Samsung's capacitor maker declares its own Moore's Law, and AI buyers are paying for it

A supplier names its own law

At the Open Compute Project APAC summit in Taipei last month, Samsung Electro-Mechanics proposed "MLCC's Law"↗: the capacitance required per GPU doubles every year ¹. Moore had transistors, and Jensen Huang claims a "Hyper Moore's Law" for whole systems; now a maker of rice-grain-sized parts claims the pattern for power ¹. It is a demand forecast from a supplier, and buyers are underwriting it with contracts.

Four months, four trillion won

Since May, the company has disclosed long-term agreements worth about 3.32 trillion won for silicon capacitors and AI-server MLCCs, with more than ten global customers ²³.

The centerpiece, disclosed on 1 September 2026↗, is 1.072 trillion won, about 780 million dollars, from a single unnamed customer and covering all of calendar 2027: the largest MLCC contract in company history, worth 9.5 percent of a year's revenue ²³⁴. Industry observers believe the client is one of the world's major AI semiconductor companies ³.

Before that deal the disclosed book stood at about 2.25 trillion won; with the record contract it reaches the 3.32-trillion-won total reported the same day ³⁵. Roughly 700 billion won more is in final negotiation with a global power and components company↗, which would take the four-month tally past 4 trillion won ⁶⁷. Korea's securities industry sees component-division operating profit doubling this year to as much as 1.3 trillion won ⁶.

Why buyers sign at triple prices

AI servers carry more than ten times the MLCCs of standard servers, at prices at least three times mobile grade ⁶⁷. TrendForce expects the high-capacity units on next-generation AI server boards to jump from 1,440 to 10,544, and certain MLCCs on NVIDIA's next Vera Rubin platform to rise from 320 to 500 a board ⁷.

Rack power is heading from today's 100 to 120kW toward a forecast 1MW ¹. Capacity is effectively sold out: utilization is nearing the high 90s percent ⁶, high-capacitance lead times run about 21 weeks as of August, five weeks longer than at the start of the year ⁸, and June book-to-bill hit 1.31, a post-pandemic high ⁹. TrendForce expects AI-server X6S prices to rise another 10 to 20 percent in the fourth quarter, an estimate the company has not confirmed ⁸.

Two suppliers, so buyers pay the premium

The bill lands downstream. Rack builders and equipment OEMs absorb the triple pricing and the 21-week waits, plus a forecast 25 to 30 percent on consumer-grade parts, because only a handful of manufacturers can make components that survive high temperature and voltage around the clock ⁵⁷⁸. Of those, only Samsung Electro-Mechanics and Murata are realistically positioned to meet AI demand ⁵.

A forecast wearing Moore's clothing

MLCC's Law is a supplier's claim, not physics. The company says it holds more than 40 percent of the AI-server MLCC market, split with Japan's Murata ³⁵. Murata is cutting part numbers across nine series, with final orders to March 2028↗, to free capacity for higher-margin work; industry sources say the cuts will not hand Samsung direct orders, because the discontinued lines are not the ones it is chasing ⁹.

Goldman Sachs sizes the AI-server MLCC market at 5.8 billion dollars by 2030, up from 1.4 billion in 2025 ²⁵. Real growth, but a rounding error beside GPU budgets. If board designers route around the constraint, the law dies quietly.

Watch Busan

The bet behind the law is about 15 trillion won through 2040 for a Busan production and R&D hub; embedded MLCCs and low-ESL ceramics are slated for next year ¹⁷. The tell: whether the 700-billion-won agreement closes, and whether Vera Rubin boards ship at 21-week lead times. The AI boom spent two years paying the chipmakers; the toll has moved to a part smaller than a grain of rice.