Economic Observer reporter Zheng Chenye

In the past three months, the top three MLCC suppliers worldwide have completed relay price increases.

In March, Murata Manufacturing Co., Ltd., the world's largest MLCC supplier, officially notified customers that it would raise prices for AI servers and automotive grade MLCCs by 15% to 35%, effective from April 1st. This is Murata's first large-scale price adjustment in three years. Subsequently, Sun Power followed suit across the board in May, and Samsung Electric also announced an increase in consumer electronics MLCC prices starting from June 1st.

MLCC， That is, multi-layer ceramic capacitors, known as "electronic industry rice" in the industry, are shipped nearly five trillion pieces annually worldwide. Each piece is usually only two millimeters long and over one millimeter wide, and its function is to be attached to a circuit board for instantaneous voltage regulation of the chip. The usage of a mobile phone is about one thousand, the usage of a high-end electric vehicle is about tens of thousands, and the usage of an AI server cabinet exceeds 400000.

The reason listed by Murata in the price increase letter is the rise in silver prices: MLCC electrodes and pastes require silver and nickel, and silver prices have more than doubled in the past year. The increase in raw material prices is not uncommon in the MLCC industry. Jiang Wei, the market manager of a capacitor manufacturer in southern China, told the Economic Observer that the 2018 round of MLCC price increases was also accompanied by an increase in raw material costs. Dealers hoarded goods and pushed up channel prices by more than ten times, but the market came and went quickly.

Jiang Wei said that this round of price increases is completely different from the previous round. "In the past, when there was a shortage, the quantity was not enough, and adding production lines could solve it. This time, downstream manufacturers are forcing manufacturers to cross over to produce products, and before the yield rate is fully achieved, large-scale production scheduling is required. Making a high-end product for AI can occupy as much production capacity as four ordinary products, and no matter how many production lines are added, it cannot catch up.

An industry insider told Economic Observer that the delivery cycle of high-end MLCC has been extended from 8 to 12 weeks to 16 to 24 weeks.

From 2000 to 400000 pieces

The power consumption of each generation of Nvidia AI chips is significantly increasing - public information shows that the H100 single card consumes about 700 watts, and by the latest generation GB300, it has reached about 1400 watts.

When the power doubles, the instantaneous current fluctuation during chip operation also intensifies, and the response speed of the power module cannot keep up with the millisecond level current mutation. High capacitance value (i.e. the amount of charge that a single capacitor can store, the higher the capacitance value, the stronger the voltage stabilization ability) MLCCs need to be densely arranged around the chip for voltage compensation.

The role of MLCC here is to stabilize the power supply voltage of the chip: every operation of the chip is accompanied by a sharp fluctuation in current. If the voltage cannot keep up with this fluctuation, it may result in calculation errors, training interruptions, or even direct damage to the chip. The higher the chip power, the more violent the current changes, and the more MLCCs need to be attached to stabilize the voltage nearby, with higher capacity requirements.

A securities analyst who has long been concerned about the semiconductor industry chain told reporters that during his recent visits to the industry chain, he learned that a regular server motherboard requires about 2000 MLCCs, while the usage of AI server motherboards ranges from 15000 to 25000, with a difference of about ten times between the two. Based on the Nvidia GB200 NVL72 full cabinet calculation, the MLCC usage exceeds 400000 pieces.

With each upgrade of Nvidia's GPU platform, the usage of MLCC around a single GPU increases exponentially. The securities analyst told reporters that taking the MLCC product specification, which has the largest usage in AI servers, as an example, in the H100 era, the peripheral usage of each GPU exceeded 200, which increased to nearly 500 in B200, nearly 1500 in GB200, and nearly 5000 in the latest Vera Rubin platform.

That is to say, the single GPU usage of the same MLCC has increased by more than 20 times for the fourth generation product.

In addition to the increase in quantity, the capacity specifications of MLCC are also rapidly rising. The unit of measurement for capacitance is microfarads. The larger the number, the more charge a single MLCC can store and the stronger its voltage regulation ability. A few years ago, the mainstream capacity of MLCC used in AI servers was still 22 microfarads, but now it has risen to 47 microfarads. Nvidia's latest Vera Rubin platform has begun to extensively use products with 100 microfarads.

Previously, MLCC with 100 microfarads had almost no large-scale application scenarios, and Vera Rubin was the first platform to turn it into a standard configuration, using about 250 GPU peripherals. At present, the industrial chain is preparing for the next generation of 330 microfarads products. With each increase in capacity, the corresponding production difficulty and capacity consumption will significantly increase.

In a recent research report released by China International Capital Corporation, it was estimated that the global annual demand for MLCCs will approach five trillion by 2025, and the MLCCs consumed by AI servers in a year will only account for 2% to 3% of this total.

Jiang Wei told reporters that making a high-end AI product requires the production capacity of four ordinary products. AI servers only use 2% to 3% of the global MLCC in terms of quantity, but they account for nearly 10% of the overall production capacity. During the interview, the reporter learned that this proportion may increase to 15% to 20% next year - MLCC for mobile phones, automobiles, and industrial equipment and MLCC for AI servers share the same batch of production lines. The more capacity AI orders occupy, the less will be left for other downstream industries.

This is the reason why the entire industry feels the tight supply of MLCC: the annual shipment volume of five trillion pieces is large enough, but AI servers are squeezing out more and more production capacity from other downstream sources.

In February of this year, Murata's president, Keiju Nakajima, publicly stated that the number of inquiries from customers regarding high-end MLCC was twice Murata's existing production capacity, which was "completely unable to meet". He also stated that high-end MLCC cannot rapidly expand production, and the tight supply situation in the next two years may continue.

In fact, as early as an investor exchange event at the end of 2025, Murata's management had already raised the annual compound growth rate forecast for AI servers using MLCC from 18% to 30% from 2025 to 2030.

The financial reports of Samsung Electric and Sunac also confirm this demand intensity.

In the first quarter of 2026, Samsung Electric's revenue reached 3.21 trillion Korean won, a year-on-year increase of 17%, breaking through the 3 trillion Korean won mark for the first time in a single quarter. At its performance briefing at the end of April, the company stated that it is in discussions with several AI technology giants regarding long-term MLCC supply contracts, and its equipment investment this year will double compared to last year.

In the fiscal year 2025 (ending in March 2026), the net profit of Sun Induced Power increased by 536% year-on-year, and the order to shipment ratio of the capacitor department reached 1.31, marking three consecutive quarters of growth. The company expects AI server MLCC revenue to grow by approximately 80% in the 2026 fiscal year.

In a research report released by China International Capital Corporation in March 2026, based on GPU and ASIC chip power calculations, it is expected that the demand for AI server MLCC will increase by 87% and 88% respectively in 2026 and 2027, driving global server MLCC demand growth of 49% and 61%. Moreover, this growth rate does not take into account the increasing volume of self-developed AI chips (ASICs) being developed by cloud vendors - ASIC stocking by companies such as Google and Amazon is also driving MLCC demand.

The process cannot keep up with the iteration speed

The basic structure of MLCC is a stack of extremely thin ceramic plates and metal electrodes alternately stacked together, and finally sintered into a whole. The more layers there are, the more charges can be stored, and the corresponding capacitance value is larger. Standard consumer electronics MLCC has approximately 50 to 100 layers, while high-capacity products used in AI servers have an average of 500 layers, and some ultra high capacity products have more than 1300 layers.

Jiang Wei told reporters that if there is a difference of five times in the number of layers, the workload of the stacking process will differ by five times, and the actual efficiency difference will be much greater than five times. In addition, as the number of layers increases, the external dimensions cannot change, as the space left for each MLCC on the circuit board is fixed. This also means that each layer must be made thinner.

The thinnest single-layer ceramic film currently has a thickness of 0.3 to 0.4 micrometers (1 micrometer is equivalent to one thousandth of a millimeter), approximately three to four hundred nanometers. Jiang Wei said that with a thickness of 300 nanometers, there are only three grains in the vertical direction, which is close to the limit that ceramic materials can withstand. In addition, the strength of the ceramic film decreases after thinning, and it is easy to crush or cause damage to the internal electrodes during stacking, resulting in a risk of short circuit. To compensate for the strength, more colloids need to be added to each layer.

But if there are too many colloids, the breathability will deteriorate, and the bubbles trapped between the layers during the stacking process will be difficult to remove, requiring a significant extension of the holding time for each layer. The actual difference between the stacking rhythm of one hundred layers and five hundred layers is much greater than five times.

After stacking, it enters sintering. The excess colloid must be thoroughly removed before sintering, as the carbon, hydrogen, and oxygen components will turn into gas at high temperatures. If not fully removed, it will swell and crack from the inside of the product. An increase in the number of glue dispensing steps will prolong the required time. In addition, sintering itself is more complex: the ceramic layer expands at high temperatures, the nickel electrode layer shrinks at high temperatures, and each layer generates internal stress. The more layers there are, the greater the accumulated internal stress, and the temperature ramp must be slower.

Jiang Wei told reporters that if the temperature rises too quickly, stress concentration will directly lead to cracking. The standard MLCC takes approximately 27 days to complete the 13 processes from feeding to discharging the finished product, while the AI oriented ultra-high capacity products take over 50 days, resulting in a nearly doubled production cycle.

In addition, achieving ceramic membranes at the level of less than one micron also puts higher demands on upstream raw materials. The ceramic powder particles used in conventional MLCC typically have a diameter of 200 to 300 nanometers, while AI server grade products require ultrafine powders below 120 nanometers or even 80 nanometers. At the same time, the particle diameter of nickel powder must also match the ceramic powder, and the two must shrink synchronously during sintering. Mismatching can lead to delamination or cracking.

That is to say, every step from powder to slurry to sintering process needs to be re adapted for high-end products.

At the same time as doubling the production cycle, the yield rate also needs to be discounted. The yield rate of standard consumer electronics MLCC is usually above 99%, and the defect rate increases exponentially after reaching 500 layers. During the interview, the reporter learned that the yield rate of ultra-high capacity MLCC products is currently about 40%, and the yield rate of some small-sized ultra-high capacity specification products is even only a dozen percentage points.

So, overall, the production capacity required to make an AI ultra-high capacity MLCC is at least equivalent to four standard products, and some specifications are even higher. Moreover, with the continuous upgrading of Nvidia platform, the capacity and stack level of AI server MLCC are constantly increasing.

Each MLCC leader is expanding production, but the speed still cannot keep up with the growth in demand. Public information shows that Murata recently announced an additional investment of about 80 billion yen in MLCC, to be implemented in two fiscal years, with an expected increase in production capacity of 10% to 15%; Previously, Murata had invested about 47 billion yen to build a new factory in Izumo City, Shimane Prefecture, which is expected to start production in 2026. Jiang Wei also told reporters that Samsung Electric's Tianjin factory will expand production by about 20% this year, but the equipment will not be in place until the end of June, and the release of production capacity will not be until the end of August; The scale of the new factory in the Philippines is about 1.5 times the existing production capacity, and the earliest output will be in early 2028.

In Jiang Wei's view, the biggest bottleneck for manufacturers to expand production is space, and the lack of space to place equipment is the most direct constraint. Next is the equipment delivery time, with a lead time of 16 months for the casting machine and approximately 10 months for the sintering furnace and laminating machine. It takes at least one and a half years to build a new factory from scratch and release production capacity. Even if new equipment is added to the existing factory, it takes three to four months to debug and run the production line to normal output. In addition, the debugging of the new production line is much more complex than the old line, and it is difficult to determine which link is the cause of the yield problem when all the equipment is new.

Jiang Wei believes that the current MLCC price increase cycle is fundamentally different from the previous one. During the price increase cycle from 2018 to 2019, the three leading MLCC companies in Japan collectively shifted from consumer electronics to automotive specifications, resulting in a shortage of consumer electronics. Dealers hoarded goods in large quantities, and channel prices increased by more than ten times, which was a combination of quantity gap and speculative hoarding; The driving force for this current round lies in the speed of product iteration - previously, MLCC manufacturers would produce new products first, and downstream demand would only rise three to five years later, giving manufacturers ample time to improve yield and material processes. However, now that Nvidia switches to a new generation platform every year, MLCC manufacturers have to start mass production before their yield is stable, which is equivalent to producing the next generation of products across generations.

During the interview, the reporter learned that several industry insiders hold a similar view: the last round of price increases was driven by hoarding and supply-demand mismatch, which lasted for about a year and a half before falling back; This round of price increases is a structural shortage caused by technological iteration on the production capacity side, and expanding production capacity cannot solve the problem of process not keeping up with the iteration speed.

Chain Reaction

After MLCC began to raise prices, downstream reactions quickly differentiated.

TrendForce consulting analyst Chen Weisheng told Economic Observer that in the first quarter of 2026, the global MLCC industry showed a clear polarization: the demand for high-end MLCC related to AI grew against the trend, while the demand for mid to low end consumer electronics MLCC remained low due to the industry's off-season and rising costs.

An industry chain insider told reporters that the delivery time for AI high-capacity MLCCs has been extended from 8 to 12 weeks to 16 to 24 weeks, and some product specifications have started to receive limited orders. During the interview, the reporter also learned that MLCC has risen to the top three in the overall material cost of AI servers, and the price increase has a direct impact on the overall cost.

The feeling on the automotive side is equally clear. Mr. Shen, the product director of a car chip company in Shenzhen, told reporters that he will feel the tightening of high-end MLCC supply from the end of 2025, and the delivery cycle of Japanese manufacturers is becoming increasingly unstable.

Sun Power has been reducing some of its vehicle production capacity and shifting towards supplying AI servers. Jiang Wei also stated that the continuous pressure of AI servers on high-end production capacity is being transmitted to the automotive supply chain. For example, Samsung Electric's car specification orders on the manufacturing side have more than doubled year-on-year, partly due to the share transfer caused by the squeezing out of production capacity by Japanese manufacturers.

Preventive stocking of AI high-capacity products has also emerged on the channel side. Mr. Shen emphasized to reporters that currently, global distributors are actively reducing inventory for general products, but increasing procurement for high-capacity AI related products, which is also reinforcing the expectation of shortages.

Mr. Shen also stated that there is a fundamental difference between the current inventory level and the MLCC price surge in 2018: the current channel supplier inventory is about 1.7 months, while the end customer inventory is only two to three weeks; During the peak of dealer stockpiling in 2018-2019, channel inventory was piled up for six or seven months.

Chen Weisheng told reporters that after Murata and Samsung Electric concentrated their high-end production capacity towards AI, the spillover of mid to low end orders began in the fourth quarter of 2025 and significantly accelerated in the second quarter of this year. The driving force for production conversion is very direct - the average unit price of AI server MLCC is about nine to ten times that of consumer electronics MLCC, and the gross profit margin of ultra-high capacity products can reach over 60%, which is completely different from the profit margin of consumer electronics MLCC.

In fact, the increment brought by the overflow orders of industry leaders has been clearly reflected in the financial reports of domestic companies.

In the first quarter of 2026, Sanhuan Group (300408. SZ) achieved a revenue of 2.681 billion yuan, a year-on-year increase of 46%; Net profit was 791 million yuan, a year-on-year increase of 48%. As a result, Sanhuan Group has become the fastest growing MLCC company in China during this cycle. According to relevant financial reports, the company's high-capacity MLCC products have been supplied in bulk for AI server applications, with a dielectric layer thickness exceeding 1 micron. The ceramic powder and slurry are 100% self-sufficient, and over 90% of the production equipment is independently developed and manufactured.

In addition, another company in the industry, Fenghua High tech (000636. SZ), had a revenue of 1.515 billion yuan in the first quarter of 2026, a year-on-year increase of 19%; The net profit was 88.56 million yuan, a year-on-year increase of 37%, reversing the situation of "increasing income without increasing profits" in 2025. The management of the company stated at the performance briefing held in May 2026 that the product has been introduced into the supply chain of top domestic AI server customers. The high-end capacitor base in Xianghe Industrial Park completed the project in April, and the monthly production capacity of MLCC exceeded 50 billion pieces.

The upstream material end also benefits. Boqian New Materials (605376. SH) has a net profit of 219 million yuan in 2025, a year-on-year increase of over 150%; In the first quarter of 2026, the revenue was 410 million yuan, a year-on-year increase of 64%. The 80 nanometer MLCC nickel powder produced by the company is one of the few products in the world that can be supplied in bulk for AI server level applications.

The overflow of orders from industry leaders and product price increases have opened up profit improvement space for domestic manufacturers, but this space has clear boundaries.

At present, domestic manufacturers can achieve the level of "high capacity", and both Sanhuan Group and Fenghua High tech have mass production capabilities, while downstream customers are also accelerating verification. But for "ultra-high capacity" products beyond this level, such as specifications of 100 microfarads and higher capacity values, currently only Murata, Samsung Electric, and Sunac are supplying globally.

An industry insider told reporters that the yield gap of domestic manufacturers in ultra-high capacity products may take about three to four years to narrow. The barrier to ultra-high capacity MLCC lies not in a single link, but in the full chain collaboration from powder formulation to stacking process to sintering control. Any shortcoming in any link will drag down the overall yield.

Of course, after Japanese and Korean manufacturers collectively raised prices by 15% to 35%, downstream customers' willingness to find alternative suppliers has also significantly increased. During the interview, the reporter also learned that several domestic new energy vehicle companies have accelerated the certification process of domestic MLCC since the beginning of this year.

Chen Weisheng emphasized to reporters that the competitive landscape of the global MLCC high-end market is still highly concentrated, with the top three suppliers accounting for over 85% of the market share. In his opinion, the current price increase cycle has opened a relatively longer window of opportunity for domestic manufacturers, but whether domestic manufacturers can fully seize this window depends on whether the iteration speed of processes and materials can keep up with the pace of Nvidia's product upgrades.