Buzzberg Cup Live

Memory-on-Logic: Re-Rating Memory Beyond the Cycle

Damnang · Damnang’s Substack · July 09, 2026 at 10:12 · ⏱ 14 min read  | Read on Substack ↗
Summary
The memory industry is undergoing a structural shift from commodity DRAM/HBM to custom silicon, driven by HBM base dies moving to logic processes (HBM4), customer logic entering the base die (cHBM), and eventually memory stacked directly on compute dies (memory-on-logic). This transition makes memory companies custom silicon partners rather than commodity suppliers, which should structurally improve revenue and margin profiles—yet the market still prices them as cyclical commodity stocks, creating upside that the author argues is not yet reflected in valuations.
  • •From HBM4, the HBM base die moves from a DRAM process to a logic process: SK hynix uses TSMC 12nm, Samsung uses its own 4nm, Micron stays on DRAM through HBM4 but plans to move to TSMC from HBM4E.
  • •cHBM (custom HBM) departs from JEDEC standards by embedding customer-specific circuitry into the base die, making the part non-interchangeable and shifting the business model from commodity to custom silicon.
  • •Memory-on-logic stacks DRAM directly on top of the compute die, eliminating the interposer; a 2025 Georgia Tech/SK hynix paper modeled 64x throughput and 3x energy efficiency vs. 2.5D HBM.
  • •Qualcomm’s HBC (High Bandwidth Compute), unveiled June 2026, uses LPDDR stacked on a dedicated accelerator die in a memory-on-logic structure, slated for mid-2027 launch with customer deployments from Humain and Microsoft.
  • •All three memory makers (SK hynix, Samsung, Micron) have placed custom HBM on official roadmaps, and Marvell is co-developing custom HBM architectures with all three.
  • •The article identifies heat as the largest unsolved problem for memory-on-logic: DRAM’s temperature sensitivity conflicts with placing it atop a high-power compute die.
  • •The progression from standard HBM → cHBM → memory-on-logic is not linear; cHBM is the near-term commercialization path, while memory-on-logic will coexist once thermal and yield issues are resolved.
  • •The author argues the market still values memory stocks on cyclical bit growth and ASP, ignoring the structural shift toward higher-margin custom silicon revenue.
Read time 14 min
Length 14,977 chars
Category finance
Ideas
Damnang Substack author, Damnang’s Substack
Micron is one of the three memory makers moving toward custom HBM; though it keeps the base die on DRAM through HBM4, it plans to move to TSMC logic process from HBM4E. The article implies all three w
Micron is one of the three memory makers moving toward custom HBM; though it keeps the base die on DRAM through HBM4, it plans to move to TSMC logic process from HBM4E. The article implies all three will benefit from the structural shift away from commodity pricing. Micron is publicly traded and directly exposed to the HBM custom transition. Risk: Micron's decision to stay DRAM-based for HBM4 could leave it behind on base die performance vs. SK hynix/Samsung; execution on custom HBM roadmap is unproven.
MU
Damnang Substack author, Damnang’s Substack
TSMC is the logic foundry for SK hynix's HBM4 base die (12nm) and the likely partner for Micron from HBM4E onward. The article also mentions TSMC's wafer-bonding technology being discussed for memory-
TSMC is the logic foundry for SK hynix's HBM4 base die (12nm) and the likely partner for Micron from HBM4E onward. The article also mentions TSMC's wafer-bonding technology being discussed for memory-on-logic with SK hynix and NVIDIA. As the industry moves base dies to logic processes, TSMC captures foundry revenue for those dies and potentially for packaging services. Risk: HBM base die wafer volumes are relatively small compared to TSMC's overall business; the impact may be marginal unless memory-on-logic drives much larger die area demand.
TSM
Damnang Substack author, Damnang’s Substack
Qualcomm's HBC (memory-on-logic using LPDDR) is the first disclosed product with a production timeline (mid-2027). The article notes customer interest from Humain and Microsoft, and claims 6x bandwidt
Qualcomm's HBC (memory-on-logic using LPDDR) is the first disclosed product with a production timeline (mid-2027). The article notes customer interest from Humain and Microsoft, and claims 6x bandwidth per watt vs. HBM. If successful, Qualcomm gains a differentiated AI accelerator that bypasses HBM and CoWoS bottlenecks. Risk: HBC is still a year from launch; performance claims need hardware validation. Competing with NVIDIA's established ecosystem is a high hurdle.
QCOM
Damnang Substack author, Damnang’s Substack
NVIDIA is cited as the key customer for cHBM (Feynman generation 2028) and as a partner in discussions about memory-on-logic with SK hynix. The structural shift toward custom HBM and memory-on-logic b
NVIDIA is cited as the key customer for cHBM (Feynman generation 2028) and as a partner in discussions about memory-on-logic with SK hynix. The structural shift toward custom HBM and memory-on-logic benefits NVIDIA by delivering higher bandwidth and energy efficiency for its AI accelerators, potentially reducing reliance on standard HBM supply and packaging constraints. Risk: Dependence on custom silicon from memory suppliers introduces new supply chain complexity; NVIDIA's own base die design (if pursued) increases R&D burden.
NVDA
More from Damnang’s Substack
A Complete Guide to NVIDIA’s Racks, and Then SemiAnalysis
Jul 07, 11:13
The Memory Crisis: Can CXL Be the Answer?
Jul 06, 02:55
Can VCSEL CPO Shake Up the InP Optical Order?
Jul 05, 00:00
Meta Compute Misread: Where the AI Infrastructure Opportunity Really Is
Jul 02, 09:00
The Inference Chip Architecture Map: 12 Companies Sorted by Three Bottlenecks
Jul 01, 07:55

This newsletter, published July 09, 2026, features Damnang discussing MU, TSM, QCOM, NVDA. 4 trade ideas extracted by AI with direction and confidence scoring.

Speakers: Damnang  · Tickers: MU, TSM, QCOM, NVDA