The Hidden Bottleneck in AI Servers Is a Tiny Ceramic Part

The part nobody is talking about

Everyone watching the AI buildout is staring at GPUs, interposers, and HBM memory. The real bottleneck may be hiding in plain sight: a ceramic capacitor no larger than a grain of rice.

Samsung Electro-Mechanics disclosed on September 1 that it signed a 1.07-trillion-won ($780 million) long-term supply agreement with an unnamed US AI semiconductor company for MLCCs — multi-layer ceramic capacitors — intended for AI servers. It is the single largest MLCC contract in the company’s history and amounts to 9.5% of its full-year sales. The deal covers January through December next year.

The nameless buyer is almost certainly one of the major AI chip designers whose servers are saturating every data center on the planet. Samsung will not confirm which one. That silence itself tells you something: in this market, buyers are not advertising their supply deals, and suppliers are equally quiet.

AI servers eat capacitors

MLCCs are the most-used passive component in electronics. They store and release electricity, acting as tiny reservoirs that smooth voltage spikes. Every circuit board needs thousands of them. An AI server needs roughly ten times as many as a conventional server, according to industry estimates cited by Samsung’s disclosures and corroborated by Goldman Sachs research.

The math is brutal and simple. A single AI training cluster can contain tens of thousands of GPUs, each surrounded by power delivery networks that demand stable, clean current. Any fluctuation causes errors. MLCCs near the processor absorb those fluctuations. More GPUs means exponentially more capacitors, not linearly.

The AI server MLCC market was worth approximately $1.4 billion last year. Goldman Sachs projects it will reach $5.8 billion by 2030, growing at 34% annually. That is not a speculative forecast from a vendor trying to sell capacity. It is an independent bank’s reading of order books.

Who can actually make these

The constraint is not demand. It is supply. High-reliability MLCCs for AI servers must operate continuously at elevated temperatures and voltages without degrading. Manufacturing them requires precision co-firing processes, exotic ceramic materials, and yield rates that most foundries never achieve because most foundries do not make capacitors.

Globally, the number of suppliers qualified to deliver AI-server-grade MLCCs at volume is small. Samsung Electro-Mechanics, Murata, and a handful of Chinese and Japanese producers compete in this tier. The Chinese suppliers face export controls and technology restrictions that limit their ability to serve Western AI firms. Murata is the main rival, but Samsung’s disclosed contract pipeline this year alone — over 10 long-term agreements totaling 3.32 trillion won — suggests it has pulled ahead in allocation.

Samsung’s order-to-shipment ratio stands at 1.5, meaning new orders are outpacing deliveries by 50%. Inventory has fallen below 30 days of supply. The company is running its factories at above 95% utilization. This is not a healthy surplus. It is a shortage signal.

Why Korea is winning where chip design cannot

The deeper story here is structural. The AI hardware narrative has been dominated by two camps: American chip designers and Taiwanese foundries. Both are indispensable. Both are also crowded, expensive, and increasingly constrained by geography and geopolitics.

Korean component makers occupy a different lane. They do not design processors. They make the physical infrastructure that allows those processors to function. Power delivery, thermal management, high-frequency interconnects, and passive components — these are unglamorous categories. They are also non-negotiable.

Samsung Electro-Mechanics’ progression this year reads like a playbook. In May, it won a 1.56-trillion-won silicon capacitor contract. In June, a 454-billion-won MLCC deal. In August, a 295-billion-won contract. And now the 1.07-trillion-won AI server MLCC agreement. Each one locks in volume for months or years. Each one raises the bar for competitors who lack production scale.

Murata has responded by expanding capacity in Japan and Thailand. Chinese producers like Sunlord and Fenghua are pushing into the high-reliability segment, but their access to Western AI firms is structurally limited by US export policy. That leaves Samsung and Murata as the primary beneficiaries of the AI server buildout, and Samsung’s contract funnel this year suggests it is capturing a disproportionate share.

The pricing cycle has just begun

Samsung already raised MLCC prices once this year. Industry sources indicate a second increase is likely before year-end, driven by the same供需 imbalance that produced the first round. High-end AI server MLCCs are particularly prone to price escalation because their manufacturing complexity does not scale easily. You cannot simply add another production line and expect yield to hold.

This pricing power is rare in components. Most passive-component makers compete on cost. Samsung is currently competing on allocation. When your buyer cannot source the part anywhere else, you set the terms.

The contract structure itself also signals a shift. These are no longer spot purchases. They are long-term agreements — annual commitments that lock in volume and price well in advance. That protects Samsung from downturns but also ties it to the AI cycle. If data center spending decelerates, those contracts become expensive obligations. If spending accelerates, they become profit machines.

The current trajectory points toward the latter. Major cloud providers have collectively committed hundreds of billions to AI infrastructure through 2027. None of those systems ship without MLCCs. None of those MLCCs will come from China-based suppliers for Western buyers. That leaves a narrow funnel, and Samsung is sitting in it.

What happens next

Expect Samsung to announce additional capacity expansions at its plants in Busan, Sejong, and Vietnam. The company has already committed capital to those sites. The question is whether new lines can come online fast enough to close the gap between orders and deliveries.

The answer matters beyond Samsung. Murata’s stock, TDK’s stock, and even the valuation multiples of AI chip designers all partially depend on whether the passive-component layer can keep pace with the active one. A GPU is useless without stable power. A power delivery network is useless without capacitors that do not fail under thermal stress.

The AI hardware stack is only as strong as its weakest component. Right now, that component is a ceramic disc no bigger than a fingernail, and Samsung Electro-Mechanics is the closest thing the market has to a kingmaker.