Memory Share Hits 49% of CSP Spend as AI Shortages Persist to 2028

Woofun AI reports that JPMorgan issued a global memory report on August 9, projecting that the global memory shortage will persist until 2028 due to sustained demand from AI servers. The core thesis identifies a structural shift in capital expenditure, where memory transitions from a minor component to a dominant cost driver, fundamentally altering the economic landscape for both manufacturers and cloud providers.

The reallocation of AI capital expenditures marks a decisive break from historical norms. Previously, memory accounted for less than 10% of Cloud Service Providers' (CSPs) hardware spending.

However, this share is projected to climb to nearly half of total hardware budgets. High Bandwidth Memory (HBM) and server DRAM are no longer peripheral components but central elements of AI infrastructure.

This shift implies that AI capital expenditures are increasingly defined by memory acquisition rather than solely by GPU purchases, raising the total cost of ownership for entire server systems.

JPMorgan raised its Total Addressable Market (TAM) forecasts for the global memory sector for the 2026–2028 fiscal years by 4% to 8%. The report predicts that the supply deficit will intensify in 2027 before showing slight improvement in 2028, though the cumulative shortage from earlier periods will not be fully resolved. These projections are derived from seller-side models and should be interpreted as indicative trends rather than established industry facts, reflecting the inherent uncertainty in long-term supply chain forecasting.

The share of memory in CSP spending is expected to reach 31% by 2026 and surge to 49% in 2027. This means that for every $100 spent on hardware, approximately $49 will be allocated to memory-related products. Figure 5 in the report further illustrates that under this model scenario, the ratio could escalate to 60% by 2028. This trajectory underscores the growing financial weight of memory in AI infrastructure, challenging traditional budgeting assumptions for cloud providers.

AI servers require not only increased volumes of HBM but also specialized solutions like SOCAMM, tailored for AI CPUs and accelerators. NVIDIA's Vera CPU now features SOCAMM capacities reduced from 1.5TB per chip to 768GB. The Rubin Ultra series may also see reductions in computing dies, with HBM4E configurations in some SKUs lowered from 16-Hi or 12-Hi to 12-Hi or 8-Hi. The Rubin series might offer 288GB and 192GB configurations instead, reflecting a strategic adjustment to manage costs and supply constraints.

Under a conservative assumption regarding memory per card, JPMorgan reduced its HBM bit demand forecasts for 2026–2028 by 4% to 19%.

However, the annual supply-demand gap for DRAM is still expected to be around -3% in 2026 and widen to -7% in 2027, before supply growth surpasses demand by about 3 percentage points in 2028. For NAND, the gaps are estimated at -3%, -5%, and -1% respectively. HBM's supply-demand gap is projected to remain tight at -15%, -14%, and -22% for 2026–2028. Consequently, HBM prices are predicted to rise 42% year-on-year in 2027, with identical specifications increasing by 30% to 40%. By 2028, price increases for identical specs may fall below 10%, but product upgrades could drive the average mixed price up by 22% year-on-year.

Long-term agreements (LTAs) are reshaping supplier power dynamics. Samsung Electronics and Micron have disclosed LTA advance payments amounting to 20% to 25% of contract value, with some agreements covering 50% to 70% of total shipment volume. Most LTAs relate to server and AI memory, featuring flexible pricing. JPMorgan estimates that CSP and AI-related demands may account for over 70% of total bit volume and over 85% of revenue from these contracts. For suppliers like Samsung Electronics, SK Hynix, and Micron, LTAs improve order visibility and cash flow, smoothing out price increase curves, although the actual coverage ratios and pricing mechanisms remain opaque.

To achieve supply-demand balance for DRAM by 2028, the industry needs to add about 5.5EB in supply, corresponding to a wafer production capacity of around 299,000 wafers per month. For NAND, an additional 76EB is needed, equivalent to 44,000 wafers per month. Building new wafer factories takes 2 to 2.5 years from groundbreaking to full capacity. SK Hynix's M15X facility is already contributing, with Phase 1 of the Yongin cluster expected to ramp up from March 2027.

Samsung's P4 is under expansion, with P5 starting in the fourth quarter of 2027. Micron's Boise ID1 and PSMC P5 are expected to start in the second quarter of 2027. Even with these projects, global DRAM monthly wafer production capacity is estimated to rise from 1.9 million wafers at the end of 2025 to 2.85 million wafers by the end of 2028, still insufficient for balance. HBM-related wafers may account for 32% of total DRAM capacity by 2028, consuming 3 to 4 times the wafer resources of regular DRAM.

Chinese manufacturers present a distinct variable. JPMorgan estimates that CXMT's DRAM production capacity and bit share will reach 16% and 11% by 2028, but it will lag by 2 to 3 years in high-density server DRAM and HBM3E products. Thus, its impact on high-end AI memory is limited in the short term. In contrast, YMTC's NAND bit output per wafer has caught up with leaders. By 2028, its global NAND production capacity and bit share may both approach 16%, indicating greater supply pressure for NAND than for DRAM.

Regarding shareholder returns, Samsung Electronics and SK Hynix allocate 50% of cumulative free cash flow to returns. Samsung's plan covers 2024–2026, while SK Hynix's runs until 2027. JPMorgan models Samsung's annual cash yield at 8% in 2026 and 2027, totaling 16% over two years. SK Hynix's yield may rise to 16.7% in 2027. The combined cash return rate for these two companies over the next two years is expected to reach 16% to 20%, potentially catalyzing valuation recovery.

However, memory stocks corrected 25% since the third quarter of 2026, driven by earnings revisions and configuration concerns.

The strategic implication is a reallocation within AI capital expenditures, where memory shifts from a minor component to a core element accounting for nearly half of hardware spending. This creates higher profits for manufacturers but rising costs for cloud providers. A memory value share exceeding 50% is unsustainable, potentially forcing CSPs to increase budgets, reduce per-server configurations, or adopt tiered solutions like CXL, enterprise SSDs, and HBF. The critical observation is not just whether shortages persist until 2028, but which factor—CSP budget growth, memory specification optimization, or new wafer production capacity—will first restore supply-demand balance.