Last time we covered Analog Chips. The week before, selective etching. This week the bottleneck is a capacitor.
A GB300 system contains tens of thousands of multilayer ceramic capacitors. At rack scale, that becomes hundreds of thousands.
Individually, they cost almost nothing. Collectively, their value per system is rising sharply as accelerator power density increases.
Morgan Stanley's teardown puts MLCC content at $1,530 per GB300 rack and $4,320 per VR200 rack, almost an increase of 182%.
But the interesting part is not that "MLCCs are in shortage."
Most MLCCs are not.
Consumer-grade parts remain well supplied and the price increase is soft. The constraint sits in a much narrower category of high-capacitance, low-inductance parts that can survive the voltage, thermal and power-delivery requirements of AI accelerators.
That is the constraint we are looking at this week.
What is an MLCC?
A capacitor stores electrical charge and releases it on demand. On a GPU board that job is unglamorous and completely non-negotiable, because an accelerator's power draw can swing by hundreds of amps in microseconds and something has to absorb the shock.
Without it the voltage sags and the chip browns out mid-calculation.
A multilayer ceramic capacitor (or MLCC) does this by stacking. Thin sheets of ceramic (the dielectric i.e. the insulating material that actually holds the charge) alternate with layers of nickel, the electrodes that carry current in and out. The whole thing is printed, pressed and fired into a single solid block about the size of a grain of sand. The more layers you fit inside that block, the more charge it holds in the same footprint.
Composition of an MLCC module (Source: Elecinsight)
That stacking is the entire engineering race, and it has been pushed remarkably far. High-end parts that are required for AI accelerators today run a thousand or more layers, each ceramic sheet under a micron thick, fired at temperature without warping the block or letting the nickel react with the ceramic beside it. One bad layer out of a thousand and the part is scrap.
What Actually Got Constrained
The shortage is not in ceramic capacitors generally. The MLCC market is splitting in two.
At the commodity end, supply remains ample. Consumer electronics demand is weak enough that aggregate MLCC pricing can still look soft.
At the high end, the picture is the opposite.
AI accelerators require capacitors with unusually high capacitance in a very small footprint, with low ESL (a measure of how much the part fights a fast change in current, where lower is better when the load swings in microseconds) and that stay stable under heat and voltage.
These parts sit throughout the accelerator's power-delivery network, where they absorb rapid changes in current and keep voltage stable as compute load moves.
Building them consistently at scale is difficult. Qualification makes the supply base narrower still.
The result is that suppliers are moving manufacturing capacity away from lower-value consumer grades and toward AI and automotive specifications.
Only a handful of companies build that part at production volume:
Murata (Japan)— roughly 45% of AI-server-grade supply.
Samsung Electro-Mechanics (South Korea) — roughly 40%.
Taiyo Yuden and Kyocera (both Japan) — qualified at the grade, dividing what little is left.
What Demand Looks Like
You do not need to rely on one indicator to see the squeeze. Several independent measures are moving in the same direction.
Lead times have stretched: High-capacitance grades that previously quoted in weeks are now quoting in months, with some specifications difficult to source at all.
Factories are running close to full: High-end server capacity across all three major suppliers is running at 90% utilization, with Murata near 95%, while shipments have reached multi-year highs.
Orders are still arriving faster than product can leave the factory: Taiyo Yuden's capacitor book-to-bill reached 1.7 in the quarter reported on August 5, up from 1.3, meaning orders were running materially ahead of sales. Orders rose 41% sequentially and backlog rose 78%. Murata's orders-to-backlog ratio has already passed its 2018 peak, the worst shortage in the industry's modern history.
And pricing power has returned: After years in which MLCC economics mostly moved toward lower prices and better performance, suppliers are now pushing through increases on high-end grades. Murata took AI-server and high-end automotive grades up 15–35% in the spring. SEMCO raised its entire book 30% effective August 1, no specifications exempted. Taiyo Yuden follows on September 1
Together, they point to something harder to dismiss: the AI-grade segment is consuming qualified capacity faster than suppliers can add it.
Demand is leading tightness in the MLCC constraint (Source:Tessara)
Why Supply Cannot Catch Up Quickly
The constraint is difficult to solve because adding "MLCC capacity" is not the same thing as adding the specific capacity AI systems need:
New lines take years, not quarters: High-end MLCC manufacturing requires extremely precise printing, stacking and firing across hundreds or thousands of microscopic layers. The equipment base is concentrated, installation takes time, and new output still has to pass customer qualification. That makes the effective supply response much slower than simply ordering another production line.
And the powder underneath is its own oligopoly: High-purity dielectric powders, nickel powders and other materials used in advanced MLCCs come from a relatively concentrated upstream supply chain. Barium titanate especially, which is one of the chemical compounds used to make the ceramic layer, is the single largest material cost in a high-capacitance MLCC, at 35–45% of the total, and the top five suppliers control most of it, with Sakai Chemical dominant at the high end.
So when AI demand accelerates, suppliers cannot immediately redirect unlimited commodity capacity into server-grade product.
Therefore, the bottleneck is not MLCC capacity. It is qualified high-end capacity.
And that is still growing much more slowly than demand.
Tessara currently reads the MLCC constraint as Tight and still tightening, based on the combination of utilization, lead times, order growth, supplier capacity plans and downstream AI demand.
MLCC Constraint Page on Tessara. Currently sits in the Tight region
But the constraint score itself is less important than what sits underneath it.
Tessara treats MLCC as a live research object rather than a static thesis: which evidence supports tightness, which companies are exposed, whether the evidence is strengthening or weakening, and what would cause the view to change.
This article is the snapshot. The MLCC page on Tessara is the model that keeps moving after publication.
Who Captures the Economics?
A constrained component does not automatically make every supplier a good investment.
Three things matter separately:
1. How directly the company's revenue is exposed to the constrained product.
2. Whether tightness is converting into price, mix and margin.
3. How much of that improvement the market has already priced in.
That produces three different setups across the major suppliers.
Bottleneck Holders of the MLCC constraint as shown on Tessara
Murata: the strongest strategic position but still spending heavily to defend it.
Murata holds the largest share of AI-grade MLCC supply and has substantial exposure to the constraint. The operating improvement is already visible, but management is also committing heavily to new capacity.
That creates the key tension in the stock. The company owns the chokepoint, but a meaningful portion of the cash generated from it is being reinvested into capacity that takes time to come online.
The ¥80bn set aside for data-center MLCC lines produces nothing until late 2027, and management is funding a build phase rather than harvesting one.
Watch: whether capacity expansion begins to outrun demand growth, or whether utilization stays high even as new lines ramp.SEMCO: the strongest recent earnings conversion, with more expectations already embedded
This is no longer a story about future pricing power. Q2 gross margin reached 24.4% against 20.3% a year ago, operating margin 12.7% against 7.6%, and operating profit doubled year over year.
The company has also locked the demand side down, signing MLCC long-term agreements with more than ten customers including hyperscalers, with roughly KRW 745bn of AI-server LTAs disclosed in July alone.So the investment question is no longer whether tightness shows up in earnings. It is how much further the economics can improve against a share price that has already run several times the group.
Watch: pricing realization on the LTAs, and the Calamba plant in the Philippines. Contracted volume the company cannot physically supply is not revenue.Taiyo Yuden: the cleanest expression, both ways
More of Taiyo Yuden's revenue depends on capacitors than either Murata or SEMCO, making it the cleanest directional bet on continued tightness.
That creates greater upside if high-end MLCC pricing and mix continue improving, but also greater sensitivity if the constraint normalizes.
The latest order data are particularly important: capacitor book-to-bill has moved close to 1.72, showing demand arriving faster than current sales.
Watch: whether that order imbalance translates into mix improvement and margins as more capacity shifts toward AI and automotive grades.
The Contradiction
There is one awkward piece of evidence in this thesis. Aggregate Korean MLCC export prices are falling. If the market is constrained, why?
"If aggregate MLCC pricing is falling, what evidence still supports a Tight read for AI-grade MLCCs, and what would falsify it?"
Research Tab on Tessara. Confirms that the commodity mix is diluting AI-grade pricing when looking at exports.
The answer is that the aggregate data are mixing two markets moving in opposite directions.
Commodity MLCC pricing remains weak enough to pull the headline number lower. At the same time, suppliers are reallocating capacity toward higher-spec AI and automotive grades, where utilization, orders and lead times remain elevated.
That means falling average selling prices do not yet contradict the AI constraint.
But the thesis is falsifiable. Any of the following would matter:
High-end lead times normalize despite continued AI-server growth.
Supplier book-to-bill falls sustainably below 1.
New X6S/X7R capacity qualifies faster than expected.
AI rack deployment slips enough to release component allocations.
Silicon capacitors or other architectures displace more MLCC content than expected.
Those are the indicators that matter more than aggregate MLCC pricing. They are also what Tessara watches as the constraint evolves.
So, the MLCC story is therefore less about a shortage of capacitors than a shortage of qualified capacitance in the places AI systems increasingly need it.
For now, the evidence still points toward tighter high-end supply, continued mix improvement for the major suppliers, and a widening gap between AI and commodity MLCCs.
But that conclusion should change when the evidence does.
Tessara continuously tracks the MLCC constraint, the companies exposed to it, the evidence behind the current read, and the signals that would cause it to reverse.
The Week Ahead
Wednesday, Aug 19
Analog Devices (ADI) — Analog Semiconductors: We wanted data center growth holding its pace. It did, and it beat: revenue $4.02bn, up 40% and $110m above consensus, EPS $3.45 against $3.34, with optical and power up more than 100% inside the segment.
Wolfspeed (WOLF) — SiC Power Devices: AI data center revenue grew ~20% sequentially, exactly as asked. It made no difference: revenue missed consensus by $74m, gross margin stayed at −19.9%, shares fell 17%.
Read our post and pre-call briefs here and stay prepared.
See where the AI buildout goes next
This issue, we named high-capacitance MLCCs as the chokepoint and two companies, Murata and Samsung Electro-Mechanics, as the ones that control the grade AI servers actually need.
In Tessara terminal, you can track the MLCC constraint and the barium titanate powder chain running beneath it, and see which of 400+ public names are most exposed to the squeeze.
“From extensive sector and company deep dives to thoughtful market updates, it has become a go-to resource for our investment team” - Portfolio Manager, Titan Global Capital Management
See you next week,
Teng & Arvind
This article is for informational and research purposes only. It is not financial advice, investment advice, or a recommendation to buy or sell any security. Tessara Research does not publish price targets. The views expressed here reflect our analysis at the time of publication and may change as new evidence arrives. Readers should do their own research and consult a qualified financial adviser before making investment decisions.