Murata's ¥1 Trillion MLCC Bet: AI Server Demand Will Outlast Its New Capacity by Years

Murata Manufacturing, the world's largest maker of multilayer ceramic capacitors (MLCCs), will spend about ¥1 trillion (roughly $6.3 billion) between fiscal 2026 and fiscal 2030 to expand production of the tiny components that stabilize power inside AI servers. Evertiq reported the investment announcement on October 30, the same day the company disclosed its results. The headline number is large, but the more telling figures sit underneath it: Murata says it wants MLCC sales in fiscal 2030 to be at least double fiscal 2025 levels, while its own plan raises production capacity by only about 60% over the same period.

That gap is the story. A sales target that outruns unit capacity can only be closed with a richer product mix and firmer prices, which means Murata is not planning to end the tightness in AI-grade capacitors. It is planning to supply more of them at higher value, while demand from GPU racks is forecast to grow faster still.

President Norio Nakajima unveiled the program on October 30 alongside the company's results for the six months through September. He described AI data centers and optical communications as being in a major demand expansion cycle and called AI-related demand the biggest challenge and opportunity in Murata's history, according to EE Times Japan's account of the briefing. Nikkei Asia reported on November 1 with additional detail on why the parts have become so scarce.

Where the ¥1 Trillion Goes and When It Arrives

Murata split the budget into roughly ¥350 billion for buildings and ¥650 billion for production equipment. The largest disclosed project is about ¥150 billion at Okayama Murata in Setouchi, which will become the company's third Japanese MLCC base alongside Fukui and Izumo. Construction of the MLCC production building starts in April 2027 and is scheduled to finish at the end of 2028, with a separate building for ceramic sheets and raw materials due in 2030 and a target workforce of 1,500 to 2,000 people.

Elsewhere, Murata plans a new components plant in Yasugi, Shimane, run by Izumo Murata, with timing and budget still undecided; a new production building in Thailand due in 2030; and new equipment in an unused section of an existing Philippine facility by 2027. Existing plants in Japan, Southeast Asia, and China will keep receiving equipment each year. Murata has not disclosed how most of the budget divides among those sites, or how much of the new output is already committed to customers.

The schedule matters more than the total. Murata's earlier plan grew MLCC capacity about 10% a year. The new plan roughly doubles that pace to about 20% a year in fiscal 2027 and 2028, then slows, reaching about 60% more capacity than today by fiscal 2030. Because the Okayama building will not be finished until late 2028, near-term relief has to come from equipment squeezed into existing floor space. That is the fastest lever available: a new building at Okayama takes about 20 months from groundbreaking to completion before any production line can be qualified inside it, while equipment added to an operating plant skips that wait. Murata has said its overseas investment will ramp up from fiscal 2028. The company's utilization was roughly 90% to 95% in both of the first two fiscal quarters, leaving little slack to absorb surprises.

The financial scale is unusual for a component maker. The plan averages about ¥200 billion a year and equals roughly 1.9 times the ¥520 billion operating profit Murata now forecasts for the fiscal year ending March 2027, a forecast it raised from ¥430 billion on the same day. First-half revenue rose 20.8% to ¥1.09 trillion, and capacitors grew to more than half of total sales.

Read more: Nvidia AI server prices rise more than 15% as memory costs hit the AI stack

Why AI Servers Need So Many MLCCs

An MLCC is a stack of extremely thin ceramic dielectric layers, usually barium titanate, interleaved with nickel electrode layers and capped with copper terminations. Capacitance rises with the number of layers and the active area between them, so the competitive frontier is fitting more and thinner layers into the same footprint; mass-produced parts already stack more than 1,500 layers.

In an AI server, those parts act as local energy reservoirs. A modern GPU can swing from light load to full load in microseconds, and the voltage regulators feeding it cannot respond instantly. Arrays of MLCCs placed around each GPU, high-bandwidth memory stack, and network chip supply current during those transients and filter high-frequency noise, keeping the core voltage within tolerance. Each processor and memory stack needs its own capacitor network, tuned for low impedance into the hundreds of megahertz.

Two design shifts are multiplying the count. Rack power delivery is moving to 48-volt distribution with conversion stages placed close to the processor package, which concentrates current transients and pushes dense capacitor arrays onto substrates, interposers, and power-module boards, often in 0201 case sizes or smaller. At the facility level, Nvidia is steering the industry toward 800-volt DC distribution, which requires higher-voltage capacitors in resonant converter stages, snubbers, and DC-link filtering.

The result is steep growth in content per rack. Nikkei-cited estimates put a single AI server at 15,000 to 25,000 MLCCs, up from 10,000 to 20,000 in earlier reporting, though some coverage applied the figure to a rack rather than a server. A Morgan Stanley analysis estimated MLCC content at about $4,320 per Vera Rubin NVL72 rack versus about $1,530 for a GB300 rack, an increase of roughly 182%. Nvidia has not published these counts, so they should be read as supply-chain estimates.

The hardest parts are not the ones produced in the greatest volume. Murata says demand is strongest for small, high-capacitance products, high-voltage power-supply parts, and memory-related components, and that rising processor power forces more capacitance into limited board space, increasing production load per part. High-capacitance ceramic parts also lose effective capacitance under applied DC voltage, a phenomenon called DC-bias derating, so engineers routinely specify more capacitors than the nominal values suggest.

Sales Double on 60% More Capacity: The Arithmetic of a Seller's Market

The gap between Murata's sales and capacity targets follows directly from that product shift. A high-capacitance 0201 part for a GPU power rail carries far more value than a commodity capacitor for a phone, and it consumes more of a factory's thin-layer printing and stacking capacity. Murata's AI data-center MLCC segment is on track to more than double in this fiscal year alone, according to EE Times Japan's account of the October 30 briefing.

Demand forecasts run far ahead of a 60% capacity increase. In December 2025, Murata projected that MLCC demand from AI servers would reach 3.3 times fiscal 2025 levels by fiscal 2030, according to Kyodo News reporting summarized by TrendForce. Goldman Sachs forecasts the AI-server MLCC market growing from roughly $1.3 billion in 2025 to $5.8 billion in 2030. Murata's broader portfolio still serves phones, cars, and PCs, so overall capacity does not need to match AI growth one for one. But the numbers imply that AI-grade parts will absorb a rising share of new output, and that buyers of commodity parts may feel the squeeze too.

Prices and lead times are already moving. Distributor prices for AI-grade MLCCs have risen sharply this year while popular parts have seen lead times stretch significantly, reflecting utilization running at roughly 90% to 95% across Murata's lines. Murata has publicly favored price discipline: Executive Deputy President Masanori Minamide said in a September interview, as reported, that frequent price changes could attract new competitors and damage long-term value.

Read more: ASML revenue surges 55% to $11.7B as CEO warns EUV supply falls short through 2028

Samsung Electro-Mechanics Locks In Customers With Long-Term Contracts

Murata's main rival is taking a different route to the same shortage. On the same day as Murata's announcement, Samsung Electro-Mechanics disclosed an 879.3 billion won AI-server MLCC supply contract with an unnamed global company, covering deliveries from December 21, 2026 through December 21, 2027. It was the company's eighth long-term supply contract disclosed this year, bringing the 2026 total to about 5.18 trillion won, and the company expects more than half of its 2027 MLCC sales to come from such agreements.

Samsung Electro-Mechanics claims more than 40% of the AI-server MLCC market, a company figure that has not been independently audited. Murata holds an estimated 40% of the overall MLCC market, but no reliable public figure establishes which supplier leads in AI-server parts specifically. Reported distributor lead times hint at the different strategies: an iM Securities comparison cited standard lead times of 61 weeks for Samsung Electro-Mechanics, 28 weeks for Yageo, and 17 weeks for Murata. Contracted volume gives the Korean supplier revenue visibility; Murata's shorter lead times and capacity plan position it to win orders from buyers who cannot wait.

The competition is not limited to ceramics. Samsung Electro-Mechanics has begun mass-producing silicon capacitors embedded in package substrates for AI chips, a technology that targets the decoupling positions closest to the die. Murata itself lists embedded-in-substrate MLCCs and higher voltage ratings among the technologies it is developing, and is separately entering AI-server power modules.

Late-Arriving Plants and an Uncertain AI Capex Peak

The plan's biggest risk is timing in both directions. Murata previously expected AI infrastructure investment to peak around 2028, but management now sees the buildout running to around 2030, partly because power supply constraints have slowed AI data center construction. Goldman Sachs similarly forecasts the AI-server MLCC market growing through 2030. If the cycle turns earlier than expected, the Okayama and Thailand buildings could arrive just as demand cools, a pattern that followed the 2018 MLCC shortage. If the buildout runs longer, the 20%-a-year ramp in 2027 and 2028 may not be enough, and AI server builders will keep competing for allocation. Murata frames the post-2030 market as a hedge, citing edge devices such as smart glasses, drones, autonomous vehicles, and humanoid robots as the next demand wave.

Order data suggests the near term is tight. Murata's ratio of orders to shipments in the April-to-June quarter was 1.27, above the 1.25 level at which the 2018 shortage began, according to TrendForce figures.

For AI infrastructure planners, Murata's announcement reframes the capacitor question. The world's largest supplier is adding capacity at twice its previous pace, but it is building a business plan around scarcity-level value rather than a return to cheap, abundant parts. Until the Okayama lines start producing at the end of 2028, the practical constraint on Vera Rubin-class buildouts will be how much of Murata's roughly 20% annual capacity growth, and Samsung Electro-Mechanics' contracted output, a given buyer can secure in advance.