Electrostatic RAM vs DRAM: In the era of AI computing power, capacitors are retracing the path of memory value reassessment
```
There is a huge gap in market expectations regarding AI hardware: capital is fixated on GPU shipments, but overlooks the "energy wall" hidden behind the explosive growth in computing power.
On June 20, Guojin Securities pointed out in its latest research report that capacitors are becoming the "power RAM" of AI computing systems—just as HBM serves as a buffer for data, capacitors act as a buffer for energy. Both are highly analogous in terms of system roles, tiered architectures, and price-volume drivers. More importantly, market calculations of capacitor demand have long remained in the "white zone" (inside PSU), while the real system-level incremental demand occurs massively in the overlooked "gray zone" (the multi-stage voltage reduction and power distribution between the grid and the cabinet).
The report emphasizes that the elasticity of capacitor demand has completely broken free from the old logic of "linear extrapolation based on GPU count." As AI cabinet power moves toward the megawatt level and power-supply architectures upgrade to 800V high-voltage DC, the capacitor industry is ushering in a reevaluation with both volume and price rising. On the supply side, faced with bottlenecks in high-end materials (aluminum foil, activated carbon), high energy consumption and environmental restrictions, and overseas industry giants adopting only 10%-20% conservative expansion strategies, domestic capacitor manufacturers with high self-sufficiency in upstream materials and strong engineering capabilities are seizing a historic window to capture global market share and pricing power.
This means the demand ceiling for capacitors is systematically underestimated, and domestic manufacturers with high upstream material self-sufficiency may benefit from both share expansion and increased pricing power.
Core Analogy: Capacitors are not just a rhetorical metaphor, but are physically the "power RAM"
Capacitors and DRAM are highly analogous in system roles, tiered structures, and price-volume drivers, which determines the inevitability of exploding demand:
Instant buffering analogy: DRAM provides buffering for computation through cyclical "refresh"; capacitors provide energy buffering for instantaneous GPU load through "charging and discharging." The stronger the computing power, the greater the dependency on "fast and close" power buffering, and insufficient past solutions will trigger supplementary demand growth in new-generation platforms.Tiered structure analogy: DRAM has a caching hierarchy of "register-SRAM-DRAM"; capacitors have a ladder network tiered by "time constant." The closer to the compute core, the faster the response, the higher the unit value, and the steeper the technology barriers.Demand-driven analogy: DRAM faces a "capacity wall," capacitors face a "power wall." Capacitor demand is collectively driven by "GPU number × per-card power/value × system complexity × price factor," with elasticity significantly higher than GPU shipment growth.
Time Constant Ladder: Six Types of Capacitors, Each in their Place, Build a Panoramic Buffer Network
AI power upgrades do not come from a single category scaling up, but from the synchronous densification of the entire buffer network. Along the path from chip to power grid, six types of capacitors form a tight defense line:
Silicon Capacitor (sub-nanosecond, inside package): Analogous to "register." Thickness can be reduced to 100㎛ or less, positioned closest to the die to suppress highest-frequency voltage disturbances; highest technical barrier and unit value density.
MLCC (nanosecond, board-level): Analogous to "SRAM cache." Greatest quantity, densely distributed around GPU, undertaking high-frequency decoupling.
MLPC (microsecond, board-level): Intermediate layer between MLCC and large electrolytic capacitors. In high-end scenarios such as 125℃ high temperature and 4000-hour lifespan, can replace multiple MLCCs in high capacitance range, offering a window for domestic manufacturers to substitute.
Screw-type aluminum electrolytic capacitor (microsecond to millisecond, power module): Analogous to "DRAM main memory." Direct beneficiary of AI power supply voltage increase (from 5.5kW to 18.5kW). To cope with high-frequency power fluctuations, capacity demand is at least 1.5 times traditional levels (e.g., from 1000μF to 1500μF), with single-unit value increasing by an order of magnitude.
Supercapacitor and lithium-ion capacitor (milliseconds to seconds, cabinet side): Undertaking backup power and power buffering. Since batteries lose efficiency and decay significantly in long-term high-frequency charge/discharge, supercapacitors are becoming a "must have" from "optional."
Film capacitor (high-voltage DC bus): As power architectures move to 800V HVDC and solid-state transformers (SST), film capacitors absorb high-voltage ripple on the bus side, indispensable in high-voltage architectures.
White and Gray Zone: Market Systematically Underestimates Real Demand
This is the most investment-worthy cognitive gap in the report.
Source of the cognitive gap: Most demand calculations currently focus on PSU internals (the "white zone"). There is a great disparity in market understanding of capacitor configurations under PSU, white zone, micro zone, Power Rack, HVDC, SST, 400V, 800V, and other architectures, and calculation standards may differ by several to ten times. Even within power supply development companies, how to solve AI power fluctuation is still under discussion.
Where is the gray zone: Along the power supply path from grid to cabinet, there are steps including grid access, multi-stage voltage reduction, gray zone distribution, local capacitor boxes, and Power Rack. As cabinet power moves toward the megawatt level and architectures shift to 800V HVDC and SST, large quantities of capacitors are also needed in these "gray zone" stages (multi-stage voltage reduction, power distribution, capacitor boxes) for voltage stabilization and buffering. Some gray zone consumption has previously not been fully counted, but is already in use.
Meaning of system-level expansion: Looking at both white and gray zones together, the expansion of capacitor demand brought by AI computing is system-level—covering both GPU board-level and power modules (white zone densification), as well as data center power supply, distribution, and redundancy (gray zone rollout). Calculating only by the white zone will systematically underestimate real demand.
Logic of Volume and Price Increase: Material Constraints, Expansion Barriers, and Window for Domestic Substitution
Logic of volume: Demand elasticity significantly exceeds GPU shipment growth.
GPU shipments maintain a high annual growth rate. Single GPU power and power supply complexity lift per-card value, combined with redundancy and structural coefficients, overall demand elasticity is significantly higher than GPU count growth. New-generation platforms (VR platforms) require higher power consumption, system complexity, and power-supply stability. Some customers with relatively weaker system optimization ability are inclined to increase capacitor redundancy, further amplifying usage.
Logic of price: Two systems operate in parallel.
Traditional aluminum electrolytic capacitors are affected by material cost, electricity price, and environmental factors, entering a price-rise channel in the second half of the year. Major Japanese manufacturers have issued letter notices of price adjustments; high-end AI-related new products are not merely "price increases," but "repricing"—with higher requirements from customers for voltage, capacitance, lifespan, and reliability, and product specification changes, pricing should be based on new products. In AI server systems, the absolute amount of capacitors is very small compared to the whole machine price, yet they are extremely critical for power-supply stability. Customers value supply assurance and reliability over raw price, giving high-end capacitor products stronger bargaining power.
Constraint: High-end materials are the real bottleneck.
AI servers' high-voltage, high-specific-capacity electrode foil (formed foil) supply is tight; the core bottleneck for supercapacitors is upstream activated carbon material, and domestic substitution will require time to nurture; MLPC is mostly stuck on process and material system understanding. Material bottlenecks mean "seeing demand" does not equal "supply can keep up."
Threshold: Expansion is a comprehensive hurdle.
Electrode foil's corrosion and forming steps are high energy-consuming, highly sensitive to electricity price and energy supply; stable, low-cost power (such as direct green electricity) is becoming critical for front-end capacity rollout. At the same time, new domestic production capacities require obtaining capacitor capacity indices in some regions, with environmental and energy consumption indicators as actual thresholds.
Window: Overseas expansion is conservative; domestic manufacturers are seizing share and pricing power.
Japanese MLCC and passive component manufacturers are culturally conservative toward capacity expansion, generally planning for 10%-20% growth, with rapid doubling expansion being very difficult. By contrast, domestic manufacturers with high self-sufficiency in upstream materials, strong engineering construction, and capital expenditure efficiency have stronger supply assurance and expansion capabilities during this tight supply-demand window, thereby seizing market share and gaining pricing power.
Risk Warning and DisclaimerThe market involves risks; investment must be cautious. This article does not constitute personal investment advice, nor does it take into account individual users' specific investment objectives, financial status, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their own situation. Invest accordingly, and bear your own responsibility. ```