Multilayer Ceramic Capacitors (MLCCs) are manufactured by stacking ceramic dielectric sheets with printed internal electrodes in a staggered pattern, followed by a single high-temperature co-firing process to form a ceramic chip. Metal layers (external electrodes) are then applied to both ends of the chip, creating a structure similar to a monolithic block-hence the name "monolithic capacitor."
MLCCs offer a wide capacitance range, excellent frequency characteristics, high temperature and voltage resistance, compact size, long service life, and low cost. They are primarily used in oscillation, coupling, filtering, and bypass circuits in electronic systems. Today, MLCCs have become the most dominant and most widely used passive surface-mount components in electronic devices, finding broad applications in automobiles, smartphones, computers, consumer electronics, home appliances, and other communication equipment and products.
AI servers consist of numerous graphics processors working in parallel. As these chips switch on and off, instantaneous currents fluctuate sharply, causing voltage instability. Moreover, newer graphics cards have ever-increasing power consumption. To stabilize voltage, provide instantaneous charge supplementation, and filter out current noise, dense arrays of MLCCs must be placed around the GPUs and each power supply module. Additionally, AI servers employ a multi-stage step-down power delivery architecture. The current mainstream GB200 server platform uses a multi-level voltage conversion scheme: mains power is first converted by the data center's AC-DC rectifier to a 48 V DC bus voltage in the rack, then stepped down to 12 V via onboard power modules, and finally reduced to the ultra-low 0.8 V operating voltage required by the GPU cores through the core voltage regulator circuit. This multi-level conversion creates complex power supply conditions with significant voltage and current fluctuations. As the next-generation Rubin architecture gradually adopts 800 V high-voltage rack solutions, the number of voltage conversion stages may further increase. Each power conversion point requires corresponding capacitors, directly driving a surge in MLCC demand.
MLCC ceramic powders are the core dielectric raw materials for MLCCs-especially those used in AI servers. Their physicochemical properties directly determine whether the products can withstand the high-temperature, high-frequency harsh conditions of computing hardware, and they serve as the fundamental technological enabler for achieving miniaturization, high capacitance, low ESR, and wide-temperature stability in MLCCs. Ceramic powders are divided into base powders and formulated powders. Base powders use high-purity nano-barium titanate as the main body of the ceramic dielectric. Formulated powders incorporate dopants such as rare-earth elements (yttrium, holmium, dysprosium, etc.) to ensure insulation, as well as additives like magnesium, manganese, vanadium, chromium, molybdenum, and tungsten to guarantee temperature stability and reliability. These additives must be uniformly distributed with the barium titanate powder to control the microstructure and electrical characteristics of the dielectric ceramic material during sintering.
MLCCs used in AI servers are held to much stricter standards than those for ordinary consumer electronics like smartphones and home appliances. This imposes a dual driving force on ceramic powders:
First, finer particle size requirements drive ceramic powders toward nanoscale, high-purity upgrades. Powders must be made finer, pushing manufacturers toward nanoscale and high-purity directions. The dielectric layers in AI-server MLCCs are extremely thin; currently, mainstream particle sizes are controlled around 100 nanometers, whereas powders for general consumer electronics are mostly above 200 nanometers. To produce such ultra-fine powders, manufacturers must thoroughly master the nano-barium titanate fabrication process, precisely control particle size uniformity, and ensure good powder dispersion; otherwise, fine particles tend to agglomerate, introducing defects inside the capacitor that directly impair performance.
Second, a multiplicative increase in layer counts drives a sharp rise in ceramic powder consumption and its cost share. For high-capacitance MLCCs used in AI applications, the external case size remains unchanged. To pack more capacitance, more ceramic layers must be stacked. MLCCs in ordinary consumer electronics typically have 50 to 100 layers, while high-capacitance versions for AI servers commonly stack 500 to 1,000 layers, with top-tier products reaching 1,300–1,600 layers. Each layer's raw material is formulated from barium titanate powder. As the layer count increases severalfold, the ceramic powder required per capacitor grows substantially. As a result, ceramic powder now accounts for approximately 40% of the total cost of such high-capacitance capacitors, making it the largest single raw-material cost component.