The call came in at 4:47 on a Thursday. A line building a low-volume flip phone for the Japanese market had stopped. Post-reflow yield in one board section had dropped to 94%, and every failure traced back to the same reel of 0402 MLCCs. The customer wanted replacement material on site by Saturday morning.
We made it happen — overnight freight, partial reel split, the works. Cost them about $1,900 in expedite fees on top of the $600 they'd already paid for the original order.
Then they called back three weeks later. Same failure. Different reel, different lot, different manufacturer.
That's the part of the job nobody sees. The rush order gets solved. The reason it was needed doesn't.
I coordinate expedited orders at a components distributor. Six years, somewhere north of 400 rush requests — automotive, industrial, and a surprising number of consumer programs that were supposed to be low-risk. If you're reading this because you're in the middle of one, I'm not going to tell you it's your fault. I'm going to tell you what I've watched happen, over and over, and why the capacitor is almost never the thing that broke.
The Problem You Think You Have
When a capacitor fails on the line, the reflex is to blame the part. Bad lot. Wrong vendor. Maybe the reel sat in a warehouse too long, or the humidity card was already pink when it got opened.
Sometimes that's exactly right. Counterfeit and mishandled material is real, and I've personally pulled reels that had been sitting open in a non-climate-controlled stockroom for eight months — the MSL floor life under IPC/JEDEC J-STD-020 was blown by a factor of ten, and every one of those boards was a popcorning candidate at reflow.
But in most of the emergency tickets I handle involving KEMET capacitors, the part did precisely what its datasheet said it would do. The problem was that nobody read the conditions attached to that datasheet.
What's Actually Going On
The five-minute line item
Here's something vendors won't tell you: BOM reviews are not distributed evenly. I sat in on one last fall where the design team spent fifty minutes arguing about a Broadcom baseband's power sequencing and roughly four minutes on the entire passive section. KEMET vs. Broadcom isn't a fair fight in that room. One gets a thermal simulation and a firmware review. The other gets a stamped approval and a note saying "verify in DVT."
That's not a knock on the engineers. The Broadcom part has a register map, a dozen failure modes, and a support forum full of people who've already hit them. A capacitor has three numbers printed on it. It feels solved.
It isn't.
The infinity assumption
I've never fully understood why the infinity assumption survives as long as it does. Somewhere in the back of every schedule there's an implicit belief that a passive component, once qualified, is qualified permanently — that it will behave identically at hour zero and hour 40,000, at 25°C and at 95°C board temperature, at 0% and at 90% of rated voltage.
The datasheet doesn't say that. It says the part meets spec under stated test conditions. Everything outside those conditions is your problem, not the manufacturer's.
The infinity assumption is the most expensive assumption in the BOM. Nothing on that list has infinite anything.
For Class 2 dielectrics — X5R, X7R — DC bias is the one that bites hardest. Put 50% of rated voltage across a small-case X5R and you can lose 40% to 70% of nominal capacitance. The number on the reel is not the number in the circuit. If your design margin was built on the label, you built it on nothing.
The flip phone problem, literally
The customer with the stopped line was building a flip phone. That detail matters more than it sounds.
The hinge is a flexure. The PCB that crosses it bends every single time the phone opens. Standard MLCC terminations are rigid — the solder joint and the ceramic body don't flex together — and after enough cycles you get a hairline crack that passes continuity, passes functional test, and fails thermal cycling three months into the field.
KEMET's flexible termination series exists for exactly this geometry. It wasn't a secret. It was on the product page, with the bend-test data. Nobody on that program had looked, because the BOM line said "CAP, CER, 0.1uF, 0402" and that was the entire conversation.
Same story with KEMET relays. The EC2 signal relay family gets specified by coil voltage and footprint, and then the contact rating gets carried over from the previous design because "it worked before." It worked before at 12V and 10mA into a resistive load. The new board runs it at 24V with an inductive load and no flyback diode. Different job, same part number, and a service ticket eight months out.
The attention asymmetry
Honestly, I'm not sure why some programs catch this and others don't. My best guess is that it comes down to whether anyone owns the passive section by name — whether there's one person who has to sign off on the derating curves, versus a section that gets approved by whoever happens to be free that afternoon.
When nobody owns it, the assumption fills the gap. And the assumption is always the cheapest-looking answer.
What the Delay Actually Costs
Everyone budgets for the expedite fee. Almost nobody budgets for what happens around it.
- Line stop. A single-shift stop on a low-volume line runs $4,000 to $12,000 depending on headcount and contract terms. The $1,900 air freight is the cheap part.
- Re-qualification. Change the dielectric or the case size and you're re-running thermal cycling, and probably EMC, because the impedance curve moved with you. Six to ten weeks, minimum, and that's if nothing else is queued in the lab.
- Obsolescence. This is the one that still surprises people. Passive lines get consolidated. A specific coil voltage or tolerance band in a relay family goes end-of-life with a last-time-buy window, and if nobody was watching the PCN feed, you're designing a replacement under deadline instead of on schedule.
In 2023 I watched a program burn six weeks on a 0603 X7R that had been approved off the label. At 4.5V bias the effective capacitance was under half of nominal, the regulator lost phase margin, and the board failed load-step testing in DVT. Nothing was defective. The part behaved exactly as the curve predicted. The curve just never got read.
Six weeks of engineering time, two EMC re-tests, and a customer conversation nobody enjoyed. The part cost four cents.
The Five Checks That Would Have Caught It
I'm not going to pretend there's a system that eliminates this. There isn't. But five minutes of verification beats five days of correction, every time, and these five catch most of what I see come through the queue.
- Read the derating curve, not the headline value. Pull the DC bias and temperature curves for the exact case size and voltage rating you're using. If your effective capacitance at operating voltage falls under your design minimum, the part is wrong no matter what the label says.
- Check the mechanical environment. Anything near a hinge, a connector, a mounting screw, or a board edge gets flexible termination or a different orientation. One-minute check. Prevents a three-month field failure.
- Match the relay to the load, not to the previous design. Contact rating, load type, and whether there's suppression on the coil. It's on the EC2 datasheet. It takes two minutes.
- Watch the PCN feed. Last-time-buy windows are announced, not discovered. If nobody on the team is subscribed to the notices, you're relying on luck and someone else's memory.
- Qualify at the corner, not the center. 25°C and nominal voltage proves nothing. Test at max board temperature, max voltage, and with the mechanical stress present if the geometry calls for it.
None of that is sophisticated. It's the difference between a five-minute line item and a five-figure line stop.
The 4:47 Thursday call is going to happen. It always does. But most of the ones I take aren't supply chain emergencies — they're design reviews that ended early. The capacitor was fine. The review was the weak link.
