JPMorgan Analyzes AI Hardware Costs, Memory Shortage May Last Until 2028

TL;DR
· JPMorgan has raised its global memory TAM forecast for 2026 to 2028 by 4% to 8%, expecting shortages to worsen in 2027, with marginal relief in 2028 but not completely disappearing.
· According to the report model, the proportion of memory in CSP hardware capital expenditures may rise from less than 10% before the AI era to 31% in 2026 and 49% in 2027.
· SOCAMM and HBM specification reductions can alleviate delivery pressures, but are insufficient to reverse the shortage, with new DRAM capacity remaining the biggest constraint on the supply side.

JPMorgan raised its memory market size forecast in a global storage report released on August 9, judging that AI servers continue to drive up the value of memory, and the global memory shortage may extend until 2028.

The most striking aspect of this report is not just the bullish outlook on storage prices, but the internal reallocation of AI capital expenditures: memory is rising from a supporting item of less than 10% in the past to a core component that may account for nearly half of CSP hardware spending.

For storage manufacturers, this means continued increases in revenue, profit, and bargaining power; for cloud providers, it means that AI capital expenditures are not just about purchasing GPUs, but HBM, server DRAM, and related memory solutions are also rapidly raising the overall system costs.

JPMorgan has raised its global memory TAM forecast for the fiscal years 2026 to 2028 by 4% to 8%. The report expects the degree of shortage to worsen in 2027, with slight improvement in 2028, but the previously accumulated absolute gap will not completely disappear. These figures are all derived from seller models and should not be directly viewed as already occurring industry facts.

In 2027, nearly half of CSP hardware spending may flow to memory

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

After the second-quarter earnings, the consensus on capital expenditures for the three major cloud service providers and Meta continues to be raised. Meanwhile, based on Omdia hardware capital expenditure data, JPMorgan estimates that the proportion of the memory market size for CSP has risen from less than 10% before the AI era to 31% in 2026, and is expected to further reach 49% in 2027.

This means that by 2027, for every $100 spent on hardware capital expenditures by CSP, about $49 may correspond to memory-related products. Figure 5 further shows that in this model scenario, this proportion may even rise to 60% by 2028.

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

AI servers not only require more HBM but also need more server DRAM, SOCAMM, and other memory solutions deployed around AI CPUs and accelerators. The more GPUs and AI CPUs are shipped, the harder it becomes to continue using the traditional server-era configuration ratios for supporting memory.

However, JPMorgan also acknowledges that a memory value proportion exceeding 50% may lead investors to question the sustainability of AI capital expenditures. If cloud providers cannot further increase hardware budgets in the future, they may continue to lower individual memory specifications or turn to lower-cost tiered memory and storage solutions.

Therefore, the rising proportion of memory is not only a positive for storage manufacturers but also means that the cost and investment return pressure of CSP's AI servers are rising simultaneously.

Specification reductions still fail to reverse the supply-demand gap

A significant reason for the recent volatility in storage stocks is market concerns that the reduction in memory configurations for AI servers indicates a cooling demand.

The report states 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 downgrade the HBM4E configuration from 16-Hi or 12-Hi to 12-Hi or 8-Hi in some SKUs; Rubin may introduce configurations of 288GB and 192GB.

Under more conservative single-card memory assumptions, JPMorgan has lowered its HBM bit demand forecast for 2026 to 2028 by 4% to 19%.

However, in the analytical framework of the report, such adjustments are mainly a way for customers to cope with supply shortages: lowering the memory configuration of individual chips or systems to support more GPU, CPU, and server deliveries, rather than indicating that AI training and inference suddenly no longer require memory.

JPMorgan also acknowledges that specification reductions will indeed cut unit memory demand, narrowing the supply-demand growth gap. However, the shipment volume of new AI chips and servers may still offset the impact of declining individual memory capacities.

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

This means that the marginal supply-demand situation for DRAM may improve in 2028, while NAND will still have a slight gap. However, due to the previous accumulation of shortages, JPMorgan judges that both products will not truly restore supply-demand balance by 2028.

JPMorgan's August model shows that the annual supply-demand growth gap for DRAM will drop to -7% in 2027 and improve to +3% in 2028; NAND will improve from -5% in 2027 to -1% in 2028, but the previously accumulated absolute gap has not completely disappeared.

HBM follows the same logic. JPMorgan estimates that the updated HBM supply-demand gap will be approximately -15%, -14%, and -22% for 2026 to 2028 respectively. Compared to the May model, the specification reductions have narrowed the degree of shortage, but have not turned the market into surplus.

Under more conservative configuration assumptions, some versions of Rubin Ultra's HBM capacity have been reduced from 1024GB or 768GB to 576GB or 384GB, and Rubin may be configured with 192GB or 288GB; after specification reductions, the HBM market has still not turned into surplus.

HBM Shortage Continues, Prices Keep Rising

In the scenario of continued shortages, HBM prices are one of the most closely watched figures in the report.

JPMorgan expects the average price of mixed HBM to rise by 42% year-on-year in 2027, with prices for products of the same specifications expected to rise by 30% to 40%. By 2028, the price increase for products of the same specifications may drop to within 10%, but product structure upgrades may still drive a 22% year-on-year increase in the mixed average price.

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

These figures are all model predictions from the report and do not represent a unified price list determined by suppliers.

One reason supporting prices is that the pace of HBM product upgrades has slowed down. The lifecycle of 8-Hi products has been extended, the ramp-up speed of 12-Hi has decreased, and the adoption of 16-Hi has been delayed. In the context of continued supply constraints, this may actually strengthen the bargaining power of major suppliers.

At the same time, long-term agreements (LTA) with major customers are also beginning to change the way traditional memory cycles fluctuate.

The report shows that the LTA prepayments disclosed by Samsung Electronics and Micron account for about 20% to 25% of the contract value, with some agreements covering or contributing to shipment volume ratios of about 50% to 70%. Most LTAs focus on server and AI memory and adopt differentiated, flexible pricing structures.

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

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

For suppliers like Samsung Electronics, SK Hynix, and Micron, LTAs can enhance the visibility of orders and cash flow and smooth the price increase curve. However, the report also acknowledges that the pricing structure of LTAs is not transparent, and the actual coverage ratios and flexible pricing mechanisms remain to be observed.

If CSP capital expenditures slow down or AI servers continue to reduce memory configurations, the pace of contract execution and final prices may still be adjusted. Therefore, LTAs can reduce the fluctuations of memory cycles but cannot completely eliminate the cycle.

New Capacity Fails to Meet Demand, Shortages Stuck at Wafer Supply

The key premise for the shortage to continue until 2028 is that new supply cannot be delivered in time.

JPMorgan estimates that to achieve balance in DRAM supply and demand by 2028, the industry needs to add approximately 5.5EB of supply, corresponding to about 299,000 wafers per month; NAND needs to add approximately 76EB of supply, corresponding to about 44,000 wafers per month.

JPMorgan estimates that to achieve supply-demand balance by 2028, DRAM needs to add approximately 5.5EB of supply and 299,000 wafers per month; NAND needs to add approximately 76EB of supply and 44,000 wafers per month.

The report believes that even if all publicly announced expansion projects are included in the model, the new supply will still be insufficient to completely fill the gap.

New wafer fabs are not short-cycle actions. It typically takes 2 to 2.5 years from groundbreaking to ramping up production. SK Hynix's M15X has begun to gradually contribute capacity, with the first phase of the Yongin cluster expected to ramp up starting in March 2027; Samsung's P4 is still expanding, with the new P5 expected to start wafer production in the fourth quarter of 2027; Micron's Boise ID1 and PSMC P5 are expected to gradually contribute capacity starting in the second quarter of 2027.

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

HBM will also squeeze traditional DRAM supply. By the end of 2028, HBM-related wafers may account for about 32% of total DRAM capacity. Since the die penalty for HBM is about 3 to 4 times that of ordinary DRAM, producing the same capacity of HBM will consume more wafer resources, further limiting the supply of traditional server and consumer-grade DRAM.

The impact of Chinese manufacturers needs to be discussed separately.

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

In contrast, the report believes that YMTC's NAND unit wafer bit output has roughly caught up with industry leaders. By 2028, its global NAND capacity and bit share may approach 16%, thus the pressure from new Chinese supply on NAND is higher than that on DRAM.

Can High Profits Be Translated into Shareholder Returns

Another clue that is easier to trade in the capital market is the potential shareholder returns of storage manufacturers.

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

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

JPMorgan expects that the shareholder returns of memory manufacturers over the next two to three years may be significantly higher than historical levels, but the related yield is based on assumptions of free cash flow, special dividends, and buybacks.

This could serve as a catalyst for the valuation recovery of storage stocks. The memory stocks tracked by the report have outperformed in the AI supply chain for four consecutive quarters, but have since corrected about 25% since the third quarter of 2026, mainly due to short-term profit forecast downgrades, a slowdown in CSP AI capital expenditure pace, and market concerns about memory configuration optimization.

However, special dividends and buybacks ultimately depend on free cash flow, memory prices, and management decisions, and cannot be viewed as locked-in returns. If AI capital expenditures or LTA execution fall short of expectations, shareholder returns may also be discounted.

The report from JPMorgan reveals not just another round of memory price increases, but an internal reallocation of AI capital expenditures: memory is rising from a supporting item of less than 10% to a core component that may account for nearly half of hardware spending.

For storage manufacturers, this is an opportunity for increased profits and bargaining power; for cloud providers, it means that AI server costs continue to rise. However, a memory value proportion close to or exceeding 50% is also difficult to maintain in the long term, which may force cloud providers to further increase capital expenditures, lower individual configurations, or adopt tiered memory solutions such as CXL, enterprise SSDs, and HBF.

Therefore, what truly needs to be observed is not just whether the shortage can extend to 2028, but which of CSP budget growth, memory specification optimization, and new wafer capacity will change the current supply-demand balance first.