How a Rice-Grain-Sized Part Landed Trillion-Won AI Deals

This piece explains a trend in the electronic components industry. It is not a recommendation to buy or sell any particular stock. The investor reactions, earnings forecasts, and contract-size estimates mentioned here are drawn from the judgment and reporting of each cited source. Please make your own investment decisions at your own responsibility.

---

An obscure little part just became the star of trillion-won contracts

Reader, have you ever heard of MLCCs? Crack open a smartphone and you’ll find tiny specks packed onto the circuit board. A good number of those specks are MLCCs. It’s a component nobody ever mentions by name in daily life — yet lately, it’s the subject of contracts worth trillions of won.

Around September 20th, reports out of Korea and China one after another said that Samsung Electro-Mechanics’ long-term supply contract would exceed ₩4 trillion (~$3.1B). Two days later, on September 22nd, the Austrian semiconductor substrate maker AT&S announced it was expanding its Malaysian plant on the back of a long-term deal with Marvell, the US chip design company. One side makes capacitors the size of a grain of rice; the other makes the substrate that chips sit on. Both are components that go into AI servers, and both are locking in years’ worth of volume through contracts signed well in advance.

The same week, similar chatter was making the rounds in the securities industry. A securities firm’s electronic-components analyst team met with institutional investors in Hong Kong and Singapore from September 14th to 18th and left behind a trip memo. The single most-discussed topic, apparently, was MLCC price hikes. But there’s one odd detail buried in that memo: Murata Manufacturing, Japan’s number one MLCC maker, reportedly said it would “follow if competitors raise prices.” In other words, even with orders piling up, the market leader isn’t willing to move first.

Today I want to connect these three scenes into one story. You don’t need to know what an MLCC is going in — we’ll start right there.

MLCC, the Grain of Rice Inside Every Electronic Device

MLCC is short for Multi-Layer Ceramic Capacitor. A capacitor¹ is a component that holds electricity briefly and releases it when needed. In the industry, people often compare it to a dam. Chips sometimes draw a large burst of current in an instant. If the chip had to wait for power to arrive from a distant source, voltage would fluctuate—so a capacitor sitting right next to the chip releases its stored charge first. Filtering out noise that creeps in through the circuit is also this component’s job.

The “multi-layer ceramic” part of the name describes how it’s made. A metal electrode is printed onto a thin ceramic film, and hundreds of these layers are stacked and fired together. The more layers you pack in, the more charge you can store in the same physical size. The materials involved include ceramic powders like barium titanate and nickel powder. The smaller the size and the greater the number of layers, the harder the manufacturing becomes—and if even one layer has a defect, the entire unit gets scrapped.

A single unit is small and cheap. But staggering quantities go into each device. A single smartphone contains more than 1,000 of them. In June of this year alone, Murata shipped 140 billion MLCCs, and Samsung Electro-Mechanics shipped 98 billion. That’s not a year’s worth—that’s one month. Because it’s used in such enormous volume, the component has earned the nickname “the rice of the electronics industry.”

It helps to think of the market as split in two. General-purpose MLCCs used in smartphones and home appliances have many manufacturers, and customers can switch suppliers easily. Chinese companies are also catching up quickly in this segment. High-spec MLCCs used in AI servers and automobiles are a different story. They need to withstand high temperatures and high voltages while packing large capacitance into a tiny footprint—and only a handful of companies, Murata and Samsung Electro-Mechanics among them, can produce these reliably at scale. Nearly everything happening in the market this year started in this second segment.

Why AI Servers Use So Many MLCCs

It comes down to how GPUs draw power. GPUs repeatedly pull huge amounts of current in short bursts, then release it. Every time that happens, you need a dense array of capacitors standing by to keep the voltage from wobbling. And since servers run hot and operate 24 hours a day without a break, those capacitors also need to withstand higher temperatures and voltages.

The numbers tell the story. According to estimates from the Chinese securities firm CICC, a typical server motherboard uses 1,800–2,500 MLCCs, while an AI server motherboard packing eight GPUs uses 15,000–25,000. For a system like Nvidia’s GB200 NVL72, which bundles dozens of servers into a single cabinet, the estimate runs to roughly 440,000 units. Industry sources suggest next-generation cabinets could need somewhere around 600,000.

It’s not just the count that’s rising. Some analyses put the unit price of AI-server-grade MLCCs at more than three times that of mobile-grade ones. Because they require far more stacked layers, the same factory, running for the same amount of time, can produce fewer of them. So if demand grows tenfold, the strain on a factory’s capacity grows by more than tenfold. That’s why adding capacity doesn’t let manufacturers catch up quickly — even with new production lines, it takes time to get yields right on a process that precisely stacks hundreds of layers.

Goldman Sachs projects the AI-server MLCC market will grow from $1.4 billion in 2025 to $5.8 billion in 2030, expanding 34% a year. That’s still a small slice of the overall MLCC market. But it’s a small slice that’s already shaking up factory utilization rates and pricing across the board.

What happened to MLCC prices this year

Early in the year, things were quiet. In March, Murata notified customers it would raise prices starting April 1, but the target was other passive components like ferrite beads and inductors. The company cited rising silver prices as the cause. In April, electronics trade outlet TheElec reported that the industry was waiting for Murata’s signal on MLCC pricing — conventional wisdom held that once the market leader moved, everyone else would watch customer reactions before following suit.

Demand heated up before prices did. At the end of June, the book-to-bill ratios² for Murata, Samsung Electro-Mechanics, and Taiyo Yuden stood at 1.30, 1.31, and 1.25 respectively — the highest since the pandemic. That meant incoming orders were running 25–31% ahead of outgoing shipments.

It was a Korean company that moved first on price lists. In late July, Samsung Electro-Mechanics told its sales partners it would raise prices by 30% across all MLCC products under certain sales codes, effective for shipments from August 1. Taiyo Yuden, the world’s No. 3 maker, followed on July 23, announcing its own price adjustment effective for shipments from September 1 — and added that even if customers accepted the increase, delivery dates still couldn’t be guaranteed. Analyst notes put the price hikes among non-Japanese makers at 20–50%.

Why Did the Market Leader Say It Would “Follow”?

You’d think if demand outstrips supply, the number one player would be first to raise prices. Murata said the opposite. The reason traces back to something this industry went through a few years ago.

In 2020-2021, as COVID-driven demand for IT devices surged, MLCC prices rose by an average of about 23%. But when demand cooled in 2022 and inventories piled up, prices snapped back quickly—some general-purpose products returned to their pre-hike levels. Component makers had to cut factory utilization and absorb margin pressure, all while facing backlash from device makers. An industry source cited by DealSite (The Elec) said that ever since, MLCC makers have taken a much more conservative approach to pricing—adding that with so many players in the field, it’s hard to create a semiconductor-style supercycle.

This is where the difference between the two markets resurfaces. In the general-purpose segment, if you raise prices alone, customers simply go elsewhere. Whoever moves first absorbs the customer backlash and the loss of share—and the bigger you are, the more volume you stand to lose. But if a competitor raises prices first, the follower can share the backlash under the cover of “the market price went up.” So Murata’s comment looks less like hesitation and more like a deliberate choice about sequencing.

That doesn’t mean Murata is sitting still. There have been reports that it recently sent notices to customers about phasing out production of certain lower-spec products. Rather than touching the price list first, it’s shifting capacity away from low-margin general-purpose lines and toward higher-spec ones. Even without raising prices, if supply shrinks, market prices rise as a consequence.

The analyst team expects that even if Japanese makers do move to raise prices, the increase will land around 20-30%, not above 50%. Investors, too, have reportedly been saying repeatedly that even if prices go up, they should go up gradually.

Samsung Electro-Mechanics’ ₩4 trillion (~$2.9B), broken down

What moved more than price here were the contracts. Starting in May, Samsung Electro-Mechanics has been disclosing a string of long-term agreements (LTA)³.

• May: Silicon capacitors, roughly ₩1.557 trillion (~$1.13B) (2027–2028, 2 years)
• June: MLCCs for AI servers, ₩454 billion (~$330M) (2027)
• July: MLCCs for AI servers, ₩295.1 billion (~$214M) (2027)
• September 1: MLCCs for AI servers, ₩1.0722 trillion (~$779M) (2027) — the largest MLCC long-term contract in the company’s history

Reports tallied the deals disclosed from May through early September at ₩3.32 trillion (~$2.4B). Add in roughly ₩700 billion (~$509M) worth of AI-server MLCC contracts still in late-stage negotiation, and if that closes, the total crosses ₩4 trillion in just four months — that was the gist of the articles circulating around September 20. Some Chinese outlets wrote as if it were already done, saying Samsung had “swept up ₩4 trillion,” but as of reporting on September 22, the deal was still under negotiation.

One more thing worth flagging: not all of that ₩4 trillion is MLCC. The single largest chunk is actually a silicon capacitor⁴ contract. Silicon capacitors sit right beneath the chip package to reduce power loss — functionally similar to MLCCs, but manufactured more like semiconductors. Add up just the disclosed MLCC contracts and you get about ₩1.82 trillion (~$1.32B). Still not a small number — one calculation puts the September 1 contract alone at roughly 21% of Samsung Electro-Mechanics’ entire MLCC business revenue from last year.

The counterparties are shifting too. The June contract was with a server manufacturer, the July contract with a server assembler. iM Securities identified the ultimate buyer behind the September contract as a semiconductor platform company. The counterparty for the ₩700 billion deal still under negotiation is reportedly a global power and electronic components company. In other words, the chase for MLCC supply has moved from server assemblers all the way up to chip designers and power-component makers securing their own stock directly.

Factory conditions are tight too. Industry estimates put Samsung Electro-Mechanics’ MLCC plant utilization in the high 90% range — effectively running at full capacity. In July, the company disclosed plans to invest about ₩15 trillion (~$10.9B) in its Busan plant through 2040, building its first mass-production line for AI-server package substrates and high-value MLCCs, along with an R&D hub. Add the roughly ₩8 trillion (~$5.8B) package-substrate investment announced the day before for its Sejong plant, and the total comes to ₩23 trillion (~$16.7B). Total capex over the past decade was about ₩10.3 trillion (~$7.5B) — so this more than doubles it. That said, the ₩15 trillion is spread across a 15-year plan, and it’s not earmarked for MLCCs alone.

Contracts Cascade Down the Component Stack

If you line up this year’s long-term AI infrastructure contracts layer by layer, a pattern emerges. First, memory got locked in. I covered the contract structure SanDisk disclosed back in August.

SanDisk’s 4-Year Deals, Backed by a $16.5 Billion Guarantee · Issue 183 · INLEVEL9Eight customers demanded 4-plus-year supply contracts and a $16.5 billion financial guarantee upfrontinlevel9.com

Next came MLCCs, and this week it’s package substrates. AT&S announced an expanded partnership with Marvell on September 22. Marvell designs custom AI chips for cloud companies. AT&S had previously announced an expansion of its Kulim, Malaysia plant, mentioning only that there was an additional customer involved — and this week it was revealed that customer is Marvell. The plan is to fill out the interior equipment of the Kulim Plant 2 facility and build new facilities for substrate cores and advanced packaging. AT&S said this expansion is backed by long-term commitments from customers, though it didn’t disclose the contract value.

A package substrate⁵ is the board that connects a chip to the main board. Modern AI chips are increasingly built by bundling multiple smaller chips (chiplets) into a single package, and AT&S explained that this shift is why substrates are getting bigger and gaining more layers. AT&S’s competitors in this market are Japan’s Ibiden and Samsung Electro-Mechanics, which has announced substrate investments in Busan and Sejong. Samsung Electro-Mechanics finds itself at the center of the long-term-contract competition in both the MLCC and package substrate markets.

Line up the three layers of contracts side by side, and the differences come into focus.

	Memory (SanDisk)	MLCC (Samsung Electro-Mechanics)	Package Substrates (AT&S)
Contract term	Weighted average of 4+ years	1 year (through 2027)	Undisclosed
What the customer put up	$16.5 billion financial guarantee	No deposit or advance payment	Undisclosed
Supplier’s move	Said no new plants	Plans ₩15 trillion (~$10.8 billion) investment in Busan through 2040	Expanding plants on the basis of customer commitments

On the memory side, the customer put money on the line, and the supplier held firm on not building more plants. On the MLCC side, the customer locked in only one year of volume and put up no money at all — yet the supplier rolled out a 15-year investment plan. Who’s more desperate shows up right there in the contract terms. That said, this comparison is based only on publicly disclosed conditions; we can’t know what other clauses might be buried in the contracts.

Why Worry About Stock Prices When Prices Are Rising

Another thing that stands out in analyst notes is investor anxiety: MLCC companies might end up following the same path as memory chip makers. Memory companies raised prices sharply—and then saw their valuations⁶ fall anyway.

It sounds counterintuitive, but the logic is simple. When component prices climb steeply, profits surge. But the market often reads a cyclical industry’s rapid price increase as a signal that profits are nearing their peak. Since stock prices don’t rise in step with the profit growth, the price-to-earnings multiple actually shrinks. That’s exactly what happened in August, when Western Digital posted record-high earnings—and its stock still dropped about 16%.

The outlook for MLCCs is similarly steep. Market estimates suggest that the operating profit of Samsung Electro-Mechanics’ MLCC division will jump from roughly ₩600 billion (~$432 million) last year to ₩1.3 trillion (~$936 million) this year, and surpass ₩4 trillion (~$2.9 billion) by 2028. The steeper the numbers climb, the more investors start asking when they’ll turn. That’s where the investor refrain—that prices should go up gradually, not all at once—comes from.

AI Servers Aren’t Just Buying Chips

The name that came up most often after MLCCs in these memos was TDK. They note that investor sentiment is shifting from “a good company” to “a solid, high-quality company.” Back in April–June, a common pair trade⁷ was buying Murata and shorting TDK, but this time more investors are buying both.

The list of things investors are pinning their hopes on all share a similar character: mid-size batteries for data centers, HDD-related components, aluminum electrolytic capacitors, thin-film power inductors, high-voltage/high-temperature MLCCs, and DC-link film capacitors. Most of these are components that store, filter, or stabilize electricity. As AI data centers multiply, orders for these parts follow right behind chips and memory. Apparently there were also plenty of questions about AI-server business opportunities for Nippon Chemi-Con and Nichicon, both makers of aluminum electrolytic capacitors.

Views on package-substrate maker Ibiden were split. Its business outlook looks very strong, but the stock has already priced in results through fiscal year 2030, so on a 1–2 year forward-earnings basis it looks expensive. Many hedge funds had apparently gone short on Ibiden. Even if additional capacity-expansion plans are announced, they won’t start contributing to earnings until fiscal year 2031, and many felt it was too early for the stock to reflect that. AT&S’s deal with Marvell will likely face the same question. It takes time for a factory to actually start running and turn into revenue.

Oswarld’s Lens

I think what this story reveals is less about price than about the contract itself. Pricing power doesn’t come simply from strong demand. It comes from whether customers can walk away once prices rise, and from how long and how much money they’re willing to commit. Murata choosing the order of its price hikes, Samsung Electro-Mechanics moving first on its price list, and customers locking in a year’s worth of volume — these are different answers to the same question.

What catches my attention is the asymmetry in duration. Customers committed to one year; suppliers rolled out 15-year investment plans. In memory chips, suppliers only released volume once customers put down deposits. In MLCCs, it’s still the suppliers carrying the bigger risk. If demand cools the way it did in 2022, we’ll see who ends up absorbing that gap. If the next round of contracts comes with longer terms or upfront-payment clauses, that’s a signal customers have grown more anxious.

If you’re building devices, there’s one thing worth checking: the cheapest part on your bill of materials can be the one that holds up your launch schedule. Taiyo Yuden said that even at higher prices, it can’t guarantee delivery times. With chip designers and even power-component makers now scrambling to secure parts directly, small hardware teams end up at the back of the line. This is the moment to check delivery timelines and whether substitute parts can even pass certification — before worrying about the price increase itself.

One distinction I want to draw out: the analyst memo quoted today is a snapshot of the mood among investors one brokerage met over the course of a week. The ₩4 trillion (~$2.9 billion) figure, too, is a forecast that includes contracts still under negotiation. The trend is clear, but headline numbers are worth pulling apart before you take them at face value.

Closing

A component the size of a grain of rice has landed on the list of AI infrastructure bottlenecks, and even the substrates that chips sit on are now being locked down by contract. Three things to watch next: whether Murata actually changes its price list, whether Samsung Electro-Mechanics’ ₩700 billion (~$506 million) contract shows up in a disclosure, and how the terms and advance-payment conditions of the next contract shift.

---

💬 Has a single component ever delayed a product or project you were working on? Tell me which part it was in the comments.

📨 If someone around you has been asking “why are electronic component prices going up these days,” send them this piece.

---

References & Further Reading

• AT&S, AT&S and Marvell Technology expand collaboration to support next-generation AI infrastructure, 2026-09-22. The company’s own explanation of expanded cooling plans and shifting substrate demand.
• Money Today, Riding the AI Wave Toward a ₩4 Trillion (~$2.9B) Jackpot…Samsung Electro-Mechanics Nears Additional ₩700 Billion (~$500M) MLCC Order, 2026-09-20. Cumulative contract size, utilization rates, and the division’s profit outlook.
• Digital Times, Samsung Electro-Mechanics Nears Another MLCC Deal With Big Tech…On Track to Break ₩4 Trillion (~$2.9B) This Year, 2026-09-20. The nature of the counterparty in the ongoing negotiation.
• Hankyung, Samsung Electro-Mechanics to Invest ₩15 Trillion (~$10.8B) in Busan…Building a Mother Line for Package Substrates and MLCCs, 2026-07-03. The timeline and target of the Busan investment.
• Seoul Economic Daily, Samsung Electro-Mechanics Announces ₩15 Trillion (~$10.8B) Busan Expansion…Total Spend Reaches ₩23 Trillion (~$16.5B), 2026-07-03. The Sejong investment, the past decade of capex, and the silicon capacitor contract.
• eToday, iM Securities: “Samsung Electro-Mechanics Lands Over ₩1 Trillion (~$720M) in MLCC Orders, and Its Pricing Power Just Got Stronger”, 2026-09. A brokerage’s read on the shift in counterparties.
• HuffPost Korea, Samsung Electro-Mechanics Secures ₩1.0722 Trillion (~$770M) Worth of AI-Server MLCC Orders, 2026-09-01. The contract term and upfront-payment conditions.
• Financial Today, Samsung Electro-Mechanics Nears ₩4 Trillion (~$2.9B) in Long-Term MLCC Contracts, 2026-09. Murata’s notice discontinuing certain specs and the difficulty of the process involved.
• TrendForce, Samsung Electro-Mechanics Lifts MLCC Prices 30% Starting Aug. 1, 2026-07-29. The terms of Samsung Electro-Mechanics’ price hike, the book-to-bill ratio, and June shipment volumes.
• The Elec, Samsung Electro-Mechanics to Raise MLCC Prices by 30% Starting in August, 2026-07-30. Taiyo Yuden’s own price-hike notice and comments on delivery lead times.
• The Elec, Murata Seen Setting Direction on MLCC Pricing as Demand Recovers, 2026-04-13. The 2020–2022 price spike and its subsequent reversal.
• TrendForce, MLCC Giant Murata Reportedly Confirms April 1 Price Hike on Key Components, 2026-03-17. The specific items covered by Murata’s March hike and per-smartphone MLCC counts.
• DoNews, AI Servers Driving a Surge in MLCC Demand, 2026. CCID Consulting’s estimates of MLCC content by server type.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Capacitor: A component that briefly stores electricity and then releases it. It stabilizes voltage and filters out circuit noise. In Korean it’s also commonly called a “condenser” in industry parlance. ↩

2. Book-to-Bill ratio (BB ratio): Orders received during a period divided by shipments made during the same period. A ratio above 1 means orders are outpacing shipments — a sign supply is tight. ↩

3. Long-Term Agreement (LTA): Rather than resetting price and volume every quarter, this type of contract locks in terms and quantities for a year or more in advance. ↩

4. Silicon Capacitor: A capacitor built on a silicon wafer using semiconductor fabrication processes. Because it can be made extremely thin, it’s mounted directly beneath a chip package and used to stabilize power around high-performance AI chips and HBM. ↩

5. Package Substrate: A board that electrically connects a chip to the mainboard, spreading out the chip’s densely packed connection terminals to a pitch the mainboard can accommodate. High-spec versions for AI chips are typically what’s called FC-BGA. ↩

6. Valuation: A measure of how a company’s worth is priced relative to its earnings or assets. The price-to-earnings ratio (PER) — share price divided by earnings per share — is the most common example. ↩

7. Pair trade: An investment strategy that buys one of two similar stocks while short-selling the other, betting on the relative performance gap between them. ↩