AI demand now squeezing MLCCs and optics

The new bottleneck in artificial intelligence data centers is not just the graphics processor or the memory chip. It is also tiny parts like multilayer ceramic capacitors, which smooth power, and optical parts, which turn electrical signals into light between racks. (murata.com) (coherent.com) A multilayer ceramic capacitor is basically a stack of ceramic and metal sheets that acts like a tiny shock absorber for electricity. Murata says these parts support stable operation in communications infrastructure, industrial equipment, and vehicles, which is why they sit all over server boards and power circuits. (murata.com) Artificial intelligence servers use far more of those parts than ordinary servers because their power draw changes fast when graphics processors switch workloads. Samsung Electro-Mechanics said in February 2026 that artificial intelligence servers use more than 10 to 15 times as many multilayer ceramic capacitors as general-purpose servers. (samsungsem.com) Murata gave a similar picture at its December 2025 investor briefing. The company said the latest artificial intelligence accelerator boards commonly install 15,000 to 25,000 multilayer ceramic capacitors. (murata.com) That is why a shortage in this market looks strange from the outside. Consumer-grade capacitors can still be relatively available, while high-capacitance server and automotive parts are the ones getting tight, because the hard-to-make end of the catalog is where artificial intelligence and cars are both buying. (trendforce.com) (fusionww.com) The supplier names keep repeating because the market is concentrated. Mordor Intelligence lists Murata, Samsung Electro-Mechanics, Taiyo Yuden, Yageo, and TDK among the major multilayer ceramic capacitor makers, so pressure at a few companies can ripple through a lot of hardware programs. (mordorintelligence.com) The pricing pressure is no longer just rumor. TrendForce reported on February 24, 2026 that Samsung Electro-Mechanics was weighing multilayer ceramic capacitor price increases starting in April, with stronger demand from artificial intelligence and electric vehicles supporting higher average selling prices in the second quarter. (trendforce.com) Optics are the second squeeze point. Optical transceivers are the modules that let data center switches and servers send data as pulses of light through fiber, and artificial intelligence clusters need huge numbers of them because thousands of chips have to talk to each other at once. (coherent.com) (lightcounting.com) That market is already scaling into very large numbers. LightCounting said in its January 2025 outlook that optical transceivers for artificial intelligence clusters would grow from $5 billion in 2024 to more than $10 billion in 2026, driven by new Nvidia cluster designs that require a lot of transceivers. (lightcounting.com) Lumentum has been unusually explicit about how full that pipe is. In its February 3, 2026 earnings call, the company said “virtually every AI network” uses its technology either through direct hyperscaler partnerships or through equipment makers, and in March 2026 it said it was expanding products for scale-out, scale-up, and scale-across artificial intelligence infrastructure. (fool.com) (lumentum.com) Lumentum’s own component roadmap shows why lead times can stretch before the end customer notices. The company said in February 2026 that it is shipping 200 gigabits per second per lane electro-absorption modulated lasers at scale for 800 gigabit and 1.6 terabit transceivers, and that production capacity for those lasers was expanding year over year to meet hyperscaler demand. (lumentum.com) Put together, the pattern is simple: artificial intelligence spending has moved from the obvious parts of the bill of materials into the hidden ones. When each server needs tens of thousands of capacitors and each cluster needs a wall of optics, shortages stop being a memory story and start looking like a full-stack hardware story. (murata.com) (samsungsem.com) (lightcounting.com)