The central AI investment theme is shifting from computing power to memory-system constraints

Overview

This report conducts thematic research on next-generation memory. Its core view is that AI has entered a new phase in which the constraint is shifting from computing expansion to system efficiency. As generative AI and agentic AI drive demand for tokens, context windows, and data access, memory capacity, bandwidth, and cost are becoming key bottlenecks to AI scalability.

Core views

The report argues that the “memory wall” is not simply a supply issue, nor can it be solved merely by adding DRAM or HBM capacity. Instead, it is a full-stack challenge involving data storage, movement, access, packaging, interconnects, materials, and system co-design. Investment opportunities therefore extend beyond traditional memory-chip manufacturers to interfaces, connectivity, advanced packaging, system design, and materials and equipment across the broader ecosystem.

Analysis framework

The research demonstrates the memory bottleneck through the mismatch between AI demand growth, cloud capital-expenditure structure, memory prices, bandwidth improvements, and token growth. It also draws on the disruptive impact of 3D NAND on traditional storage architectures and supply chains, developing a six-path framework to assess the commercialization potential and investment implications of next-generation memory technologies.

Methodology notes

• Thematic investment frameworkSix-path memory innovation framework

  Design, process, packaging, peripherals, integration, and materials

  This framework expands memory innovation beyond traditional process nodes to system-level innovation and is used to assess which technology pathways have scalability and commercial relevance.

• Industry constraint analysisMemory wall

  Computing power improves faster than memory capacity, bandwidth, and cost, constraining system performance

  The report uses this concept to explain that even with stronger computing capabilities, AI models remain constrained by data-supply speed, capacity, and economics.

• Historical analogy3D NAND disruption case study

  3D NAND overcame the physical scaling limits of 2D planar NAND through vertical stacking

  The report uses the industry adoption of 3D NAND to illustrate how major memory innovations can reshape cost curves, end-market applications, and supply-chain structures.

Asset mapping & comparison

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

• Samsung Electronics

  Traditional memory leader and AI-memory beneficiary

  Strengths

  A relatively complete positioning across HBM mass production, 3D DRAM, thousand-layer NAND, hybrid copper bonding, CXL, and customized memory.

  Weaknesses

  In the near term, it remains exposed to changes in the memory cycle and supply-demand visibility.

  Comparison

  Compared with earlier-stage ecosystem companies, Samsung is more of a scaled mainstream-memory beneficiary.

  Risks

  If year-over-year price growth plateaus by the end of 2026, near-term catalysts may be insufficient.

• Kioxia

  Beneficiary of AI SSD and NAND systems

  Strengths

  Its CM, GP, and LC series address bandwidth, latency, IOPS, and capacity requirements, with some products able to alleviate HBM capacity constraints.

  Weaknesses

  It is relatively dependent on the realization of AI storage demand and the pace of product commercialization.

  Comparison

  Compared with DRAM/HBM manufacturers, Kioxia places greater emphasis on NAND and SSD moving up the AI data hierarchy.

  Risks

  If hyperscaler adoption of storage architectures is slower than expected, earnings upside may be delayed.

• Micron

  Beneficiary of DRAM, HBM, and AI data-center demand

  Strengths

  The report believes its HBM execution is underestimated and that AI data-center demand could drive a multiyear earnings upcycle.

  Weaknesses

  Valuation expansion depends on the sustainability of HBM share, gross margins, and long-term supply-demand tightness.

  Comparison

  Like Samsung, it offers core exposure to the mainstream memory cycle and AI demand.

  Risks

  If supply-demand tightness eases or customer agreements are realized below expectations, earnings revisions may slow.

• SanDisk

  Beneficiary of AI NAND and high-capacity storage

  Strengths

  The report believes inference demand will push NAND into higher memory tiers for LLM KV-cache and context-window storage.

  Weaknesses

  It depends on a structural shift toward cloud becoming the largest end market for NAND.

  Comparison

  Compared with the traditional PC/mobile NAND cycle, SanDisk’s investment thesis relies more on hyperscaler demand and new commercial agreements.

  Risks

  If AI inference storage demand falls short of expectations, structural re-rating potential may be limited.

• Montage Technology

  Beneficiary of memory interfaces, connectivity, and expanding AI-server content

  Strengths

  It has exposure to the “plumbing” of next-generation memory systems, including CXL, MRDIMM, and DDR6.

  Weaknesses

  Some technologies remain in the early-adoption phase, making the commercialization pace uncertain.

  Comparison

  Compared with traditional memory manufacturers, Montage represents more differentiated ecosystem- and interface-side alpha.

  Risks

  Standards implementation, customer qualification, and system-deployment progress may fall below expectations.

• GigaDevice, Winbond, APMemory

  Beneficiaries of the advanced-packaging, 3D-stacking, and edge-AI specialized-memory ecosystem

  Strengths

  They may benefit from wafer-on-wafer stacking, specialized memory applications, bandwidth improvements, power-efficiency gains, and form-factor optimization.

  Weaknesses

  Scalability and customer adoption still require validation.

  Comparison

  They are more early-stage and niche-application focused than mainstream DRAM/NAND manufacturers.

  Risks

  Technical mass-production challenges, costs, and end-market demand remain uncertain.

• Renesas, Marvell

  Names related to the next-generation interconnect and connectivity ecosystem

  Strengths

  The report lists them as examples related to next-generation interconnects.

  Weaknesses

  The report excerpt does not provide detailed financial valuation or standalone rating analysis.

  Comparison

  They belong to the system-connectivity layer of memory-bottleneck solutions rather than the storage-chip core itself.

  Risks

  Competition among interconnect standards and changes in system-architecture roadmaps may affect the degree of benefit.

Key data

• Cloud memory spending forecastUS$418bn by 2030Based on current forecasts, implying approximately 8% CAGR from 2026.
• Agentic AI contribution to incremental DRAM demand26-77% by 2030The report believes agentic AI will materially increase memory demand.
• Memory as a share of cloud capital expenditure40% in 2027e vs 12% in 2023Shows that memory is shifting from a peripheral cost item to a core AI infrastructure expenditure.
• DDR5 single-channel bandwidth improvement+14%, from 44.8GB/s in 2024 to 51.2GB/s in 2026Bandwidth improvements are materially lagging token growth.
• GCP token growth10T/month in Apr 2024 to 3,200T in Jun 2026The report uses token growth of more than 320x to illustrate system-level pressure.
• TAM of emerging memory technologies, excluding HBMUS$23.0bn base case by 2030; bull/bear US$41.4bn/US$16.9bnDriven primarily by packaging- and peripheral-related pathways.
• TAM of emerging memory technologies, including HBMUS$276bn base case by 2030; bull/bear US$342bn/US$160bnIncluding HBM significantly expands the potential market size.
• Micron target priceUS$1,200The report maintains Overweight, valuing the company at 30x through-cycle earnings power of US$40.
• SanDisk target priceUS$1,750The report maintains Overweight, valuing the company at 28x through-cycle EPS of US$62.50.
• Montage Technology target-price adjustment6809.HK: HK$310 to HK$432; 688008.SS: Rmb274 to Rmb377Reflects opportunities from the expansion of CXL, MRDIMM, DDR6, and AI-server content.

Impact & implications

The investment implication is that the AI semiconductor theme may shift from “who owns the fastest GPU” to “who can move, store, and access data most efficiently.” Traditional memory leaders remain direct beneficiaries, but more differentiated alpha may come from underpriced areas such as advanced packaging, interfaces, interconnects, CXL expansion, MRDIMM bandwidth improvements, PIM/CIM architectures, and specialized storage solutions.

Risks

• Software-layer optimization may reduce the need for incremental hardware intensity.
• A slowdown in AI demand due to weaker-than-expected commercialization or monetization would undermine memory-demand assumptions.
• Next-generation memory technologies are complex, and many solutions have not yet completed large-scale validation, creating execution risk.
• Memory stocks remain subject in the near term to pricing momentum, supply discipline, and cyclical catalysts.
• The standardization and customer-adoption pace of ecosystem technologies such as CXL, MRDIMM, DDR6, PIM/CIM, and advanced packaging remain uncertain.

What to watch

• Adoption progress for next-generation memory standards.
• Launches and customer qualification of new connectivity solutions such as CXL.
• Commercialization evidence for architectures including 3D DRAM, advanced HBM, wafer-on-wafer, MRDIMM, and DDR6.
• Whether system-level AI deployments show that memory bottlenecks are being alleviated.
• Price trends, supply-demand tightness, and changes in cloud capital-expenditure structure for DRAM, NAND, and HBM.
• Earnings revisions and order visibility for key names including Samsung, Kioxia, Micron, SanDisk, and Montage Technology.