBM supply-demand gap will be approximately -15%, -14%, and -22% in 2026, 2027, and 2028, respectively. Compared to the May model, the downward revision in specifications has narrowed the extent of the shortage, but the market has not yet shifted into surplus.

Under more conservative configuration assumptions, the HBM capacity for certain Rubin Ultra versions has been reduced from 1024GB or 768GB to 576GB or 384GB, while Rubin may be configured with 192GB or 288GB; even after these specification reductions, the HBM market has not turned into a surplus.

HBM shortage continues, prices keep rising

In the scenario where shortages persist, HBM prices are among the most closely watched figures in the report.

JPMorgan expects the average price of hybrid HBM to rise 42% year-over-year in 2027, with prices for products of the same specification potentially increasing by 30% to 40%. By 2028, the price increase for products of the same specification may fall below 10%, but product structure upgrades could still drive the average hybrid price up by 22% year-over-year.

JPMorgan expects the HBM supply-demand gap to be approximately -15%, -14%, and -22% in 2026, 2027, and 2028 respectively, with cumulative shortages expanding; the average price of hybrid HBM may rise by 42% in 2027 and another 22% in 2028 driven by product upgrades.

These figures are model predictions from the report and not a confirmed, standardized price list from the vendor.

One reason for the support price is that the product upgrade cycle for HBM has slowed down. The lifecycle of 8-Hi products has been extended, the ramp-up speed of 12-Hi has decreased, and the adoption timeline for 16-Hi has been delayed. Under continued supply constraints, this may instead strengthen the bargaining power of major suppliers.

Meanwhile, long-term agreements with large customers, known as LTAs, are also beginning to alter the traditional volatility of memory cycles.

The report shows that Samsung Electronics and Micron disclosed that their LTA advance payments account for approximately 20% to 25% of the contract value, with certain agreements covering or contributing to shipment volumes of about 50% to 70%. Most LTAs focus on server and AI memory and feature differentiated, flexible pricing structures.

The LTAs disclosed by major memory manufacturers vary in terms of duration, shipment coverage, advance payments, and pricing mechanisms, but overall contract terms are more favorable to suppliers and primarily target server and AI memory.

JPMorgan estimates that CSP and AI-related demand could account for more than 70% of total bit volume and over 85% of revenue in related contracts. Server memory also enjoys a significant price and margin premium compared to non-server applications.

For suppliers such as Samsung Electronics, SK Hynix, and Micron, LTAs can enhance visibility into orders and cash flow and smooth out price increase curves. However, the report also acknowledges that LTA pricing structures lack transparency, and the actual coverage ratio and flexible pricing mechanisms remain to be seen.

If CSP capital expenditures slow down, or if AI servers continue to reduce memory configurations, the pace of contract execution and final pricing may still be adjusted. Therefore, LTAs can reduce volatility in the memory cycle but cannot eliminate it entirely.

New production capacity can't keep up with demand, with shortages stemming from wafer supply constraints.

The continuation of the shortage until 2028 hinges on the difficulty of new supply coming online in a timely manner.

JPMorgan estimates that to balance DRAM supply and demand by 2028, the industry needs an additional supply of approximately 5.5 EB, equivalent to about 299,000 wafers per month; for NAND, an additional supply of approximately 76 EB is required, equivalent to about 44,000 wafers per month.

JPMorgan estimates that to achieve supply-demand balance by 2028, DRAM requires an additional supply of approximately 5.5 EB and a capacity increase of 299,000 wafers per month; NAND requires an additional supply of approximately 76 EB and a capacity increase of 44,000 wafers per month.

The report concludes that even when including all publicly announced expansion projects in the model, the additional supply still falls short of fully closing the gap.

Building a new wafer fab is not a short-cycle endeavor. It typically takes 2 to 2.5 years from groundbreaking to ramping up to full production. SK Hynix’s M15X has already begun contributing capacity incrementally; the first phase of the Yongin cluster is expected to ramp up starting March 2027; Samsung’s P4 is still expanding, with the new P5 facility projected to begin wafer output in the fourth quarter of 2027; Micron’s Boise ID1 and PSMC’s P5 are both expected to gradually contribute capacity starting in the second quarter of 2027.

Even if these projects proceed as scheduled, JPMorgan expects global DRAM monthly wafer capacity to rise only from approximately 1.9 million wafers at the end of 2025 to 2.85 million wafers by the end of 2028, still below the level needed to achieve supply-demand balance.

HBM itself will also squeeze traditional DRAM supply. By the end of 2028, HBM-related wafers may account for approximately 32% of total DRAM capacity. Due to HBM’s die penalty being about 3 to 4 times that of standard DRAM, producing equivalent capacity in HBM consumes more wafer resources, further limiting supply for traditional server and consumer-grade DRAM.

The impact brought by Chinese manufacturers needs to be discussed separately.

JPMorgan expects CXMT's DRAM capacity and bit share to reach approximately 16% and 11% by 2028, respectively, but it still lags behind leading manufacturers by about 2 to 3 years in high-density server DRAM and products above HBM3E. Therefore, its capacity expansion will have limited direct impact on the high-end AI memory market in the short term.

In contrast, the report concludes that YMTC's NAND bits per wafer output has largely caught up with industry leaders. By 2028, its global NAND capacity and bit share could approach 16%, meaning the additional supply pressure on NAND from China is higher than that on DRAM.

Can high profits be converted into shareholder returns?

Another clue that capital markets favor easier trading is the potential shareholder returns from storage manufacturers.

Both Samsung Electronics and SK Hynix's current shareholder return policies require allocating 50% of cumulative free cash flow to shareholder returns. Samsung's current plan covers 2024 through 2026, while SK Hynix's plan extends through 2027.

According to the Morgan Stanley model, Samsung Electronics' annual cash yield for 2026 and 2027 may both be around 8%, totaling approximately 16% over the two years; SK Hynix's cash yield in 2027 could rise to approximately 16.7%. The report therefore expects the cumulative cash return for both companies over the next two years to reach approximately 16% to 20%.

JPMorgan expects that memory manufacturers' shareholder returns over the next two to three years could significantly exceed historical levels, but these yields are based on assumptions regarding free cash flow, special dividends, and buybacks.

This could serve as a catalyst for the valuation rebound of memory stocks. After outperforming for four consecutive quarters along the AI industry chain, memory stocks have seen a pullback of approximately 25% since Q3 2026, with pressure primarily stemming from downward revisions to short-term earnings forecasts, a slowdown in CSP AI capital expenditure timing, and market concerns over memory configuration optimization.

However, special dividends and buybacks ultimately depend on free cash flow, memory prices, and management decisions, and should not be viewed as guaranteed returns. If AI capital expenditures or LTA execution fall below expectations, shareholder returns may also be reduced.

This JPMorgan report reveals not just another round of memory price increases, but a reallocation within AI capital expenditures: memory is rising from a minor component accounting for less than 10% in the past to a core component potentially accounting for nearly half of hardware spending.

For storage vendors, this is an opportunity for increased profits and stronger bargaining power; for cloud vendors, it means continued rising costs for AI servers. However, memory accounting for nearly or even over 50% of total value is difficult to sustain long-term, potentially forcing cloud vendors to further increase capital expenditures, reduce per-server configurations, or adopt tiered memory solutions such as CXL, enterprise SSDs, and HBF.

Therefore, what truly needs to be monitored is not just whether the shortage will last until 2028, but which of the following will first shift the current supply-demand balance: CSP budget growth, memory specification optimizations, or new wafer capacity.

Click to learn about the open positions at BlockBeats

Welcome to the official BlockBeats community:

Telegram subscription group: https://t.me/theblockbeats

Telegram group: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia