Four major high-capacitance MLCC manufacturers announce synchronized price hikes!

01. From 15 cents to 50 cents: In-stock items will be sold first, and original factory items will be replenished later.

Let’s start with the spot market. Prices began rising in late February, with the first wave of hikes ranging from 15% to 20%. By June, average price increases for high-capacitance components had reached 40% to 50%, and certain part numbers even surged two to three times higher.

The most representative example is the 1206 package 47μF MLCC: it was priced at ¥0.15 at the end of February, climbed to ¥0.35 by June, with some channel quotations hitting ¥0.50.

Will prices keep rising from current levels? As long as supply gaps persist, stockpilers will continue driving up prices, and original manufacturers generally will not intervene to curb the uptrend.

Price adjustments from OEMs unfolded at a slower pace, yet they have not ceased since late April.

The price adjustment mechanism works simply as follows: manufacturers raise prices confidently whenever supply falls short of demand or supply gaps emerge. Price revisions generally take place once per quarter.

Current operating rates remain robust: Murata’s overall capacity utilization exceeds 90%, while its high-end high-capacitance lines hit above 95%; Taiyo Yuden averages over 85%; Taiwanese suppliers register an average utilization rate above 80%, with high-end products for AI and automotive applications reaching more than 90%.

The Book-to-Bill (BB) Ratio stands at 1.5 and 1.3 for Japanese manufacturers, and 1.2 for Taiwanese players. The figure was merely 1.1–1.3 last year. Orders are genuine.

02  Rubin Requires 600,000 Pieces; Demand Only Drops to 540,000 Even With Specification Downgrade

On the demand side, NVIDIA serves as the primary growth driver. Each cabinet of the GB300 generation consumes approximately 450,000 MLCCs, while the figure for Rubin will exceed 600,000 pieces.

There have been recent rumors that Rubin may undergo specification downgrades: switching from HBM4E back to HBM3, with power consumption reduced from 2,300W to 1,800W. This remains unconfirmed and is merely a hypothesis. The speculation stems from rising shipments of GB200 and GB300 alongside insufficient HBM4 production capacity, which could push customers to continue adopting the previous-generation memory solution.

Even if such spec downgrades materialize, we can quantify the impact: a 20% drop in overall system power consumption corresponds to a 10% reduction in MLCC usage, bringing consumption down from 600,000 to 540,000 pieces. The demand magnitude remains substantial.

Another point worth noting: while volume declines by 10%, value falls by 20%. The cutbacks are mainly concentrated around GPUs, CPUs and HBM, consisting entirely of small-form-factor high-capacitance components with relatively high unit prices.

The price of the 0402 package 20μF MLCC stood at only 6–7 cents before the price hike, and has now risen to 11–12 cents; the 47μF variant commands an even higher price.

The part number mix on the board side breaks down as follows: 22μF accounts for more than half, 10μF makes up 25%, and sizes of 47μF and above represent 25–30%.

The next generation will see an even more dramatic increase. Preliminary estimates for Rubin Ultra show MLCC consumption reaching 3–5 times that of Rubin, roughly 3 million pieces, with the average unit price rising from ¥0.05 to ¥0.07.

The incremental demand comes from two sources. First, a natural rise in high-capacitance usage driven by higher power consumption. Second, purely new additions: vertical power delivery routed upward from beneath will require an extra set of RF capacitors or silicon capacitors embedded into the PCB. These feature compact form factors suitable for embedding. Once the 800 V DC high-voltage architecture is implemented, additional high-voltage-resistant components will also be required. This is not a substitution, but a net addition of components.

The optical module segment also merits discussion. One 800G module requires approximately 200 RF MLCCs at a unit price of ¥0.20–0.30. Demand rises linearly for 1.6T modules, which need a minimum of 500 pieces.

Silicon capacitors will capture some market share going forward: one silicon capacitor can replace four RF MLCCs, supporting frequencies up to 100 GHz. They carry a slight price premium but offer smaller size and superior stability.

Samsung Electro-Mechanics has secured silicon capacitor orders worth ¥6.8 billion.

Unable to expand: Equipment waits a year, rewiring results in a 50% loss.

Demand outlook is well defined, so why is capacity unable to keep up?

First, factory facilities must be constructed from scratch.

Second, equipment. High-end precision machinery needs to be ordered from Japanese suppliers, with lead times exceeding one year.

When these two constraints coincide, a standard new production line takes 12 to 18 months to complete. Under urgent circumstances, manufacturers can only repurpose existing lines, which comes at a steep cost with an average yield trade-off ratio of 1:4.

A production line originally capable of outputting 30,000 automotive-grade pieces can deliver fewer than 10,000 pieces after conversion for server high-capacitance MLCCs. Lines manufacturing standard components such as 105°C and 106°C types with an original throughput of 50,000 pieces also drop below 10,000 pieces post conversion.

Murata serves as a ready example. It had been carrying out line conversion work, yet halted operations in June and July, remaining suspended for two months.

The challenge is not technical, but customer-related. Long-term clients with partnerships spanning more than a decade have voiced strong opposition when notified that mid-to-low-end product supply would be discontinued, forcing them to source from Taiwanese and mainland Chinese manufacturers. Murata cannot withstand such backlash.

The suspension further widens the supply gap.

Its own capacity expansion is also subject to delays: factory construction will only finish in August, equipment installation follows in September, and commissioning requires another one to two months. Limited output of high-capacitance products will not be available until the fourth quarter at the earliest.

Yield rate is another major hurdle. For grades including 22μF, 47μF and 100μF, Murata achieves yields above 80%; Samsung Electro-Mechanics reaches just over 70%, while less competitive players see yields below 50%.

Furthermore, newly built lines cannot go into stable operation immediately after installation. It takes two to three months to tune the production line to normal operating conditions, followed by another three to four months of production ramp-up. The whole process spans roughly six months. This phase cannot be shortened, as operators lack hands-on experience with these products.

Current monthly capacity ranking among major suppliers is as follows:

Murata: 150 billion pieces;

Samsung Electro-Mechanics: 110 billion pieces;

Taiyo Yuden: 90 billion pieces;

Yageo: 80 billion pieces.

Yageo plans to add an extra 10 billion pieces of monthly capacity this year, with funding and factory facilities already secured in Taiwan.

Upstream stuck in nickel powder

For ultra-high-capacity MLCCs like those used in NVIDIA servers, the requirements for nickel powder have far surpassed automotive-grade standards. Automotive-grade products only need particle sizes of 200–300 nanometers, whereas high-capacitance server products require sizes below 120 nanometers, with purity starting at four nines (99.99%) and even reaching five nines (99.999%). Achieving such ultrafine and ultrapure specifications is extremely challenging. Currently, Japanese manufacturers are the best in this field, and everyone has to queue up for their products, which naturally drives prices up—this is a direct reason for the recent cost increases in the upstream supply chain.

There are two domestic Chinese players. BQ New Materials achieved 80 nm last year and introduced 60 nm in the first half of this year, placing them among the world’s top tier. They mainly supply Samsung Electro-Mechanics, but their output still falls short of total demand. In terms of capacity, their production and sales were about 1,450 tons last year, and this year they are targeting at least 2,200 tons, potentially reaching around 2,300 tons by year-end if all goes well, with unit prices around ¥1,600. High-end small-particle-size powders account for over half of their mix. They are expanding capacity by 600–1,000 tons this year, with an additional 600 tons planned for next year and the year after.

Shandong Sinocera currently produces 150 nm powder; they have not yet achieved the most advanced ultrafine grades, but this year they have begun supplying automotive-grade products, and Samsung Electro-Mechanics is also using their material. Their production process is the same as that of their Japanese counterparts, so future breakthroughs are possible, but none have occurred so far.

On the Japanese side, Shoei Chemical produces 100 nm powder and has a target of 60 nm that has not yet been realized. However, 100 nm is already sufficient for server applications, and they mainly supply Murata. When Samsung Electro-Mechanics faces shortages, they also purchase from Shoei. Interestingly, Murata has invested in Shoei—essentially saying, "Expand more capacity and lock in more supply for me," which mirrors the relationship between Samsung Electro-Mechanics and BQ New Materials, where Samsung supports BQ to secure its own supply chain.

Where are domestic manufacturers located?

The progress of the three manufacturers varies.

Three Circle Group (Sanhuan Group) has cleared two major hurdles this year: mass production of the 0805 package 47μF capacitor (which was not possible last year) and the 1206 package 100μF capacitor – both are large‑size, high‑capacitance products. They are currently sending samples to both Huawei Kunpeng and Inspur simultaneously, and this sampling process has been ongoing for over three months. Such validation typically takes five to six months to reach a conclusion; if no major issues such as crashes or system freezes occur after installation on the end‑user side, orders will follow. Their capacity expansion from this year to next year focuses mainly on large‑size high‑capacitance products and automotive‑grade components. In addition, they have developed the 01005 miniaturized component but have not yet started mass production. If they succeed with this, their entire communications product line will be fully established.

Fenghua Advanced Technology (Fenghua Gaoke) has its strength in automotive‑grade products. Last year, it secured over 30% of the automotive‑grade orders among major domestic manufacturers, and automotive‑grade products already account for more than 15% of its own output value, which is expected to exceed 20% this year. Its automotive‑grade resistors, inductors, and other components have also been widely adopted by domestic automotive manufacturers. Their medium‑ and high‑voltage products perform well, reaching 500 to 1,000 volts, and the next step is to see whether they can enter the 800‑volt high‑voltage platform. They are also sending samples of large‑size high‑capacitance products to Huawei.

Yageo relies on KEMET, which it acquired at the end of 2020. Automotive‑grade MLCCs are its strongest segment, accounting for over 20% of its total output value; when combined with industrial‑grade high‑capacitance products, the share totals approximately 26%, while the remaining more than 70% is still mid‑to‑low capacitance.

Last year, AI‑related revenue accounted for 12% of its business, and by July this year that figure had already reached 16%. Google and Amazon are currently signing long‑term contracts with Yageo. However, it should be noted that the performance gaps are mainly reflected in two areas. The first is high‑temperature tolerance. 

Ordinary large‑size high‑capacitance MLCCs only need to withstand 85°C, but locations inside servers where temperatures exceed 85°C require components rated for 105°C or 125°C; if the temperature tolerance is insufficient, the capacitance value will drop significantly.

The second gap is high‑voltage tolerance. For an 800‑volt platform, it is not enough for a component to merely withstand 800V – MLCCs have a DC bias issue, meaning that capacitance drops as voltage and temperature change. Therefore, a margin of over 40% must be reserved, and the withstand voltage needs to reach 1000V, preferably 1200V or above, before customers feel confident using the parts. TDK can achieve 1000‑2000V, so their products are not a problem.

There is also a layer of opportunity. On the NVIDIA side, large‑size high‑capacitance products have not yet been made available to Taiwanese manufacturers; they can only supply mid‑to‑low‑end materials rated below 105°C. When Taiwanese manufacturers send samples to major overseas customers, the validation cycle takes a minimum of one year and up to two years. Meanwhile, Taiwanese ODM manufacturers – such as Foxconn Industrial Internet, Inventec, and Quanta – have considerable influence over supplier selection during contract manufacturing, so they naturally give priority to Taiwanese suppliers.