A multilayer ceramic capacitor is one of the least exotic components inside an AI server. That has not stopped it from becoming increasingly difficult to secure.
In September 2026, Samsung Electro-Mechanics announced a KRW 1.07 trillion contract to supply AI-server MLCCs during 2027, its largest long-term MLCC agreement to date. By then, distributors had already reported lengthening lead times and stock-outs that had started in high-capacitance server parts and were spreading into more conventional specifications.
The interesting part is not another component shortage. It is why a component manufactured by the billions is becoming more important as computing advances.
AI does not fundamentally change what an MLCC does. It changes the electrical environment around it. Accelerators draw more power, current changes faster, power must pass through more conversion stages, and less board area is available to stabilize that power. The result is a system that needs more capacitance in more places, often with tougher voltage, temperature, size and reliability requirements.
That combination makes MLCC demand increasingly tied to the architecture of AI computing rather than simply to the cycle in smartphones and PCs.
AI creates a power-delivery problem before it creates a capacitor problem
MLCCs perform several unglamorous but essential jobs. They filter electrical noise, smooth voltage, decouple one circuit from another and provide small reservoirs of charge when processors suddenly demand current.
Those functions become harder as accelerator power rises.
A modern AI system does not deliver electricity directly from a data-center power supply into a GPU. Power moves through several voltage domains and conversion stages before reaching logic operating at very low voltage and very high current. Capacitors are required around those conversion stages and close to processors to keep voltage stable when loads change within extremely short periods.
This creates two different MLCC opportunities at the same time.
Further upstream, higher-voltage products are needed as data centers move toward higher-voltage distribution architectures. TDK, for example, expects future data-center power units to move toward 400–800V and is expanding products including high-voltage MLCCs, film capacitors and aluminum electrolytic capacitors.
Closer to the accelerator, the problem reverses. Voltage falls while current rises sharply. Here, compact high-capacitance MLCCs become useful because they combine small size with low electrical resistance and inductance. Large numbers can be placed in parallel around voltage regulators, boards and accelerators to distribute current and stabilize the power rail.
AI therefore increases capacitor content through architecture, not merely through server shipments.
More compute means both more MLCCs and harder MLCCs
The number of capacitors is only half of the story.
Higher-power AI platforms are pushing suppliers toward products with more capacitance in smaller packages, higher voltage tolerance, greater temperature stability and higher reliability. Samsung Electro-Mechanics describes AI-server MLCCs as components expected to operate continuously under high-temperature and high-voltage conditions. TDK is simultaneously expanding both high-voltage products for data-center power infrastructure and low-voltage, high-capacitance products closer to computing devices.
This matters economically because MLCC manufacturing capacity is not perfectly interchangeable.
A factory capable of producing commodity capacitors does not automatically become additional supply for high-capacitance or high-reliability AI products. Materials, dielectric layer thickness, stacking, electrode design, firing, yield and customer qualification all become more demanding as specifications rise.
The industry can therefore experience plenty of nominal MLCC capacity while remaining constrained in the specifications that AI systems need most.
That distinction has become visible in the 2026 market. High-capacitance products tightened first. Suppliers then redirected more production toward server and other high-value applications. Lead times subsequently lengthened across additional product categories, while shortages began appearing in parts previously considered more conventional.
This is one reason an AI-driven shortage can spill into ordinary MLCCs even when consumer electronics demand is not especially strong. The scarce resource is not only total factory capacity. It is how suppliers choose to allocate qualified capacity.
The supply response is slower than the component looks
MLCCs are tiny. The production system behind them is not.
Leading suppliers must repeatedly print and stack extremely thin ceramic dielectric and metal-electrode layers, laminate them, cut them, fire them at high temperature, form external electrodes and test the finished components. Higher capacitance in a small footprint typically requires thinner dielectric layers and more of them.
Small changes in specification can therefore affect process difficulty and yield.
Capacity expansion also has to be matched with materials, equipment, skilled labor and customer qualification. Samsung Electro-Mechanics has accelerated additional capacity while outlining a long-term investment program for high-value MLCCs and package substrates used in AI data centers. TDK is also investing in MLCC materials and new low-voltage, high-capacitance products.
The important constraint is time. Demand can shift from one accelerator generation to another faster than high-end component capacity can be built, stabilized and qualified.
This helps explain why long-term supply agreements have become more visible. For an AI system costing tens or hundreds of thousands of dollars, losing availability because of a passive component that represents only a small fraction of system cost is an unattractive trade. Supply assurance can matter more than squeezing the final few cents from the capacitor bill.
Silicon capacitors are real competition—but not a universal replacement
The strongest challenge to the structural MLCC argument is technological substitution.
Silicon capacitors are particularly relevant. Semiconductor processes can create extremely thin capacitors with very low parasitic inductance, making them attractive inside or immediately adjacent to high-performance packages. Samsung Electro-Mechanics has already signed a large multi-year silicon-capacitor supply agreement and positions the technology for GPUs, HBM and other advanced semiconductor packages.
That sounds like the beginning of an MLCC replacement cycle. The actual architecture is more complicated.
Samsung’s own AI-server application map shows conventional MLCCs, tantalum capacitors and silicon capacitors being used within the same system. TDK describes a similarly heterogeneous power architecture in which aluminum electrolytic and film capacitors handle parts of the higher-voltage system, while MLCCs and other technologies serve lower-voltage stages closer to the compute device.
These components overlap, but they are not interchangeable in every position.
MLCCs offer a difficult-to-replicate combination of small footprint, low equivalent series resistance and inductance, fast high-frequency response, broad voltage coverage, temperature capability, mature surface-mount manufacturing and enormous production scale. Polymer and tantalum capacitors can provide attractive bulk capacitance and other electrical characteristics. Aluminum and film capacitors are useful further upstream. Silicon capacitors offer exceptional thinness and low inductance close to the die.
The likely result is not one capacitor technology winning the server. It is a hierarchy of capacitor technologies optimized for different electrical distances from the processor.
Silicon capacitors may take some of the most valuable positions near or inside advanced packages. That does not automatically remove the much larger population of board-level capacitors used across power conversion, networking, memory and peripheral circuits.
Technology can reduce MLCC count without reducing MLCC demand
There is another complication. Better MLCCs can replace several older MLCCs.
TDK’s recent development of a higher-capacitance 100V MLCC for AI-server power systems is a good example: higher capacitance density can reduce the number of components required for a particular function. Similar gains in materials and manufacturing should continue.
This puts a natural limit on simple forecasts based on ever-rising component counts.
But component efficiency and system growth move in opposite directions. A next-generation capacitor may halve the number required at one location while a new accelerator platform adds more power stages, more accelerators, higher current and additional networking elsewhere in the rack.
The structural question is therefore not whether every future AI server contains more individual MLCCs than the previous one. It is whether the amount of demanding capacitance required per unit of AI infrastructure keeps increasing faster than technology can remove discrete components.
Current product roadmaps suggest that this remains the more useful framing.
Why the cycle is spreading beyond premium MLCCs
The present market also shows how a high-end structural shift can affect a product category that still contains large commodity segments.
Japanese and Korean suppliers have been directing more production toward AI servers, automotive electronics and other high-value applications. Market data in 2026 showed lead times extending first in high-capacitance MLCCs and later across more general-purpose sizes. Some lower- and mid-range demand has consequently moved toward Taiwanese and Chinese suppliers.
This does not make every MLCC structurally scarce.
Consumer demand can weaken. Inventory can be rebuilt. Customers can qualify additional suppliers. Chinese manufacturers can close part of the technology gap. Distributor orders can also exaggerate a shortage when customers begin ordering earlier than necessary.
Those are normal cyclical forces, and they matter.
What has changed is the source of demand at the upper end. AI infrastructure is adding a large, power-intensive end market whose component requirements are increasingly different from those of consumer electronics. As leading suppliers devote more engineering and manufacturing resources to that market, the resulting mix shift can tighten supply elsewhere even without a broad consumer boom.
What could break the structural thesis
Three developments would matter more than a temporary change in distributor inventory.
First, AI infrastructure investment could slow enough for high-end component capacity to catch demand. A prolonged gap between installed compute and actual utilization would eventually reach the passive-component supply chain.
Second, power-delivery architecture could reduce discrete capacitance faster than computing power increases it. Vertical power delivery, integrated voltage regulation, advanced packaging and silicon capacitors all move some power-management functions closer to the processor. This is the most credible technological challenge to long-run MLCC intensity.
Third, new suppliers could qualify high-end products more quickly than expected. Chinese manufacturers are already benefiting from capacity reallocation in mainstream products. Moving successfully into high-capacitance, high-reliability AI specifications would change the supply economics considerably.
None requires the MLCC itself to disappear.
A commodity component is becoming an architectural component
The strongest case for MLCC growth is not that AI servers simply contain an extraordinary number of capacitors. Counts vary by platform and by what is included in the calculation.
The stronger argument is architectural.
More compute requires more power. More power creates additional voltage domains, higher current, faster transients and tougher thermal conditions. Those changes increase the need for capacitance across the system while raising the specifications of the capacitors closest to the accelerator.
Silicon, polymer, tantalum, film and aluminum capacitors will take parts of that opportunity. In several positions they are better technologies than MLCCs. Their expansion is evidence that AI power delivery is becoming more complex, not evidence that one capacitor category is about to eliminate the others.
MLCCs occupy too many useful points along that power hierarchy, at too attractive a combination of electrical performance, size, reliability and manufacturing scale, for a single substitute to remove them quickly.
AI may therefore do something unusual to one of electronics’ most ordinary components: turn a mature passive device into a continuing constraint on how much compute can be packed into a system.
Research Cut-off: 2026-09-23 16:26
Public primary sources
- Samsung Electro-Mechanics — AI Server MLCC supply contract, Sep. 1, 2026. KRW 1.0722 trillion contract for 2027 and more than 10 global LTA customers.
- Samsung Electro-Mechanics — AI Server application architecture. Shows MLCC, tantalum and silicon capacitors coexisting around CPU/GPU, voltage regulation and memory/network functions.
- Samsung Electro-Mechanics — Silicon Capacitor Lineup, Jun. 17, 2026. Describes ultra-thin structure and sub-1 pH ESL as key silicon-capacitor advantages.
- Samsung Electro-Mechanics — silicon capacitor supply agreement, May 20, 2026. Approximately KRW 1.5 trillion contract for 2027–2028, focused on high-performance packages such as GPUs and HBM.
- Samsung Electro-Mechanics — Mid- to Long-Term Investment Strategy, Jul. 3, 2026. Approximately KRW 15 trillion planned through 2040 for AI data-center package substrates and high-value MLCC capabilities in Busan.
- TDK — FY March 2026 full-year briefing. TDK describes future 400–800V data-center power architecture and simultaneous development of high-voltage and low-voltage/high-capacitance MLCCs.
- TDK — shareholder Q&A. Management identifies AI data centers as a new MLCC growth market and highlights miniaturization and higher capacitance as development priorities.
- TDK — Investor Day Q&A. Notes rising MLCC demand in high-voltage AI-data-center power lines alongside aluminum electrolytic capacitors and inductors.
Public secondary sources
- TrendForce, Sep. 10, 2026. Reports production reallocation by Japanese and Korean suppliers toward higher-end AI products, with lower- and mid-range demand increasingly shifting toward Chinese suppliers.
- TrendForce, Aug. 27, 2026. Reports prospective direct-channel price increases at Samsung Electro-Mechanics and high utilization among Taiwanese and Chinese suppliers as AI-related demand tightens supply.
- TrendForce, Jul. 9, 2026. Describes structural shortages in high-end and medium/high-capacitance MLCCs and capacity reallocation toward AI-server specifications.
- TrendForce, May 18, 2026. Uses board-level examples to show rising MLCC content as accelerator power increases; these counts were used directionally rather than reproduced in the article because platform boundaries differ.