Analyse fra Goldman Sachs argumenterer for, at MLCC’er – multilayer ceramic capacitors – kan blive den næste store flaskehals i AI-infrastrukturen, på samme måde som memory tidligere blev det. MLCC’er er små keramiske kondensatorer, der stabiliserer strømforsyningen tæt på GPU’er og ASICs i AI-servere. De fungerer som ultrahurtige mini-energilagre, der udjævner strømspidser og filtrerer elektrisk støj, hvilket er kritisk, når AI-chips skifter ekstremt hurtigt mellem høj og lav belastning. En avanceret AI-serverrack kan ifølge analysen kræve op mod 600.000 MLCC’er.
Hovedpointen er, at efterspørgslen fra AI-servere vokser langt hurtigere end industrien kan udvide kapaciteten. MLCC-markedet er relativt lille, og produktionskapaciteten vurderes kun at kunne vokse lidt over 10 pct. årligt, mens servermarkedet vokser markant hurtigere. Goldman Sachs vurderer, at MLCC’er allerede er blevet den tredjestørste komponentomkostning i AI-servere efter GPU’er og memory.
Analysen fremhæver især Murata, Samsung Electro-Mechanics, Taiyo Yuden som centrale leverandører af de avancerede lavspændings- og højkapacitets-MLCC’er, der bruges tæt på GPU’er og ASICs. TDK nævnes som stærk inden for højspændings-MLCC’er til strømforsyningskredsløb, men ikke som nuværende teknologisk leder i de lavspændingskomponenter, der er mest kritiske omkring AI-processorer.
Prisudviklingen er ifølge analysen begyndt at accelerere kraftigt. Spotpriser på visse MLCC-typer er steget 2-4 gange siden kinesisk nytår, mens enkelte efterspurgte modeller ifølge markedsaktører i Shenzhen er steget mere end 20 gange. Det skyldes både AI-servere, elbiler, hamstring, dobbeltbestillinger og stigende inputpriser på bl.a. nikkel og sølv.
Konklusionen er, at MLCC’er kan blive en strukturel mangelvare i anden halvdel af året, især fordi store AI-platforme ændrer design og øger antallet af MLCC’er pr. motherboard. Analysen peger derfor på, at investorer kan få eksponering via producenter som Murata, Taiyo Yuden, Samsung Electro-Mechanics, Yageo, Fenghua Advanced Technology og Chaozhou Three-Circle Group – eller via Goldmans asiatiske MLCC-aktiekurv.
Kort sagt: Analysen ser MLCC’er som “den nye memory-trade” i AI-cyklussen: en lavprofileret, men kritisk komponent, hvor begrænset kapacitet, eksploderende AI-efterspørgsel og begyndende prisstigninger kan skabe betydelig indtjenings- og kursopside for de førende producenter.
Uddrag fra Goldman Sachs
One month ago, we explained why “capacitors are the new memory”, laying out a detailed analysis showing how the boring, humble capacitor (technically Multi Layer Ceramic Capacitor or MLCC) will be “the next sector that generates memory-like returns, or even surpasses them” in the context of the rollout of not only Nvidia’s next generation Vera Rubin rack, but also as part of the broader AI buildout.
For those who missed it, the thesis in a nutshell: the big question everyone wants answered is what is the “next bottleneck”, i.e., which is the next sector that generates memory-like returns, or even surpasses them? The answer is found in humble, boring capacitors, or rather Multilayer Ceramic Capacitors (MLCCs).
In a note published by Goldman’s strategists, they reported that MLCCs (a roughly $15 billion market of which the server market is $1.3 billion, growing at a 80% CAGR while slowing in some other key industries such as cars and cell phones) have risen to become the third-highest cost item in AI server builds, trailing only GPUs and memory.
Multilayer Ceramic Capacitors (MLCCs) act as a tiny, ultra-fast rechargeable battery for electricity. While a normal battery stores a lot of energy and releases it slowly, an MLCC stores a tiny amount of energy and releases it in a fraction of a millisecond. They serve two main purposes:
- Power Smoothing: They absorb sudden spikes in electricity and fill in sudden drops, ensuring a smooth, steady flow of power to sensitive chips.
- Noise Filtering: They block unwanted electrical “noise” (interference) that could corrupt digital data.
How these fit into AI Infrastructure: AI servers do not process data in a smooth, predictable stream. When an AI model handles a massive calculation, the processor instantly demands a massive surge of electrical power. When the calculation finishes, the power demand drops to zero. Power supplies cannot react fast enough to these microsecond changes.
MLCCs sit directly next to the AI chips to instantly release power into them when they need it, preventing the server from crashing. Because AI chips (like NVIDIA GPUs) handle billions of tasks simultaneously, they require massive networks of these capacitors. A single advanced AI server rack requires up to 600,000 MLCCs working together to keep the system stable.
Key players: For low-voltage, high-capacity MLCCs around GPUs/ASICs in AI servers, miniaturization and increased capacity are progressing simultaneously within limited board space. Technological requirements are becoming increasingly stringent. Murata, SEMCO (Samsung Electromechanics), and Taiyo Yuden are the three suppliers, and they will continue to benefit fully from demand expansion, leading to rising prices. On the other hand, TDK does not currently possess the technology to enter the market for low-voltage, high-capacity MLCC market for GPUs/ASICs (it is awaiting materials development from its collaboration with Nippon Chemical Industrial). However, the company is receiving brisk orders for high-voltage, high-capacity MLCCs for use around power supply circuits. It can use almost the same product technology as in automotive applications such as EVs, which is expected to contribute to a rise in factory utilization rates
The bottleneck: Goldman’s Allen Chang recently highlighted that the (relatively small) MLCC industry’s production capacity is limited to a little over 10% annual growth and companies have started to hike MLCC prices, as equipment and materials are produced in-house (and constrained by in-house engineering resources). If much of this capacity increase is absorbed by the sharp rise in demand for AI servers, the current cycle could see prolonged supply/demand tightness.
The following sequential schematic of AI supply/demand bottlenecks and prices for components and materials and dated as of a few April, shows that unlike memory (DRAM and NAND), ABF Substrates and CCLs (or Copper Clad Laminate which are the foundational material used to manufacturing Printer Circuit Boards), MLCC have so far barely seen any price increases (this is important). But that is now rapidly changing, and off all AI components and materials, MLCCs (along with ABF and CCL) have the longest upside price runway. Translation: memory is yesterday’s trade – the real upside is in capacitors (more on this below).
Perhaps most importantly, we said that just like memory prices exploded in late 2025 when attention focused on that particular bottleneck in the AI supercycle rollout, so now a sense of supply/demand tightness is strengthening across the entire MLCC industry, triggered by the vast demand opportunity from AI servers (which Goldman estimates will grow by approx. 4.3x from FY25 to FY30).
While demand from AI servers and automotive applications (which have high production load) remains firm, we are starting to see moves from smartphone/PC customers to seek long-term contracts, despite decreasing volumes due to memory shortages. As MLCC industry production capacity is diverted to AI servers, concerns about future component procurement are actually growing among customers with weaker demand.
Among the other reasons cited by industry participants for higher MLCC prices are:
- High nickel and silver are affecting all segments
- Supply-demand mismatch in the high-end (high capacitance, high voltage) segment, which is used in autos & servers
- High-end MLCC lead time is over 20 weeks
- Spot/distributor prices have increased by 20%-40% for low capacitance and consumer device MLCCs due to hoarding and double booking, especially in China
- OEM contracts have not seen large price hikes yet
There is much more in the full report, including an analysis of how much more MLCC content a typical Vera Rubin rack has relative to the current Blackwell offering (assuming constant prices)…
… the key publicly traded players, and much more (we urge all subscribers to take a look).
Which brings us to today. Fast forward 4 weeks from our article when in a report from the South China Morning Post titled “Tiny capacitor, huge demand: the AI frenzy driving MLCC prices higher”, we get confirmation of what we expected: capacitor prices are exploding, in some cases blowing away memory gans.
Echoing verbatim what we wrote, the SCMP writes that “Last year, it was memory chips. This year, the hottest commodity in Shenzhen’s Huaqiangbei electronics market is a component no larger than a postage stamp.” Also echoing what we wrote, the report confirms that “these tiny capacitors are currently under the spotlight because the insatiable computing needs of artificial intelligence (AI) have caused an explosion in demand”, a surge described as “scary” by Katsuya Sase, CEO of Japanese MLCC maker Taiyo Yuden.
But most importantly, the report confirms that as we expected, prices have exploded in recent weeks, as Chinese vendors of multilayer ceramic capacitors (MLCCs) are riding a massive wave of soaring prices driven by the global artificial intelligence boom.
Because these minuscule components are vital for regulating electric currents, global manufacturing lines are scrambling to keep up with skyrocketing demand from AI server clusters and electric vehicles. The resulting supply crunch has sent the spot market into overdrive, with prices jumping two- to fourfold since Chinese New Year for high-capacitance MLCC.
“I don’t see any signs of prices cooling down,” said local distributor Wu, who mainly sells components from Japanese industry leader Murata Manufacturing to fellow traders and secondary distributors.
While high-capacity components are driving demand, the price hikes are hitting the spot market across the board. Some models have surged from 10 yuan per 1,000 units to 40 yuan (US$5.90), Wu said. But while inquiries are flooding his shop, actual transactions remain sluggish as buyers balk – for now – at the inflated price tags.
Other merchants report even steeper climbs that began accelerating through April and May. One merchant surnamed Zheng, who deals in MLCCs from South Korea’s Samsung Electro-Mechanics (SEMCO) and domestic firms like Chaozhou Three-Circle Group, said that highly sought-after models, such as the 0805 22μF MLCC, have seen prices skyrocket by over 20 times, blowing away the best gains by either DDR or NAND memory during this cycle.
Both imported and domestic products have been surging, Zheng said. Another trader of SEMCO MLCCs said that frequent stock shortages have plagued his business since April due to production struggling to keep pace with demand.
What is playing out on the chaotic floors of Huaqiangbei is what Goldman Sachs has described as “the largest and longest cycle in history”. Leading Japanese and South Korean manufacturers are currently inundated with orders, fuelling expectations of a severe, prolonged supply crunch.
The industry is increasingly facing the risk of a “structural shortage” in the second half of this year, according to a report released on Wednesday by Taipei-based market research firm TrendForce. And unlike memory, where hyperscalers have at least made backups supply chain plans, very few have noticed the far greater bottleneck that MLCCs represent.
Next-generation AI platforms are undergoing frequent design modifications in their final qualification stages, drastically multiplying the number of high-end MLCCs required per motherboard, the firm said.
The usage of one type of MLCC in Advanced Micro Device’s MI450 platform recently surged by a staggering 632% – jumping from 1,440 units per board to 10,544 units – after engineers swapped out all aluminium electrolytic and tantalum capacitors, TrendForce said.
The report noted that MLCC demand would hit a “new peak” later this year as tech giants like Google, Amazon Web Services and Meta Platforms ramp up production for their custom AI chips.
As major manufacturers prioritize long-term contracts, clients relying on the spot market face hefty price premiums and severe shipment delays, TrendForce said.
Readers who wish to piggyback on the move can either purchase the actual products, which may be challenging, or by purchasing the stocks of the industry leaders. One should certainly have exposure to the three top names in the space, Murata, Taiyo Yuden and SEMCO, other names such as Yageo (2327.TWD) and Fenghua Advanced Technology (000636.CH) and Chaozhou Three-Circle Group (300408.CH) are emerging as the frontrunners.
Alternatively on can buy Goldman’s MLCC basket of Asian stocks (GSXAMLCS). It consists of all the key players across Japan, South Korea, Taiwan, and China-A that offer differentiated exposure to the MLCC value chain, spanning pure-play manufacturers, diversified passive component producers, and upstream materials suppliers.
The basket has only recently started to rally (orange line below) but is still lagging other hot Ai themes (lines blue and purple).
2 md. adgang for
2 x 49 kr.
Få straks adgang til denne artikel og derefter 2 måneder til alle artikler på ugebrev.dk
- Alle artikler på ugebrev.dk
- Om investering, finans, ledelse, samfundsansvar, life science og Bestyrelsesguiden.dk
- Daglige nyhedsmails med nyheder og analyser
Tilbuddet gælder til 31. juni 2026. Abonnement fortsætter til normalpris på 249 kr. efter bindingsperiode på to måneder. Opsig når du vil - til udgang af den anden måned. Tilbud gælder kun, hvis du ikke har haft abonnement på ØU udgivelser de seneste tre måneder
Allerede abonnent? Log ind her
