The next constraint on AI scaling is shifting from GPUs to the memory wall

Overview

This report discusses how, as AI enters a new stage, memory is shifting from a peripheral topic to a core constraint in AI infrastructure. The report argues that GPUs determine AI runtime speed, while memory determines the boundary of AI scalability. As generative AI and agentic AI drive rising demand for HBM, DRAM, enterprise SSDs, and more, the industry must break through the "memory wall" via broader system-level innovation.

Core views

The core view is that the bottleneck in AI systems is no longer just compute, but whether data can be stored, accessed, and moved at sufficiently low cost, high bandwidth, and large capacity. Traditional node migration remains important, but it is insufficient on its own; future investment opportunities will expand from traditional memory makers to a broader ecosystem including interfaces, interconnects, advanced packaging, system design, materials, and equipment. The report believes that early-stage, less crowded parts of this ecosystem may generate stronger alpha.

Analysis framework

Starting from system constraints, the report first defines the "memory wall" in AI scaling, then uses the historical case of 3D NAND to show how disruptive memory innovation can break physical limits and reshape the supply chain. It then proposes six innovation pathways and combines TAM estimation, company exposure, valuation, and catalysts to identify potential investment opportunities.

Methodology notes

• Technology innovation frameworkSix memory innovation pathways

  Design, process, packaging, peripherals, integration, materials

  This framework breaks next-generation memory innovation into six assessable pathways to help investors judge which technologies have scalability, commercial relevance, and investment value.

• System bottleneck analysisMemory wall

  Mismatch between compute growth and memory capacity, bandwidth, and cost

  AI models and token demand are rising rapidly, but memory systems cannot provide data at the same pace, constraining overall system performance and deployment economics.

• Historical analogy3D NAND case study

  Disruptive memory architecture innovation

  3D NAND broke the physical scaling limits of 2D planar NAND through vertical stacking, increasing density and changing the supply chain; the report uses this as an analogy for the potential impact of current next-generation memory innovation.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Samsung Electronics

  Core memory supplier benefiting from DRAM/HBM demand driven by AI compute and agentic AI

  Strengths

  The report maintains OW and Top Pick, expecting DRAM prices to rise more than 20%-30% in 3Q26, with earnings revisions supporting valuation stability.

  Weaknesses

  YoY price momentum may plateau in 4Q26, potentially limiting near-term cyclical catalysts.

  Comparison

  Along with other major memory leaders, it is among the most direct beneficiaries.

  Risks

  Supply-demand visibility, price cycles, and changes in long-term supply discipline.

• Kioxia

  NAND and SSD portfolio benefits from AI demand for bandwidth, latency, IOPS, and capacity

  Strengths

  The report maintains Overweight/Top Pick; the CM, GP, and LC series address KV cache, GPU-direct low-latency, and large-capacity AI infrastructure demand, respectively.

  Weaknesses

  The pace of NAND demand shifting toward AI use cases still needs validation.

  Comparison

  Compared with traditional NAND cycles, AI infrastructure may open up higher-value use cases.

  Risks

  Commercialization pace, customer adoption, and NAND price cycles.

• Micron

  Beneficiary of the DRAM and HBM upcycle

  Strengths

  The report maintains Overweight with a target price of $1,200; HBM execution is seen as not fully priced in, and strategic customer agreements improve revenue visibility.

  Weaknesses

  Valuation remains relatively sensitive to the sustainability of AI data center demand and HBM share retention.

  Comparison

  Together with Samsung and SK hynix, it is in the core HBM/DRAM competitive landscape.

  Risks

  HBM share, gross margin delivery, capital expenditure, and the degree of supply-demand tightness.

• SanDisk

  Beneficiary of NAND demand in AI inference, KV cache, and long-context storage

  Strengths

  The report maintains Overweight with a target price of $1,750; AI may move NAND up the storage hierarchy and increase the importance of cloud customers.

  Weaknesses

  Comparable company data in the table show a case where the current price is above the target price, so the near-term upside framing should be verified against the original table.

  Comparison

  Compared with price-sensitive markets such as PC/mobile, hyperscaler demand may support higher margins.

  Risks

  Shifts in NAND demand structure, delivery of NBM agreements, and discipline around new capacity.

• Montage Technology

  Beneficiary of memory interfaces, CXL, MRDIMM, DDR6, and increasing memory content in AI servers

  Strengths

  The report treats it as a high-conviction idea, believing it has leveraged exposure to the "plumbing" layer of next-generation memory systems, and raises its target price.

  Weaknesses

  Its business is affected by standards adoption, server platform timing, and customer validation cycles.

  Comparison

  Compared with traditional memory manufacturers, Montage is more of a differentiated alpha opportunity in the connectivity and interface ecosystem.

  Risks

  CXL/MRDIMM/DDR6 commercialization slower than expected, intensifying competition, and valuation volatility.

• GigaDevice、Winbond、AP Memory

  Potential beneficiaries in advanced packaging, wafer stacking, and edge AI specialized memory

  Strengths

  Benefit from wafer-on-wafer stacking, specialized memory, and demand for higher bandwidth, energy efficiency, and form-factor optimization.

  Weaknesses

  Many technologies are still in the early adoption stage, with lower scaling certainty than mainstream DRAM/HBM.

  Comparison

  Relative to mainstream memory suppliers, these companies are more part of the "enabler ecosystem."

  Risks

  Technology maturity, customer onboarding, cost curves, and the speed of use-case expansion.

Key data

• Cloud storage spending forecastUS$418bn by 2030Based on current forecasts, the report believes the CAGR from 2026 onward will be about 8%.
• Contribution of agentic AI to incremental DRAM demand26%-77% by 2030The report believes agentic AI could become an important source of incremental DRAM demand.
• Memory as a share of cloud capex40% in 2027eThe report says this is above the 12% level before the AI-driven cloud capex boom in 2023.
• Comparison of memory bandwidth improvement versus token growthDDR5 per-channel bandwidth about +14% from 2024-2026; GCP monthly token volume up about 320x+The report uses this comparison to show that bandwidth improvement is lagging AI data demand growth.
• Emerging memory technology TAM, excluding HBMUS$23.0bn base case by 2030；bull/bear为US$41.4bn/US$16.9bnThe report says this is mainly driven by packaging and peripherals.
• Emerging memory technology TAM, including HBMUS$276bn base case by 2030；bull/bear为US$342bn/US$160bnHBM significantly enlarges the overall market size.
• Micron target price$1,200The report maintains Overweight, using a 30x valuation on through-cycle earnings power of $40.
• SanDisk target price$1,750The report maintains Overweight, using a 28x valuation on through-cycle EPS of $62.50.
• Montage Technology target price revision6809.HK: HK$310.00至HK$432.00；688008.SS: RMB274.00至RMB377.00The report views it as a key beneficiary of next-generation memory interfaces, CXL, MRDIMM, and DDR6 expansion.

Impact & implications

For investors, the main AI investment theme may expand from "who has the fastest GPU" to "who can move, store, and access data most efficiently." This means traditional memory leaders remain direct beneficiaries, but larger differentiated opportunities may lie in memory interfaces, interconnects, advanced packaging, system-level co-design, materials, and equipment. If the market begins pricing in the system-level impact of the memory wall, related ecosystem companies could see valuation re-rating.

Risks

• Software-layer optimization could reduce demand for incremental hardware intensity.
• If AI commercialization or monetization falls short of expectations, AI infrastructure and memory demand could slow.
• Many next-generation memory technologies are complex and not yet fully validated at scale, creating execution and commercialization risk.
• YoY momentum in memory prices could plateau periodically, affecting near-term cyclical catalysts.
• Supply discipline, capital expenditure, and the delivery of long-term customer agreements could alter the earnings upside path.

What to watch

• Adoption progress of next-generation memory standards.
• Launches of new connectivity solutions such as CXL and customer validation.
• Commercialization evidence for advanced architectures such as 3D DRAM, zHBM, MRDIMM, DDR6 peripherals, and PIM/CIM.
• Early cases of system-level deployment, especially in AI servers, cloud data centers, and edge AI scenarios.
• Price trends in DRAM, NAND, and HBM, as well as changes in memory's share of cloud capex.
• Earnings revisions and target price changes for core names such as Samsung, Kioxia, Micron, SanDisk, and Montage.