Are high-voltage, high-capacitance MLCCs ready to take off? The AI computing power revolution is opening up a new round of growth opportunities for high-end power components.

Last week, after visiting leading Japanese MLCC manufacturers and Taiwanese niche suppliers, it was evident that the market sentiment and industry structure are becoming increasingly optimistic, and the industry position of high-voltage, high-capacitance MLCCs has clearly moved up. In the past, MLCCs primarily fluctuated with demand from mobile phones, PCs, automotive, and industrial applications, with market attention largely focused on inventory, prices, and economic cycles. Following the rapid increase in AI server power density, data center power supply architectures are entering a reset phase, and the role of capacitors is changing accordingly. NVIDIA has publicly promoted an 800VDC data center power architecture because when rack power increases from tens of kW to over 100kW, and even develops towards higher power density, current, copper loss, and conversion losses under low-voltage distribution rapidly amplify. NVIDIA stated in a technical article that from Hopper to Blackwell, the power consumption of a single GPU increased by 75%, and the rack power density brought by a 72-GPU system increased by 3.4 times, thus forcing an upgrade of the entire power supply architecture.

Since the advent of AI servers, the business opportunities for supercapacitors have been widely discussed, and this direction itself has a clear technical background. AI workloads exhibit synchronous characteristics, where loads experience rapid and significant transient disturbances during switching between training, inference, data movement, and synchronous computing. NVIDIA points out that the power draw of an AI rack can rapidly increase from about 30% utilization to 100% and then fall back within milliseconds, so the system must be designed for peak power. Panasonic's explanation of AI data centers is also direct: supercapacitors correspond to burst loads and transient peaks, functioning to absorb instantaneous surges, stabilize bus voltage, and reduce transient stress on PSUs and UPSs. Supercapacitors have therefore re-entered the discussion of data center power architectures.

Moving forward, market attention will extend from supercapacitors to the entire power train. AI data centers are facing a comprehensive upgrade of their power supply chains, with directions including higher voltage, higher efficiency, higher power density, and higher thermal loads. This process will first drive the high-voltage, high-capacitance MLCCs already deeply embedded in power systems. This is because PSUs, PFCs, LLC resonant circuits, snubbers, bypass circuits, 48V outputs, and board-level power modules all require capacitors with higher specifications. The growth in demand for high-voltage, high-capacitance MLCCs stems from specification upgrades following changes in power architecture conditions.

TDK's website content clearly explains the role of MLCCs in data center PSUs: high-voltage MLCCs of 630V or higher are used in bypass positions, paralleled with large electrolytic capacitors to reduce ripple; high-voltage, low-loss Class 1 C0G MLCCs are used in snubber and resonant positions; for 48V output positions, large-capacitance MLCCs of 75V to 100V are being introduced to reduce part count and support high-power-density designs. This configuration reflects not merely an increase in component count, but a simultaneous upward shift in voltage, capacitance, loss, temperature resistance, and reliability requirements. Here, capacitors directly influence conversion efficiency, thermal performance, ripple control, and system stability.

The 48V power system is the first layer of demand amplification in this round. Samsung Electro-Mechanics points out that as AI server power consumption rises, the adoption rate of 48V power systems increases because they can transmit higher power at the same current. The 48V system mainly falls into two positions: one is the 48V output circuit of the Server PSU, and the other is the 48V input circuit of the power module on the server board. For these two positions, designers are starting to introduce high-capacitance 100V MLCCs, while also needing to address soft termination to reduce mechanical stress and self-temperature rise caused by high ripple current. This means the key benefits are not just increased capacitance, but also high withstand voltage structures, packaging design, and reliability verification capabilities.

Power architecture upgrades will push high-voltage, high-capacitance MLCCs into higher voltage ranges. The number of MLCCs used in AI servers has already exceeded that of regular servers by 10 to 15 times, with demand extending to ultra-high capacitance and high-voltage MLCCs. To support 120kW-class racks, future architectures will shift from directly stepping down AC to 12V or 48V, to first rectifying to 800V high-voltage DC, and then distributing power within the rack; under this architecture, demand for 100V MLCCs will increase, and demand for large-sized MLCCs ranging from 1kV to 2kV will also grow simultaneously. This section extends the benefits of high-voltage, high-capacitance MLCCs from board and module levels all the way to rack-level power architecture.

From the perspective of component specifications, this upgrade has three core directions. First, the withstand voltage is increasing, with the importance of the 100V, 630V, 1kV, and 2kV ranges all rising. Second, the demand for high capacitance values is increasing, as both 48V systems and high-frequency, high-efficiency converters require higher capacitance density. Third, requirements for low loss and high reliability are simultaneously increasing, especially in LLC resonant, snubber, and bypass positions, where Class 1 C0G dielectric, high surge tolerance, high ripple current capability, and thermal stability all become essential conditions. Once these conditions enter the AI server supply chain, the benefits will not be evenly distributed among all MLCC manufacturers, but will instead concentrate on product lines with high-voltage, high-capacitance design, material control, stacking process, and verification capabilities.

Looking at the industrial chain, supercapacitors and high-voltage, high-capacitance MLCCs correspond to different positions on the same upgrade curve. Supercapacitors handle power buffering from milliseconds to short seconds, corresponding to peak absorption, power leveling, and bus smoothing, with benefits concentrated in activated carbon, electrolyte, aluminum foil, individual cells, and modules. High-voltage, high-capacitance MLCCs handle high-efficiency power conversion and stable power supply, with benefits concentrated in high-grade ceramic powder, internal electrode materials, stacking and sintering processes, high withstand voltage structures, soft termination, and part numbers that can pass verification for AI server PSUs, power modules, and future 800VDC systems. The former absorbs transient fluctuations, while the latter supports the stable operation of the entire high-power supply system.

AI data centers are pushing capacitors back from a commodity logic to a system-specific component logic. High-voltage, high-capacitance MLCCs will be the earliest and most easily underestimated part of this main trend. As data center power supply gradually evolves from traditional AC and low-voltage distribution to 48V, 800VDC, and higher power density, the growth in capacitor demand will not just be about increased quantity, but also a simultaneous rise in specifications, positions, unit prices, and verification thresholds. The focus to track next is which part numbers have transitioned from optional to essential, and which MLCC positions are beginning to take on higher-level system responsibilities.