AI Server Single Rack Requires 600K MLCCs, Capacitor Costs Rise to Third-Largest Component
nashnova research
MLCCs have become the third-largest cost item in AI servers, behind only GPUs and memory chips. Goldman Sachs says the server MLCC segment is growing at 80% CAGR, yet industry capacity can expand by only ~10% a year — making capacitors potentially the AI component with the longest pricing upside runway.
A tiny capacitor — why is it suddenly the third-biggest AI server cost?
MLCCs — multilayer ceramic capacitors, ultra-small energy-storage components that respond in microseconds — now rank behind only GPUs and memory chips in server cost.
This means → a component once dismissed as "fractions of a cent apiece" is being repriced by AI compute demand.
A single advanced AI server rack needs up to 600,000 MLCCs placed right next to the chips. In plain terms = AI chips draw power in violent microsecond spikes that the main supply cannot match instantly; MLCCs act as tiny fast-discharge batteries glued beside the chip — they fill the power gap and filter electrical noise to prevent data corruption.
Demand growing at 80%, capacity at 10% — how big is the gap?
Goldman Sachs puts the total MLCC market at roughly $15 billion; the server segment accounts for about $1.3 billion and is expanding at an 80% CAGR.
But industry capacity can grow at most ~10% per year — equipment and critical materials must be built in-house, constrained by internal engineering resources.
This means → if AI servers keep absorbing new capacity, the supply-demand squeeze may not be a one- or two-quarter event but a structural gap lasting years.
Meanwhile, traditional demand from autos, smartphones, and PCs is softening — the MLCC industry is shifting from broad-based growth to an "AI eats all the incremental output" pattern.
MLCCs have barely risen in price — what does that actually tell us?
Goldman notes that memory (DRAM, NAND), ABF substrates, and copper-clad laminates (CCL) have all already repriced. MLCC is one of the last categories in the pricing cycle.
This means → its price upside runway is the longest among all AI components — the later the move starts, the larger the elastic potential.
In plain terms = the current "no price hike" is not a sign of slack supply; it means the pricing transmission chain has not yet reached MLCCs. Once it does, the upcycle could last longer than earlier movers.
Who captures this wave — three oligopolists, and TDK on a different track?
In the low-voltage, high-capacitance MLCCs needed around GPUs and ASICs, Murata, Samsung Electro-Mechanics (SEMCO), and Taiyo Yuden form an oligopoly and stand to benefit directly from both volume and price gains.
TDK currently lacks the technology to enter this segment; it is waiting for a joint materials R&D program with Japan Chemical Industry to deliver results.
TDK is, however, seeing strong orders for high-voltage, high-capacitance MLCCs used in power circuits — products that overlap heavily with EV automotive technology, potentially lifting factory utilization.
This reflects a key nuance: the MLCC market is not monolithic — low-voltage and high-voltage are two distinct technology tracks with very different competitive landscapes.
What to watch next — can MLCCs replicate the memory-chip pricing story?
Goldman estimates AI server demand will grow roughly 4.3× from FY2025 to FY2030.
Smartphone and PC clients, despite falling shipment volumes, have already begun seeking long-term MLCC contracts, further tightening available capacity.
In plain terms = whether MLCCs trace a memory-chip-style pricing curve comes down to a race: the speed at which AI demand absorbs new capacity vs. the pace at which the industry can expand. Right now, the demand side is running faster.
市场有风险，内容仅供研究参考，不构成投资建议。