Report Board
High-End MLCC vs HBM: AI Hardware Demand, Bottlenecks, and Expansion Difficulty
A demand-to-bottleneck report comparing high-end MLCC and HBM across AI server demand, physical constraints, expansion economics, and public-equity transmission.
Memory is a price spike; high-end MLCC is a slower inflection
Official PPI series show that electronic and storage-related price indexes are rising in 2026, while broad capacitor PPI remains much more moderate. TrendForce's public price shocks show memory moving vertically, while high-end MLCC is transitioning from narrower declines into mild price increases.
DDR5 spot
TrendForce said DDR5 2Gx8 spot prices rose 307% from the start of September 2025 to Nov. 18, while DDR4 1Gx8 rose 158%. This is the clearest memory price spike.
DRAM contract
TrendForce upgraded 1Q26 conventional DRAM contract prices to +90%-95% QoQ, NAND contract prices to +55%-60%, and enterprise SSD prices to +53%-58%. The spike moved from spot into contract.
Official storage PPI
Computer storage device manufacturing PPI rose from 60.998 in Jan 2025 to 73.772 in May 2026. The official series confirms direction, but does not replace DRAM/HBM spot quotes.
Capacitor PPI
Capacitors for electronic circuitry PPI rose from 284.385 in Jan 2025 to 306.974 in May 2026, up 7.9%; in 2026 YTD it rose only 1.4%, far from the memory spike shape.
MLCC turning point
TrendForce said Taiyo Yuden raised prices on some low-capacitance consumer/auto MLCCs by 6%-13% in April 2026 and that the overall MLCC average price decline narrowed to less than 0.5%. The real test is whether high-end MLCC turns into sustained ASP/mix improvement in 2H26.
Bottom line
High-end MLCC is not an HBM substitute and not an EMIB/CoWoS substitute. HBM is a memory bandwidth bottleneck close to the accelerator. MLCC is a power-integrity and reliability bottleneck. Both benefit from AI server complexity, but their value density, expansion economics and earnings transmission are very different.
Which bottleneck is harder
HBM is harder if the standard is near-term earnings torque. A shortage directly limits accelerator availability and customer schedules. High-end MLCC is also hard, but it is more dispersed and should not be modeled as an HBM-style revenue ramp.
Expansion comparison
High-end MLCC expansion depends on ceramic powder, thin dielectric layers, multilayer stacking, nickel-electrode co-firing, defect control and qualification. HBM expansion depends on advanced DRAM, TSV, stacking, thermal management, packaging yield and customer validation. MLCC projects are usually much smaller in dollar terms; HBM can require tens of billions of dollars across fabs and packaging.
How to read the price-cycle chart
The update uses two data layers. The first layer is official FRED/BLS monthly PPI data: computer storage device manufacturing rose 20.9% from Jan 2025 to May 2026, semiconductor and other electronic component manufacturing rose 25.4%, and capacitors for electronic circuitry rose 7.9%. The second layer is TrendForce's public industry shock data: DDR5 spot prices rose 307% from early September 2025 to Nov. 18, 2025; 1Q26 conventional DRAM contract prices were expected to rise 90%-95% QoQ; Taiyo Yuden's April MLCC price action was 6%-13%, while the overall MLCC average decline narrowed to less than 0.5%.
Price-cycle conclusion
Memory is already a spike cycle: spot jumped first, contract prices followed, and official PPI is now rising with a lag. High-end MLCC is not in that shape yet. Broad capacitor PPI is still only moderately higher, while TrendForce frames MLCC as moving from narrower deflation toward mild 2H26 price increases. For equities, memory is about profit leverage and cycle-top risk; MLCC is about proving that supplier bargaining power reaches the income statement.
Core names: bottleneck hardness, transmission and risk
This is a watchlist and verification framework, not investment advice. Ranking reflects bottleneck hardness and transmission, not current valuation.
| Layer | Core names | Purity / torque | Transmission | Bottleneck hardness | Main risk |
|---|---|---|---|---|---|
| HBM / advanced DRAM | SK hynix / MU / Samsung Electronics | High: SK hynix leader, MU U.S. high-beta, Samsung catch-up | HBM4/HBM4E qualification, longer sold-out windows, ASP and gross-margin upgrades. | Highest | Crowded expectations, customer qualification, pricing, yield and capex-cycle volatility. |
| High-end MLCC | Murata / Samsung Electro-Mechanics / Taiyo Yuden / TDK | Medium-high: AI server, automotive and high-performance power delivery raise specifications | High-end book-to-bill, price increases, AI server part disclosure and mix improvement. | Medium-high | Disclosure is not granular enough, and weak consumer MLCC demand can blur the AI signal. |
| U.S.-tradable proxies | MRAAY / MU / SSNLF / VSH | Medium-low: MRAAY is Murata ADR, MU is HBM, VSH is not a clean MLCC pure play | Use MU for AI memory, MRAAY for Murata exposure, and VSH only as a diversified passive proxy. | Imperfect | ADR liquidity, OTC constraints and weak purity. |
| Packaging companions | TSM / Ibiden / Ajinomoto / ASE / AMKR | Medium-high: HBM and AI ASICs need packaging and substrate capacity | CoWoS, substrate, interposer and test capacity remain tight. | Medium-high | Some names already price in the AI packaging shortage. |
AI hardware supply-chain report · Demand, bottleneck, capex and equity transmission · 2026-06-15
Earnings releases, announcements, filings, estimate tables, and reviewable sources.
- Core signal
- SIA/BCG frames semiconductors as more than 95% of AI data server rack value. TrendForce points to MLCC market polarization and high-end support from AI servers and automotive demand. Murata, Taiyo Yuden and Samsung Electro-Mechanics are expanding high-end MLCC, while Micron, SK hynix and Samsung are pushing HBM4/HBM4E as the 2026-2028 AI memory core.
- Current read
- The short answer: HBM has stronger certainty and earnings torque, but more crowded expectations. High-end MLCC is a real bottleneck, but it should be underwritten as high-end mix, pricing and utilization improvement rather than an HBM-like revenue explosion.
- Next question
- When AI server demand and physical bottlenecks are analyzed together, which layer is harder, which layer is harder to expand, and which equities receive the clearest transmission?
HBM is the higher-value bottleneck and has the clearer earnings-revision path.
High-end MLCC is real, but it should be modeled as mix, pricing and utilization, not an HBM-style revenue explosion.
On price cycles, memory is already spot spike, contract follow-through and lagged official PPI; high-end MLCC remains a slower turn from narrower deflation toward mild inflation.
Murata is likely one of the strongest broad high-end MLCC leaders; Samsung Electro-Mechanics is also core in AI server and automotive MLCC.
Expansion difficulty roughly ranks as low-end MLCC < high-end MLCC < HBM/advanced DRAM < full advanced-packaging delivery.
U.S.-listed expressions are imperfect: MU is HBM, MRAAY is Murata ADR, and VSH is not a clean high-end MLCC pure play.
Price cycle dashboard: official PPI shows storage/semiconductor indexes rising materially, while capacitor PPI remains moderate YTD; if capacitor PPI accelerates later, MLCC repricing may be spreading from high-end parts to the broad index.
MLCC price / book-to-bill: track high-end products, not only commodity consumer MLCC.
Murata / SEMCO / Taiyo Yuden margin mix: a true high-end shortage should show up in utilization, ASP, mix and order visibility.
HBM4 qualification: track SK hynix, Micron and Samsung HBM4/HBM4E qualification with NVIDIA and ASIC customers.
Memory capex and packaging: HBM supply depends on wafer starts, TSV, stacking, advanced packaging, substrates and CoWoS.
AI server teardown: MLCC supplier proof from Rubin or AI ASIC board teardowns would make the thesis much more concrete.