How to Select Capacitors for AI Server Power Stages

Selecting a capacitor for an AI server power stage starts with three numbers: the rail’s voltage, its power or ripple current, and the ambient temperature it runs in. Once you have those, match the rail’s role to the right capacitor type—aluminum electrolytic for 380 V rectifier and 48 V intermediate rails, polymer or MLCC for low-voltage high-current rails like 12 V and 1 V, and supercapacitors or electrolytics for hold-up—then size for ripple and lifetime. Get the order right and the spec follows.

This is the fourth article in our AI server power series. We’ve covered why AI server power demand is exploding, how capacitors work in switching power supplies, and which capacitors fit each power stage. This piece turns that allocation into a selection method.

Start With the Three Numbers of the Rail

Every power rail in an AI server gives you three selection inputs: working voltage, power or ripple current, and operating temperature. The 380 V rectifier bus, the 48 V intermediate bus, the 12 V VRM input, the 1 V GPU core rail, and the hold-up circuit each present a different combination of these three, and that combination decides the capacitor’s voltage rating, ripple capability, and lifetime grade.

An AI server’s rails typically look like this:

Voltage sets the floor for the capacitor’s rating—always derate, especially on high-voltage rails. Power translates into ripple current, which the capacitor must carry without overheating. Temperature decides whether you need an 85°C standard series or a 105°C long-life part; AI servers run hot and continuously, so 105°C is the safer default.

Match the Rail to the Capacitor Type

Once you know the rail’s numbers, its role in the power chain tells you which capacitor type to start with.

380 V rectifier bus: aluminum electrolytic

The 380 V rail holds bulk energy after rectification. Aluminum electrolytics give you the capacitance-per-volume needed to smooth that bus at a reasonable size and cost. Voltage rating should sit well above 380 V with margin for transient overshoot.

48 V intermediate bus: aluminum electrolytic or polymer

The 48 V bus feeds the downstream converters. If ripple current is moderate, aluminum electrolytics work. If ripple is heavy or ESR must stay low across temperature, polymer electrolytics are the better pick—they hold ESR stable and handle ripple better than standard aluminum at the same rating.

12 V and 1 V rails: MLCC

Low-voltage, high-current rails like the 12 V VRM input and the 1 V GPU core need low ESR and fast transient response. MLCCs deliver both. At 1 V and several hundred amps, even a few milliohms of ESR turns into heat and voltage droop, so MLCC arrays are standard here.

Hold-up circuit: supercapacitor or aluminum electrolytic

Hold-up circuits bridge the gap when mains power drops, typically 16-20 ms until backup or shutdown kicks in. Supercapacitors pack more energy per volume than electrolytics at the same voltage, so they suit tight spaces. Electrolytics work if space allows and cost matters more.

This mapping assumes you’ve already read which capacitors fit AI server power stages—that piece explains why each type suits its position. Here we’re turning that logic into a selection step.

Then Size for Ripple and Lifetime

After you’ve matched the rail to a capacitor type, three parameters pin down the part: capacitance, ESR, and lifetime rating.

Capacitance from ripple current

Ripple current flows through the capacitor at the switching frequency. The higher the ripple and the higher the frequency, the more capacitance you need—or the lower the ESR must be—to keep voltage ripple within spec. For a 1 V rail at 300 A and tens of kilohertz switching, an MLCC array in parallel brings both the capacitance and the low ESR the rail needs.

ESR matters most on low-voltage, high-current rails

At 1 V and 300 A, even 2 milliohms of ESR drops 600 mV and dissipates 180 W. That’s why GPU core rails use MLCC arrays—their ESR stays in the sub-milliohm range. On higher-voltage rails like 48 V, ESR still matters for ripple heating, but the voltage headroom is bigger.

Lifetime: 105°C long-life for continuous operation

AI servers run at high temperature continuously, not in duty cycles. A capacitor rated for 5,000 hours at 105°C will outlast one rated for 2,000 hours at 85°C in this environment. Specify long-life series upfront; the cost delta is smaller than a field failure.

Xuansn manufactures aluminum electrolytic capacitors, polymer capacitors, and supercapacitors rated for these roles. Our aluminum electrolytic range covers 380 V and 48 V rails, our polymer series suits ripple-heavy intermediate buses, our MLCC catalog includes low-ESR parts for sub-2 V rails, and our supercapacitor line handles hold-up at 48 V and above.

Selection Quick Reference

Here’s the decision path compressed: identify your rail’s voltage, power, and temperature, then map it to the table below. The recommended type and key parameters follow.

For a detailed walk-through of capacitor types and their trade-offs, see our complete guide to choosing a capacitor.

Common Selection Mistakes

Three mistakes show up often enough to call out.

Looking only at voltage rating. A capacitor rated at the right voltage but undersized for ripple current will overheat and fail early. Check the ripple spec at your actual switching frequency—datasheets often quote it at a reference frequency that isn’t yours.

Using an 85°C part in a 105°C environment. Derating rules might say it’s acceptable, but lifetime drops fast with temperature. AI servers don’t have the thermal budget to tolerate short-life capacitors; specify 105°C long-life upfront.

Underestimating ESR on low-voltage rails. At 1 V, ESR isn’t a second-order effect—it’s the dominant loss and droop mechanism. If you’re seeing voltage sag or unexpected heating on a GPU rail, ESR is usually the culprit.

Common Questions About Selecting Capacitors for AI Power

How do I size capacitance for a 1 V rail at 300 A?

Start from the allowable voltage ripple and the switching frequency. For a rail at 1 V with a few tens of millivolts ripple budget and switching in the hundreds of kilohertz, you’ll need an array of MLCCs in parallel to hit both the capacitance and the sub-milliohm ESR target. The exact count depends on the per-part ESR and capacitance, but dozens of MLCCs in parallel is common for high-power GPU rails.

Should I use aluminum or polymer electrolytics for the 48 V bus?

If ripple current is moderate and cost matters, aluminum works. If ripple is heavy or you need stable ESR across temperature swings, polymer is better—it handles ripple without the ESR drift aluminum shows as it heats. Check your ripple current against the datasheet rating at your operating temperature and frequency.

Hold-up: supercapacitor or electrolytic?

Supercapacitors pack more energy per volume, so they suit designs where space is tight. Electrolytics cost less and integrate more easily if you have the board area. Both work for 16-20 ms hold-up; the choice comes down to space and cost trade-offs in your specific layout.

This selection method builds on our earlier pieces: why AI server power demand is exploding, how capacitors work in switching power supplies, and which capacitors fit AI server power stages. For AI-specific applications beyond server power, see our coverage of AI server EMI filtering with feedthrough capacitors, 48 V and 800 V data center power with aluminum electrolytics, eVTOL DC-link film capacitors, and supercapacitor UPS for AI server racks.