The Next AI Bottleneck May Be an MLCC
The most visible components in artificial intelligence infrastructure are also the most expensive. Graphics processors dominate headlines. High-bandwidth memory attracts executive attention. Advanced networking, storage, power systems, and liquid cooling increasingly shape deployment schedules and capital plans. These are the technologies companies discuss when they describe the scale of an AI build.
But an AI system does not ship as a collection of headline components. It ships as a complete, functioning assembly. Every processor must be connected to memory. Every accelerator must receive stable power. Every board must manage noise, voltage fluctuations, heat, and transient electrical demands caused by high-performance devices switching rapidly between operating states. That work depends on thousands of smaller board-level components, including multilayer ceramic capacitors, or MLCCs. (Read More: The AI supply chain is bigger than GPUs)
An MLCC may cost only a few cents. Its commercial value may seem insignificant compared to a processor worth tens of thousands of dollars or a fully configured AI rack valued in the millions. Yet if the correct MLCC is unavailable, unapproved, or unable to meet the design’s electrical and thermal requirements, the value of every other component in the system becomes temporarily irrelevant. Without them, manufacturers cannot complete the board, commission the rack, or begin generating a return on the infrastructure investment.
That is why the developing MLCC shortage 2026 deserves more attention than it is currently receiving. The risk is not that every ceramic capacitor in every package and specification has suddenly disappeared. The more credible concern is narrower and more consequential: demand is increasing for compact, high-capacitance, high-voltage, low-loss, and thermally capable MLCCs used in AI computing, networking, and power systems, while market reporting indicates that availability and lead times for some of these higher-performance components are tightening.
This distinction matters. Broad claims create panic. Specific risk intelligence creates action. The question is not whether there is a universal shortage of all MLCCs. It is whether the specific capacitors required by a particular board, power architecture, voltage rail, or approved vendor list are available in the required volume and timeframe to support production. For a growing number of AI infrastructure programs, that question may soon become as important as processor or memory allocation.
The smallest parts can carry the largest production risk
MLCCs are among the most widely used components in modern electronics because they perform several essential functions in an extremely small footprint. At the most basic level, a capacitor stores and releases electrical energy. On a circuit board, MLCCs help stabilize voltage, filter noise, smooth power delivery, decouple sensitive devices from disturbances elsewhere in the circuit, and supply brief bursts of current when a processor’s demand changes faster than the upstream power system can respond.
Their multilayer construction allows manufacturers to place many alternating layers of ceramic dielectric and metal electrodes within a compact package. As TDK explains, the layered structure increases the effective electrode area and, therefore, the capacitance achievable within the component. This is one reason MLCCs have become indispensable wherever designers need substantial electrical performance without surrendering scarce board space.
AI boards intensify every part of that challenge. GPUs, CPUs, ASICs, memory devices, power-management ICs, networking controllers, and optical interfaces operate at high speed and demand precise, stable power. Current can change dramatically within very short intervals. If the voltage at the device falls outside its acceptable range, the result can be instability, reduced performance, data errors, or system failure. Capacitors placed close to these devices help respond to those transient demands and suppress unwanted electrical noise.
Samsung Electro-Mechanics describes the issue directly: high-performance processors in AI systems operate at very low voltages while drawing extremely high current, thereby increasing the total capacitance required to stabilize power delivery. The company also notes that high-speed network interface cards and PCIe connections operate in frequency ranges where noise suppression becomes increasingly important. In other words, MLCCs do not merely support the processor. They help create the electrical environment in which the processor can perform as designed.
This is what makes the component’s economics deceptive. Procurement teams often rank risks by unit price or annual spend. That is understandable: expensive categories deserve attention because they affect budgets, working capital, and supplier negotiations. But unit cost is not the same as production criticality. A capacitor with a negligible share of the total bill of materials can still have a binary effect on output. Either the correct part is installed, and the board can advance, or it is missing, and the assembly stops.
The result is a profound mismatch between the purchase price and the operational consequences. A low-value line item can strand expensive processors, memory, substrates, power modules, cooling hardware, and finished inventory. The two-cent component is not suddenly worth as much as the AI system. It is simply the component that determines whether the value already invested in that system can be realized.
AI servers are changing both MLCC quantity and capability
The pressure on AI server components is not driven solely by the number of new data centers under construction. Changes in the architecture of the systems within them are also driving it. AI servers concentrate more computing power, electrical current, heat, and high-speed communication into less physical space. That combination increases both the number of capacitors required and the performance expected from them.
Samsung Electro-Mechanics has stated that AI server power consumption can be five to ten times greater than that of conventional servers and that the higher power requirements lead designers to mount more MLCCs. The same source explains the physical constraint: many of these components must be positioned near the GPU, yet the available board area is limited. Smaller, higher-capacitance components therefore become especially important because they allow designers to place more effective capacitance within the available footprint.
Another Samsung product briefing says AI server boards can contain thousands of capacitors as designers densely integrate GPUs, ASICs, memory, and power devices. It also highlights a second challenge: the heat generated within densely packed AI systems increases the need for components that can operate reliably under more demanding thermal conditions. The emerging requirement is therefore not simply “more MLCCs.” It is more small, high-capacitance MLCCs with the electrical, thermal, and reliability characteristics appropriate for advanced infrastructure.
Power architecture adds another layer of demand. AI servers are increasingly adopting higher-voltage distribution within the rack and server to move greater amounts of power more efficiently. Samsung has connected the expansion of 48-volt AI server power systems with demand for high-capacitance, 100-volt MLCCs used in server power supplies and board-mounted power modules. Those applications must also account for mechanical stress, ripple current, and self-heating, not just nominal capacitance.
TDK reaches a similar conclusion in its application guidance for data-center power systems. As rack and server power density rises, power supply units and intermediate bus converters require greater efficiency, density, and reliability. TDK points to MLCC advantages, including high capacitance density, low equivalent series resistance and inductance, and the ability to distribute current and thermal load by using multiple capacitors in parallel.
These requirements show why the MLCC supply chain cannot be evaluated as one homogeneous market. Capacitance, case size, voltage rating, dielectric, tolerance, temperature behavior, termination, ripple-current capability, and effective capacitance under operating voltage can all affect whether a part is appropriate. Two devices that look similar in a purchasing database may behave differently in the actual circuit. A nominal match does not automatically make a technical substitute.
That specificity is where supply risk concentrates. Manufacturers may be able to produce enormous aggregate quantities of MLCCs while customers still struggle to obtain particular high-performance combinations. Capacity measured in total units does not necessarily equal capacity for a high-capacitance, high-voltage product in a compact case size with the correct temperature and reliability characteristics.
The product roadmaps of major manufacturers reinforce this direction. TDK introduced a 100-volt, 1-microfarad commercial MLCC in a 1608 metric case size in 2025, describing it as the industry’s highest capacitance for that voltage, size, and temperature characteristic at launch. Murata has also continued to introduce products that deliver higher capacitance and voltage capability in smaller footprints. These developments are evidence of where customer engineering requirements are moving: greater electrical capability in less space.
TDK’s 2026 full-year performance briefing adds a commercial signal. The company identified capacitors and other passive components used across AI data-center power conversion as a growth area and presented a goal of significantly expanding passive-component sales into AI data-center applications through fiscal 2031. That does not prove a shortage by itself. It shows that a major component manufacturer expects AI power systems to become an increasingly important source of demand for passive components.
Why the developing market is different from 2018
The electronics industry has seen this category become a production bottleneck before. During the 2018 MLCC crisis, demand from smartphones, automotive electronics, connected devices, and other rapidly expanding applications outpaced available capacity. Lead times extended, products moved onto allocation, prices increased, and customers searched aggressively for alternatives.
A few years back, Murata’s fiscal review described rapidly increasing capacitor demand, including strong growth in automotive electronics, and the company subsequently expanded and reinforced its MLCC production capacity. Contemporary industry accounts reported lead times exceeding 30 weeks, with some sources reporting significantly longer waits for constrained products.
There are clear similarities between that period and the MLCC shortage risk now developing in 2026. Both involve structural increases in electronic content. Both expose the difficulty of quickly adding specialized manufacturing capacity. Both demonstrate how demand can migrate toward higher-value or more technically demanding products faster than the supply base can adjust. And both can punish customers that treat passive components as interchangeable commodities until the moment they become difficult to obtain.
The differences may be even more important.
First, the source of incremental demand has changed. A broad mix of smartphones, vehicles, industrial electronics, and connected devices propelled the 2018 cycle. Those applications have not disappeared. Automotive electrification, advanced driver-assistance systems, industrial automation, telecommunications, medical equipment, and consumer products still compete for portions of the same manufacturing base. AI infrastructure is now adding another high-growth demand center, and it is doing so at the higher-performance end of the market.
Second, the design requirements are becoming more demanding. AI systems require greater capacitance near processors, more sophisticated noise suppression across high-speed interfaces, and components suited to high-density power architectures. Higher voltage, high ripple current, temperature performance, low electrical losses, and compact case sizes narrow the field of acceptable parts. The pressure is therefore not only on total output. It is on the mix of output.
Third, the industry enters this period with a different inventory structure. Market intelligence shared with Rand indicates that the channel does not appear to hold the same broad cushion of available MLCC inventory that helped absorb portions of the 2018 disruption. That observation should be treated as a directional market signal rather than a universal, auditable inventory count. However, current reports from independent distributors support concerns that availability is tightening for certain specifications.
Multiple distributors, including Rand, reported last month that MLCC supply risk was rising across AI servers, automotive, telecommunications, industrial, medical, and consumer applications, with higher-specification parts more exposed because they require more specialized manufacturing capability. Astute Group separately reported that AI-related demand was extending lead times for high-end MLCCs and identified particular pressure from AI server power systems.
These reports should not be interpreted as proof that every MLCC family is already in crisis. They do support a more precise conclusion: customers should not assume that today’s relative availability will persist across every high-capacitance or high-performance specification. By the time a constraint becomes visible in a broad industry lead-time chart, the most exposed customers may already be competing for the same approved supply.
The danger of solving the expensive problems first
AI infrastructure procurement naturally begins with the categories that appear most difficult and costly. Companies negotiate accelerator allocations, secure memory, plan high-speed network fabrics, arrange storage, design power distribution, and determine whether cooling and facility capacity can support the intended deployment. These are legitimate priorities. Without them, there is no AI system.
The mistake is treating the remaining bill of materials as a routine execution problem to be addressed later.
Traditional procurement processes often divide categories among different teams, suppliers, and contract manufacturers. High-value semiconductors may receive executive oversight, long-range forecasts, and direct supplier negotiations. Passives may be bundled into board-level procurement managed by an EMS provider, ODM, or lower-level commodity team. The OEM may see a completed assembly price without maintaining the same visibility into whether every capacitor, resistor, inductor, connector, and power device has actually been purchased and allocated to its program.
That model works when components are readily available and easily replenished. It becomes dangerous when demand accelerates. A contract manufacturer’s responsibility for buying the part is not the same as confirmed coverage. A supplier’s quotation is not the same as dedicated inventory. An approved manufacturer listed on the AVL is not the same as an available manufacturer with capacity. And inventory visible somewhere in the channel is not necessarily inventory available in the required date code, location, quantity, specification, and timeframe.
This is how expensive systems become stranded by inexpensive parts. The organization believes the difficult procurement work is complete because the processors and memory are secured. In reality, the build remains exposed to hundreds or thousands of lower-cost dependencies. When one of those dependencies fails, the company may discover the problem only after production is scheduled, premium components are already committed, and the time available to qualify an alternative has disappeared.
The economics of waiting amplify the risk. When the market begins to tighten, suppliers may shorten quote validity, reduce available quantities, or reprice inventory as replacement costs rise. A sourcing team can spend days seeking a marginally better unit price only to find that the original stock has been sold. Companies must therefore weigh the apparent savings on a low-cost component against the costs of delays, line disruptions, engineering escalations, and stranded high-value inventory.
The right objective is not to overbuy every passive component. It is to understand which low-cost components have the greatest potential to disrupt production and to manage those parts according to their operational consequences rather than their unit prices.
Where customers should look first
A useful MLCC risk review begins at the part-number level. Broad commodity labels are not enough. “Ceramic capacitors covered” may conceal exposure in a small number of specifications that have limited manufacturer capacity, a narrow approved vendor list, or no validated substitute.
Customers should first examine high-capacitance MLCCs used near GPUs, CPUs, ASICs, memory, networking controllers, and power-management devices. These locations often require substantial effective capacitance in limited board space and may be especially sensitive to changes in electrical characteristics.
The next area is the power-conversion path. That includes capacitors used in power supply units, intermediate bus converters, voltage regulator modules, 48-volt input and output stages, and the lower-voltage rails supplying processors. Higher-voltage MLCCs, parts exposed to high ripple current, and components selected for low-loss behavior warrant closer attention, as suitable alternatives may be less abundant.
Engineers must also review the thermal requirements. AI hardware places large amounts of heat into dense assemblies. A substitute that meets capacitance and package requirements at room temperature may not provide the required performance across the operating range. Teams should confirm temperature characteristics, voltage derating, DC-bias behavior, ripple-current performance, and the component’s effective, not merely nominal, capacitance under actual circuit conditions.
Physical constraints matter as well. An electrically suitable part in a larger package may not fit the board. A smaller component may behave differently under voltage or thermal stress. Termination design can affect mechanical reliability. Changes that appear minor on a spreadsheet may require engineering analysis, qualification, and customer approval.
Finally, customers should identify sole-sourced positions and AVL limitations. A common passive value may still pose exceptional risk if only one manufacturer and a single exact part number are approved. Conversely, a technically demanding component can be manageable when several qualified sources and validated alternatives are already available.
The review should answer practical questions:
- Which MLCCs can stop the build if supply is interrupted?
- How much demand is covered by purchased, program-dedicated inventory?
- Where is that inventory physically located?
- Which requirements depend on unconfirmed forecasts or supplier quotations?
- Which parts have only one approved manufacturer?
- Which same-manufacturer upgrades could be acceptable?
- Which cross-manufacturer alternatives have been technically evaluated?
- How long would engineering validation and customer approval take?
- Which parts are exposed to rising demand in AI, automotive, industrial, or communications?
- What is the financial impact of buffering each critical item compared with the cost of a stopped build?
This converts the BOM from a purchasing list into a production-risk map.
Buffering must be selective, visible, and tied to real demand
Buffer inventory is one of the most direct defenses against a passive component shortage, but it must be designed carefully. Buying every MLCC on the BOM in excess quantities can consume cash, create excess inventory, and shift risk from availability to obsolescence. Buying nothing because the components are inexpensive and historically available can leave the entire program exposed.
The correct approach is selective buffering based on criticality, lead time, demand confidence, qualification flexibility, replacement difficulty, and the cost of disruption. Components with limited approved sources, long qualification cycles, high-performance specifications, or disproportionate production impact deserve more protection than generic parts with abundant substitutes.
Visibility is equally important. Customers need to know whether buffer inventory is actually owned, where it is held, whether it is dedicated to their program, and what conditions govern its release. Inventory somewhere inside a contract manufacturer’s network may be committed to other programs. Supplier backlog may be mistaken for supply. Forecast coverage may be treated as purchase coverage. These distinctions remain invisible until demand exceeds what is truly available.
Hubbing, bonded inventory, scheduled releases, and other structured inventory programs can help align supply protection with actual consumption. The goal is not simply to accumulate parts. It is to create a controlled bridge between uncertain supply and forecast production while preserving accountability for ownership, location, quality, and allocation.
Timing matters. The best moment to establish a buffer is before every customer reaches the same conclusion. Once a shortage becomes widely accepted, available inventory is repriced, suppliers become less willing to commit, and buyers compete for a shrinking set of approved parts. A proactive plan uses market intelligence to act while choices still exist.
A broader AVL creates options, but engineering must lead
Buffering buys time. Engineering flexibility creates alternatives.
Supply chains do not balance themselves when demand shifts faster than capacity. Many companies expanded their approved vendor lists during the 2018 MLCC crisis because they discovered that narrow approvals had converted ordinary sourcing problems into production emergencies. That lesson remains relevant. An AVL with several technically validated manufacturers gives sourcing teams more supply paths and reduces dependence on a single factory, region, or product family.
However, broadening an AVL is not a clerical exercise. MLCCs are not universally interchangeable. Dielectric type, capacitance tolerance, voltage rating, DC-bias response, temperature characteristics, equivalent series resistance, equivalent series inductance, package size, termination, flex performance, and reliability requirements can all affect circuit behavior. The appropriate replacement must be evaluated against the application, not merely matched against a few catalog fields.
There is also an important distinction between an upgrade and a cross. An upgrade generally remains within the same manufacturer and part family while improving a specification such as voltage or temperature capability. Depending on the design and the customer’s approval rules, that may be easier to accept. A cross introduces a component from another manufacturer. It can provide valuable supply flexibility, but it may require deeper technical comparison, testing, documentation, and formal AVL approval.
This work is most effective before the shortage. When engineers have time, they can compare datasheets, model component behavior, test samples, examine performance under actual operating conditions, and document the rationale for approval. When a line is already threatened, every step becomes an emergency, and the organization faces pressure to accept risk it would otherwise evaluate more carefully.
The same principle applies to quality. Tight markets attract unfamiliar supply, fragmented lots, older date codes, and parts moving through nontraditional channels. Expanding sourcing options must therefore be matched by disciplined traceability, inspection, authentication, and testing. Availability without confidence is not continuity.
The next bottleneck will be the component the plan overlooked
The central lesson of the AI infrastructure buildout is that the system is limited by its least available critical dependency. That dependency can move. At one moment it may be GPUs. At another it may be memory, optical transceivers, power equipment, cooling capacity, substrates, connectors, or high-capacitance MLCCs.
This is why supply chains can no longer be managed as a collection of unrelated commodity strategies. Companies must plan compute, memory, networking, power, cooling, and board-level components in parallel because these elements create value only when they are available together. Optimizing one category while ignoring the rest of the system can create the appearance of progress without producing a deployable result.
The emerging pressure in the MLCC supply chain offers customers a window to act before a possible shortage becomes a universal emergency. That action should not begin with panic buying. It should begin with visibility: identify the capacitors that can stop the build, confirm real coverage, examine sole-source positions, map technical alternatives, estimate qualification time, and establish selective buffers where the economics justify them.
For AI infrastructure companies, this is a form of capital protection. Securing a low-cost capacitor protects the value of the processors, memory, networking, power systems, storage, engineering time, and facility capacity already committed to the project. The relevant comparison is not between the prices of one MLCC and another. It is the cost of acting early against the cost of an idle production line or delayed deployment.
Rand sees the complete BOM
At Rand Technology, we believe production continuity depends on understanding the entire system, not simply sourcing its most expensive components. The components that receive the least attention can become the constraints that matter most, especially when demand changes faster than traditional procurement processes can respond.
Rand works with OEMs, EMS providers, and leading technology companies to identify component availability risks, resolve shortages, reduce costs, manage inventory, and support products from new product introduction through end-of-life. Our global market intelligence helps customers recognize emerging pressure. Our sourcing network helps locate supply across regions and channels. Our component engineering capabilities support the evaluation of crosses, upgrades, and second-source options. Our inventory programs help customers build strategic buffers and improve visibility. At the same time, Rand Certified™ quality processes, in-house quality-control hubs, and advanced inspection and testing give customers confidence when they must secure supply outside traditional channels.
In the event of a potential shortage of passive components, those capabilities must operate together. Locating inventory is only the first question. The customer must also know whether the part is technically suitable, commercially executable, properly documented, authentic, and available on a schedule that protects the build. When an exact part cannot satisfy the full requirement, the next step may be an approved alternative, an engineering evaluation, a staged purchase, or a structured inventory program.
The MLCC shortage of 2026 may not develop evenly across all product families, manufacturers, or end markets. But the warning signs around high-capacitance MLCCs and other performance-oriented AI infrastructure components are strong enough to justify immediate review. Companies that wait for a universal shortage declaration may discover that their most important part numbers became constrained much earlier.
AI companies are investing billions to increase computing capacity. Protecting that investment requires attention at the opposite end of the cost spectrum. The next AI bottleneck may not be the processor everyone is watching. It may be the two-cent component no one thought to ask about until the build stopped.
