Every conversion stage burns power as heat, takes board space and adds a failure point. So why not pick one voltage and carry it from the utility transformer all the way to the motherboard? The answer is a conflict between two hard limits: the resistance of copper across a room, and the electrical resistance of the human body inside a cabinet.

The Rack and the Room Want Opposite Voltages

When AI server power demand pushed individual cabinets past 30 kW and toward 100 kW, the legacy power delivery model broke in two places at once. Inside the cabinet, the old 12 V distribution bus required busbars so thick they blocked airflow and still dropped too much voltage before reaching the compute trays. Out on the data hall floor, conventional three-phase AC distribution or 400 V DC feeds required massive overhead cable trays to feed a 1 MW cluster of ten racks.

Fixing both problems meant pushing voltage higher, because higher voltage moves the same watts at lower current. Yet the room and the rack stopped at completely different numbers. Facility engineers are pushing room-level distribution toward 800 V DC, borrowing heavily from electric vehicle charger and traction inverter architectures. Rack architects, on the other hand, moved from 12 V up to 48 V DC and stopped right there.

This creates a horizontal fork in the power tree. Rather than a smooth, uniform rail running through the building, the architecture pulls apart: upward in the facility hall to survive distance, and downward at the rack boundary to survive human contact. Every intermediate DC-DC conversion stage in a modern AI data center exists to bridge the gap between what the room wiring requires and what the rack mechanics allow.

The Current Math That Forces Each Choice

Basic Ohm’s law dictates why neither 12 V nor 48 V can leave the rack, and why 800 V dominates the room. Power equals voltage times current (P = V × I). When power stays fixed at 100 kW for a single high-density rack or 1 MW for a ten-rack row, raising the voltage cuts the current in exact proportion.

Current reduction is only part of the picture. Resistive heating in a conductor follows the square of the current (P = I²R). If two systems use a copper cable or busbar of the exact same cross-sectional area and length, their resistance (R) is identical. Cutting the current by a factor of four reduces the heat lost in that conductor by a factor of sixteen.

Look at the gap between 48 V and 800 V. For the same conductor size, distributing power at 48 V generates 278 times more resistive heat than distributing it at 800 V. Inside a two-meter tall server rack, a thick vertical copper busbar can carry 2,083 A over a short distance without melting or costing a fortune. Running 20,833 A at 48 V across thirty meters of data hall floor to feed a 1 MW cluster calls for a conductor closer to an industrial smelting plant than to a server hall. Only 800 V DC brings cluster-level current down to 1,250 A, a number that fits into standard overhead busways with manageable thermal loss.

48 V Is a Safety Limit, Not an Efficiency Optimum

If 800 V cuts copper losses by 278 times compared to 48 V, a reasonable question is why rack designers stopped at 48 V instead of running 100 V or 200 V down the backplane. The reason has nothing to do with electrical efficiency. It is a legal and physiological boundary.

Under international safety standard IEC 62368-1, which governs information and communication technology equipment, Energy Source Class 1 (ES1) sets the ceiling for touch-safe direct current at 60 V DC under normal operating conditions. Below 60 V DC, dry human skin offers enough resistance that accidental contact does not drive lethal current through the chest, and service personnel can work around exposed conductors without high-voltage personal protective equipment or strict lock-out/tag-out procedures.

A nominal 48 V DC system does not stay at exactly 48.0 V. When backup battery strings or rectifiers float on the bus, the actual line voltage typically sits around 52.8 V, and transient surges or regulation tolerances push it slightly higher. Choosing 48 V nominal leaves a safe margin below the 60 V DC ES1 hard ceiling.

At 10 kW or 15 kW per rack, 12 V meant roughly 1,000 A, workable with heavy power shelves. At 100 kW it demands 8,333 A. No practical connector can hot-swap a blade at that current without massive contact resistance, and the copper weight alone would exceed floor loading limits. Moving to 48 V cut rack current by a factor of four and I²R busbar losses by a factor of sixteen, right up to the highest voltage permitted for live human maintenance.

Why 800 V Only Works Outside the Rack

Outside the cabinet, nobody hot-swaps bare busbars by hand while the system is live. Room-level power runs inside enclosed overhead busways, armored conduits, and locked power distribution cabinets. Because fingers are kept out by grounded metal enclosures, the 60 V DC touch-safe rule no longer applies, freeing facility engineers to chase the copper math up to 800 V DC.

Bringing that 800 V rail directly into the server compute tray, however, hits immediate mechanical and physical walls:

• Creepage and clearance distances expand dramatically above 60 V. On a dense server motherboard or backplane, 800 V requires wide physical gaps between copper traces to prevent surface tracking and air breakdown, especially as dust and humidity accumulate over years of service.
• Direct current does not cross zero volts the way alternating current does. If a technician pulls a live 800 V DC blade from a backplane, the resulting electrical arc will not self-extinguish; it will sustain a plasma flame capable of vaporizing contacts and causing severe arc-flash injuries.
• High-voltage connectors require deep shrouding, sequenced safety interlock pins, and bulky insulation housings. In a 1U or 2U compute sled where every millimeter of rear panel width is needed for liquid cooling manifolds and high-speed optical transceivers, 800 V connectors simply take up too much space.

So 800 V stays outside the rack, or enters only a sealed, dedicated power shelf where qualified power electronics convert it down to 48 V before it ever touches the open vertical rack busbar.

What Each Voltage Does to the Capacitors

Splitting the building into an 800 V room domain and a 48 V rack domain changes the electrical stress at every conversion boundary. Capacitors sit at each of these boundaries to absorb switching ripple, hold up the rail during brief grid drops, and feed sudden load spikes. Because voltage and current invert as power moves from the wall to the silicon, the capacitor technologies must shift at every step. For a longitudinal look at how a single rail steps down, see our guide on which capacitors fit each power stage.

At the AC-to-800 V DC rectifier stage (or active front end), the primary challenge is high voltage combined with continuous switching ripple current and bulk energy storage requirements. A single 800 V bus often splits into two 400 V halves in series or uses high-voltage stacks. Here, engineers rely on large can XLW screw-terminal capacitors rated at 400 V to 500 V per cell, alongside heavy-duty polypropylene DC-link capacitors. The main constraint here is internal self-heating from ripple current: if the equivalent series resistance (ESR) is too high, core temperatures climb past 105 °C and dry out the electrolyte prematurely.

At the 800 V-to-48 V isolated DC-DC converter stage, power crosses the galvanic safety barrier that separates the dangerous room bus from the touch-safe rack bus. These high-density resonant (LLC) converters operate at hundreds of kilohertz to shrink the isolation transformer. The input side needs high-voltage, low-inductance capacitance that fits into a low-profile power shelf, while the output side suddenly faces thousands of amps of combined current at 48 V. You can read a deeper breakdown of the role of aluminum electrolytic capacitors in 48 V and 800 V data center power on our companion technical site.

Once power lands on the 48 V rack busbar and enters the server tray, the capacitor job shifts to local energy buffering. Snap-in and conductive polymer hybrid aluminum electrolytic capacitors rated at 63 V to 100 V sit at the tray input and around the 48 V intermediate bus converters (IBC). Their task is to keep the 48 V rail stiff when all eight GPUs in a node slam from idle to full load at the same millisecond, preventing voltage droop from tripping under-voltage lockouts.

Finally, at the point-of-load (POL) stage right next to the processor package, voltage drops below 1.0 V and current surges into the thousands of amps per chip. Electrolytic cans are far too slow and inductive here. This domain belongs entirely to multi-layer ceramic capacitors (MLCCs) and low-profile conductive polymer tantalum or aluminum chips. At this millimeter scale, parasitic inductance and microsecond transient response govern everything, a dynamic we covered in detail under GPU VRM decoupling.

Where the Conversion Losses Stack Up

A common objection to the 800 V-to-48 V architecture is that it still uses three major conversion blocks from the AC utility feed to the processor core: AC to 800 V DC, 800 V DC to 48 V DC, and 48 V DC to sub-1 V POL. Why does this beat the older AC to 400 V DC to 12 V to POL chain?

The difference comes down to converter physics at higher voltages and the elimination of massive distribution losses between the stages. Typical conversion efficiencies across the two chains fall in the range of the figures below (note that these are typical architectural benchmarks, not measured test data from a specific vendor):

A jump from 89.4% to 92.7% looks modest on paper until you apply it to a 10 MW data hall. That 3.3 percentage point gain saves 330 kW of continuous waste heat—and saves another 50 kW to 100 kW of chiller and pump power that would have been required to remove that heat from the building. The 800 V-to-48 V intermediate stage hits 98.5% typical efficiency because wide-bandgap semiconductors (SiC and GaN) switch 800 V with very low conduction loss, and outputting at 48 V instead of 12 V cuts secondary rectifier currents by a factor of four.

What This Means for a 2026 Build

For hardware teams and power supply buyers specifying equipment in 2026, the rack-versus-room split sets clear boundaries on component sourcing. Most halls in service still feed the rack from AC distribution; the 800 V room bus is where new high-density builds are heading rather than a retrofit every operator has already made. If your rack density stays below 30 kW, conventional three-phase AC distribution into the rack with internal 48 V rectifiers remains practical and avoids facility-wide DC retrofits. Once single-rack power passes 50 kW and clusters scale to 1 MW and beyond, keeping AC or 400 V DC in the room forces unacceptable copper cross-sections and voltage drops.

On the component side, power supply BOMs are polarizing into two distinct voltage classes. On the high-voltage side of the rack boundary, engineers need 450 V to 500 V rated screw-terminal aluminum electrolytics and 800 V to 1,200 V DC-link films designed for continuous high-ripple operation at 105 °C. On the low-voltage side inside the cabinet, the old 16 V and 25 V capacitor bins used on 12 V rails are obsolete for new designs, replaced by 63 V, 80 V, and 100 V rated snap-in and conductive polymer parts that give safe derating margins on a 48 V bus.

Where This Sits in the AI Power Series

Our previous articles in the AI power series traced why compute power is rising, how switching converters rely on energy storage, how a single voltage rail steps down from 380 V to 1 V across four capacitor layers, and how GPU decoupling responds across nanosecond-to-microsecond time windows. This article looked horizontally across the data center floor to explain why the facility room and the server rack refuse to meet at a single voltage, and how that 800 V / 48 V split shapes the converters and capacitor banks standing between them.

Common Questions

Can 48 V power a GPU directly without another conversion stage?

No. Modern GPU cores operate at roughly 0.6 V to 1.0 V because lower voltage reduces dynamic switching losses inside nanometer-scale transistors (which scale with V²). Applying 48 V directly to a processor core would instantly punch through the microscopic gate oxides and destroy the silicon. The 48 V bus must step down through intermediate bus converters and multiphase point-of-load voltage regulators right on the GPU board before entering the package.

What is the biggest obstacle to bringing 800 V directly inside the server rack?

Human maintenance safety and arc-flash physics are the primary blockers. Above 60 V DC, equipment exceeds the IEC 62368-1 ES1 touch-safe limit, meaning technicians could no longer hot-swap server blades or work inside an open cabinet without high-voltage protective gear and strict lock-out procedures. In addition, 800 V DC requires wide creepage distances that waste board space and creates sustained DC arcs if a connector is pulled under load.

How does the 48 V / 800 V split change capacitor requirements compared to older 12 V servers?

It pushes voltage ratings upward at both ends while increasing ripple current density. Inside the server tray, capacitors that used to sit on a 12 V bus (rated at 16 V or 25 V) must now survive a 48 V nominal bus (requiring 63 V to 100 V ratings) with low ESR in compact packages. At the room and power-shelf level, rectifiers and DC-DC converters feeding an 800 V DC bus require series-stacked 450 V/500 V screw-terminal aluminum electrolytics or high-voltage DC-link film capacitors capable of handling continuous kilowatt-level ripple currents without overheating.

Can an existing AC data center room be retrofitted to 800 V DC?

Yes, typically by installing sidecar rectifier cabinets at the end of each server row rather than ripping out the entire building transformer infrastructure. Facility AC cables feed the row-end rectifier cabinet, which converts three-phase AC into 800 V DC and sends it down a dedicated overhead busway to the high-density AI racks in that row. However, existing AC breakers, transfer switches, and metering equipment cannot be reused on the DC side because DC fault interruption requires specialized arc-quenching switchgear.

If you are evaluating screw-terminal aluminum electrolytic capacitors, snap-in banks, or DC-link film capacitors for 800 V power shelves and 48 V server architectures, our engineering team can help match ripple current, ESR, and lifetime specs to your thermal profile. Reach out through our contact page or email us directly at coco@xuanxcapacitors.com with your operating voltage and ripple requirements.

The post Data Center Power Architecture: 48 V Inside, 800 V Outside appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.

What holds the rail up in the gap is capacitance. Deciding where that capacitance goes means splitting the load step into time zones, because each zone has a different dominant parasitic and a different way of failing when under-provisioned.

For the wider picture of how capacitor families divide work across an AI server power chain, the four-capacitor overview runs the rail from 380 V down to 1 V. This piece stays on one rail and goes deep.

A load step outruns the regulator

Take a boost converter feeding a GPU core rail. A workload transition pushes demand from 200 A to 600 A in a few hundred nanoseconds. The VRM senses the droop, its error amplifier responds, the inductor current ramps. Between the demand and the delivery, something has to hold the rail.

How far it may sag is a design decision rather than a datasheet constant. Most accelerator designs work to two to five percent around nominal. On a 0.9 V rail, three percent is 27 mV. That number drives every sizing decision below.

Target impedance sizes the array

Convert the window into a specification with target impedance:

Z_target = ΔV_allowed / ΔI_step

At 27 mV and a 600 A step, the PDN has to look like 45 µΩ or less wherever the step has energy.

Forty-five micro-ohms is hard to reach with discrete parts. One 1210 MLCC has ESR in the low milliohms and ESL near 1 nH, and its own impedance bottoms out well above the target. The array works because many parts in parallel divide both, and because the parts are spread across frequency. Layering gets the curve under the line. A single large capacitor does not, which is the whole reason the array exists.

Four time zones on one rail

Planning a GPU decoupling array by time rather than by capacitance value keeps the parasitics honest.

Below 100 ns: inside the package

In the first tens of nanoseconds the current loop never leaves the die and its package. On-chip capacitance and the substrate planes supply the charge. Board capacitors are excluded by the inductance of the path between them and the die.

Package and interposer design get so much attention for this reason. Nothing soldered to the PCB changes what happens here. Vertical power delivery, with the regulator underneath the die rather than beside it, exists mostly to shorten this loop.

100 ns to 2 µs: the board MLCC array

Once the demand outlives the package, the board array takes over. Most of the capacitor count lives in this band, and most of the design arguments happen here. Ceramic parts win on ESL and ESR, and they can sit within a few millimetres of the die.

Two numbers drive the count: the charge the array must hold, and the parallel impedance it must present. The charge figure is easy to underestimate. Supplying 600 A for one microsecond moves 600 microcoulombs, and holding a 27 mV window while doing it takes roughly 22 mF at the die:

C = I × t / ΔV = 600 A × 1 µs / 27 mV ≈ 22 mF

Twenty-two millifarads in ceramic means more than 700 pieces of 1210 at an effective 30 µF each, on one rail. Nobody builds that. The arithmetic earns its place by showing how the job actually splits: regulator bandwidth shortens the window, and the array covers whatever remains. Treat the bank as the sole energy source and the board stops routing.

2 to 50 µs: polymer and solid electrolytics

Past a couple of microseconds the regulator is awake and the inductor current is climbing, though the loop is still catching up. Polymer tantalum and solid aluminum parts work in this band. They carry far more capacitance per unit area than ceramic, they take the ripple current, and their ESR is low enough to matter at these lower frequencies.

Misplacing parts between this band and the ceramic band is common. A polymer part where ceramic belongs adds bulk without lowering the high-frequency impedance. Ceramic where bulk belongs leaves the low-frequency end unsupported, which shows up as a slow sag under sustained load rather than a fast transient.

Above 50 µs: the regulator and its bulk input

At millisecond scale the VRM is in control. What remains downstream is bulk capacitance that rides through input transients and keeps the regulator’s own supply steady. Aluminum electrolytics still do that work on the input side, picked for capacitance per dollar and ripple handling rather than speed.

You get less capacitance than you order

A BOM that says 47 µF and an impedance curve that behaves like 30 µF is ordinary on a low-voltage rail. On a 0.9 V GPU rail, DC bias does most of the damage.

DC bias on Class II ceramic

Class II dielectrics such as X5R and X7R use barium titanate, and its permittivity falls as the applied field rises. This is the material behaving as specified, not a bad batch. Published curves show small-case parts losing over half their nominal value at a few volts, and losses past 80 percent are recorded for high-capacitance-density parts running inside their ratings.

Case size changes the shape of the curve. More dielectric volume per unit capacitance, as in a 1210 rather than an 0603, usually holds value better at the same voltage. Board area is tight on accelerator cards, and large case sizes survive anyway for this reason.

At 0.9 V the applied field stays far from the rating, so bias derating is milder there than it would be at 5 V. The trap sits on the 48 V intermediate bus and the 12 V rails feeding the regulators, where a part chosen by a blanket fifty percent voltage rule can lose most of its value. Read the bias curve for the exact part number and case size.

Ageing and temperature drift

Class II ceramics also lose capacitance with time after the last heat treatment, and recover on reflow. The drift is logarithmic and flattens, so a part measured on the bench and the same part measured a year later on a card differ by several percent. Temperature adds another shift on top. Neither moves the design as much as bias, and both explain why field measurements rarely match the BOM.

The anti-resonance peak nobody budgets for

Mixing capacitor values sounds like it broadens the impedance curve. It creates a peak instead.

Between the two self-resonant frequencies, the larger capacitance has turned inductive while the smaller one is still capacitive. In parallel they behave like a tank circuit, and impedance climbs at the frequency where that happens. A PDN can pass its target at low frequency, look clean at high frequency, and miss in the band between. If the load current has harmonic content there, the rail sees more noise than the average figure suggests.

Keeping the spread between neighbouring values modest helps, and so does using many parts of one value rather than a wide ladder. Where a peak cannot be avoided, add damping. Simulation is the only reliable check, because the peak position depends on layout parasitics rather than on the capacitor list.

Choosing the mix

Below is a starting point for a GPU rail in the 500 to 1000 A class. The values reflect what the time zones call for, not a prescription for any particular board.

For the rail-by-rail version of this table, including the 380 V bus and the 48 V intermediate stage, the selection guide for AI server power stages covers the same ground with the regulator topology included. If the term itself is new, the decoupling capacitor basics page starts further back.

What changes at 1000 W

Current density comes first. A card at 1000 W on a 0.9 V core rail pushes more than a thousand amperes through a plane, and resistive drop across that plane stops being a second-order concern. Copper thickness, via count and the number of parallel layers all move into the budget.

Space moves the other way. Capacitor count grows while the footprint beside the die shrinks, and accelerator boards are already crowded with decoupling. That pressure is what pushed vertical power delivery from a research idea into shipping products.

For the capacitor supply chain both trends point one direction. Demand shifts toward low-ESR, high-reliability parts in small cases, and toward polymer and solid types that hold their ratings under heat. Sourcing them at consistent quality is the less glamorous half of the problem. Xuansn builds aluminum electrolytic, film, solid and supercapacitor lines to customer specification, and the ceramic and solid ranges are catalogued under the multilayer ceramic and solid capacitor categories.

Where this sits in the AI power series

This is the GPU power rail installment of a series that began with why AI server power demand is rising. Earlier parts covered how capacitors work inside a switching supply, which families fit which rail, and how to size them by ripple and lifetime. The fastest end of the problem, where the regulator is too slow and the package too inductive, was the one left open.

Next comes the supply side of the same coin. The shift toward 48 V distribution and vertical delivery changes what a power designer asks a capacitor supplier for, and that changes how high-capacitance, low-ESR parts are quoted and scheduled.

Common questions

How many decoupling capacitors does a GPU need?

No fixed number exists, and any figure quoted without the rail voltage, allowed ripple and regulator bandwidth is guesswork. Size the array from the target impedance and the time window the regulator needs, then check the result against the board area you have. Once derating is applied, high-current core rails usually land in the hundreds of pieces.

Why use many small capacitors instead of one large one?

Impedance decides this, and large capacitors carry more ESL. A single large part has a low self-resonant frequency and looks inductive long before the frequencies a GPU load step occupies. Parallel small parts divide the inductance along with the capacitance, and they sit closer to the load.

Do I need to derate MLCC capacitance for DC bias?

On any Class II dielectric, yes. Use the bias curve for the exact part number and case size at the operating voltage rather than a percentage rule. The effect is smaller on a 0.9 V rail than on 12 V or 48 V, and it is never zero.

Can polymer tantalum replace MLCC in decoupling?

Not at the high-frequency end. Polymer parts carry more capacitance per volume and handle ripple well, which makes them useful in the microsecond band and on rails where bulk matters. Their ESL and ESR remain higher than ceramic, so they complement an MLCC array. The ESR explainer covers the trade-off in more detail.

Working on a GPU rail and need the capacitance side solved? Send the rail voltage, the load step and the space you have. Xuansn builds aluminum electrolytic, film, solid and supercapacitor lines to specification and quotes against your BOM.

✉ coco@xuanxcapacitors.com  |  ☎ +86-769-8166 8821  |  Send your specification

The post GPU VRM Decoupling: What a 600 A Load Step Does to the Rail appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.

This article closes our 2026 sourcing series, which started with why capacitor prices rose, moved through locking in supply before allocation, and then looked at what can and cannot be substituted. Those three were about what a buyer can do once the shortage arrives. This one sits earlier: what has to be true in the design file for sourcing to stay possible two years from now.

Three Things Decide Whether a Capacitor Can Be Replaced

By the time a distributor sends an allocation notice, the answer is usually already fixed. Three constraints decide it, and all three are set at design review.

The three are not equal. Any of them can block a swap on its own, and the mechanical one is the most expensive to undo after the fact.

Footprint Is the Lock That Rewrites the Board

Lead pitch is the easiest detail to get wrong. Radial lead electrolytic capacitors come with 5.0 mm or 7.5 mm lead spacing, and the two do not share a hole pattern. A board laid out for 5.0 mm can still take a 7.5 mm part if someone bends the leads, but on an automated insertion line that is a different process, not a different part. On a hand-built prototype nobody notices, which is exactly why the problem survives to production.

Snap-in parts have their own version of the same trap. The lead pitch is normally 10 mm, with two leads on smaller diameters and four leads on cans from roughly 30 mm upward. Two leads and four leads are not interchangeable by drilling two more holes, since the landing pattern and the mechanical anchoring differ. Screw-terminal capacitors add another layer: the stud is M5 or M6 depending on the series, and even two cans of the same nominal size can have different stud spacing and terminal shapes between makers.

Designing for a replaceable part does not mean designing for every part. It means writing constraints that more than one maker can meet:

• Give the specification a maximum diameter and height with a margin, for example 5 percent on diameter and 3 mm on height, instead of a fixed case code.
• Keep a standard lead pitch unless the space genuinely cannot take it.
• State the terminal form (two-lead snap-in, four-lead snap-in, M5 screw) as a requirement, and check which approved makers offer something in that form.

A footprint with a small margin costs nothing at layout time and is very hard to add later.

The Parameter Window You Write Becomes Your Supplier List

An ESR limit is the most common self-inflicted constraint. If comparable parts on the market sit at 25 to 30 mΩ and the specification asks for a maximum of 20 mΩ, the supplier list drops from a dozen manufacturers to one or two, and those two know it. In practice the limit is rarely deliberate. It usually comes from copying the specification of a part that was already difficult to source, and the restriction is inherited several designs later. The same happens with tolerance. Write ±10 percent where ±20 percent would work, and a large part of the industry is excluded before the first enquiry goes out.

Derating buys flexibility as a side effect. A 48 V rail can run on a 63 V part at roughly 1.3 times derating, or on a 100 V part at just over 2 times. The second option opens a much wider group of suppliers. Derating is chosen for life and reliability in the first place, and the supply benefit comes along with it.

The capacitance and voltage combination decides how scarce the part is. A 470 µF 450 V snap-in is a hard combination to source from many makers at once. The same bus can sometimes run on two 220 µF 450 V parts in parallel, or on a lower capacitance if the ripple budget allows. Both routes widen the supplier list immediately. They need board space and a current-sharing check, so they have to be considered while the layout is still open.

Certification and the Approved Vendor List

Safety capacitors show the problem most clearly. X and Y class parts are certified to IEC 60384-14, or its EN and UL equivalents, and the certificate is issued to a specific manufacturer’s part rather than to a set of parameters. Two capacitors can match on capacitance, voltage and dimensions and still not be interchangeable, because the second one carries no certificate. Changing maker means a new certification, which is measured in weeks rather than days.

Automotive adds a second layer on top: AEC-Q200 validation plus the customer’s production part approval process, followed by whatever the customer’s approved vendor list requires. None of this is avoidable, and none of it can be shortcut once a shortage is already running.

The practical response is to treat certification as a design-stage task. Where a position depends on a safety or automotive approval, the goal is two named, certified makers in the release documentation, not one maker and an intention to find another.

Writing a Specification That Leaves Room

The difference between a replaceable part and a locked one is often just how the requirement is phrased. A specification that names a parameter set with an interface and a compliance level can be met by several factories. One that names a competitor’s part number can be met by one.

A workable shape for a bulk capacitor requirement looks like this:

• Capacitance: 470 µF ±20 percent
• Voltage: 450 V, applied at or below 80 percent of rated in normal operation
• ESR: 35 mΩ maximum at 100 Hz, 20 °C
• Ripple current: 2.5 A minimum at 105 °C, 100 Hz
• Case: snap-in, 10 mm lead pitch, maximum 35 mm diameter by 50 mm height
• Compliance: EN 60384-14 X2 where the position requires it, two approved manufacturers

Nothing in that list is loose. Every line is a real requirement, and a qualified engineer can still compare parts across makers with it. That is the point: a specification built from parameters gives the buyer something to negotiate with, while a specification built from a part number gives the maker all of the leverage.

What to Ask Before the Design Review Sign-Off

These questions are worth asking while changes are still cheap:

• If this line is allocated until the second quarter of next year, what is the board allowed to change?
• Which positions already have a second, certified manufacturer? Named parts, not options that could be explored.
• Where is the specification drawn tighter than the industry norm, in ESR, tolerance, temperature or ripple?
• Does any footprint accept only one maker’s case size?
• How much free space is left around each capacitor, in diameter and height?
• Can this position be split into two smaller parts without hurting the electrical behaviour?
• Which certificates bind a part to a single manufacturer?
• What does the customer’s approved vendor list allow today?

A design that answers these before release does not need a second source during a shortage. It already has one.

Where Flexibility Is Not Worth Paying For

Not every position should be widened for sourcing. High-frequency and timing circuits are the clearest case, because capacitance and ESR feed directly into how the circuit behaves, and a substitute that measures close on a bench can still shift performance in the field. Snubber and resonant positions sit in the same group. Very small portable designs run out of space before they run out of supplier options.

For those positions the practical move is different. Instead of loosening the specification, qualify a second maker against the tight one and put a part through validation before it is needed. Most filtering, energy storage and bulk decoupling positions have more room than designers assume, and family choices often do as well: once the electrical role is fixed, several capacitor types can usually cover it, and our guide to capacitor types sets out where each one fits. The effort belongs there first.

The Part of the Problem the Buyer Cannot Fix

A price cycle, a lead-time chart and a substitution table all assume the same thing: that the design has room to move. Where it does not, none of it helps. The constraints that decide how much room exists are set long before a buyer sees a shortage, and they sit in the design file rather than in the purchase order. Those constraints are the half of sourcing that no amount of negotiation can reach.

Questions About Capacitor Second Sourcing

Can a second source share the same safety certificate?

No. Certificates for X and Y class capacitors are issued to a specific manufacturer’s part, so a second maker needs its own certificate even when the parameters match. Plan the certification lead time into the schedule, not just the sample lead time.

How much extra board space does this actually cost?

Very little in most cases. The cost sits in the constraints rather than in extra parts: a diameter and height margin, a standard lead pitch, a specification that is not drawn at the industry limit. Where a capacitance and voltage combination is genuinely scarce, splitting it into two parallel parts is the option that widens supply, and that one does need space on the board.

We are already in production. Is it too late?

Not too late, but the work is different. Rank the lines by risk, looking at single-sourced positions, long lead times and specifications that are difficult to meet, then start with the top two or three. For those, qualifying a second maker against the existing board is usually faster than redrawing the part.

Do I add a second part number, or a second maker for the same part?

Both approaches are used. A second maker for the same specification leaves the board untouched and is the cheaper route. A second part number, such as a different case size or capacitance, needs layout space and its own qualification, and it is the fallback when the original combination cannot be sourced elsewhere.

Building a capacitor specification that has to survive the next allocation cycle? Send us the parameters and the position the capacitor sits in. Our engineers will come back with the parts we can hold to that window, and with the second sources worth listing beside them. Xuansn Capacitor manufactures aluminum electrolytic, film and supercapacitor parts.

✉ coco@xuanxcapacitors.com | ☎ +86-769-8166 8821 | Send your specification

The post Designing for a Capacitor Second Source: Where Supply Flexibility Is Won or Lost appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.

Often you can. Substitutions that hold up in the field are the ones treated as engineering decisions rather than purchasing shortcuts. The capacitance printed on the label rarely decides whether a swap is safe. The electrical stress the part actually sees matters more, along with the environment it runs in and whether the replacement was ever qualified for that job.

This is the third article in our 2026 supply series. The first explained why prices are rising, and the second covered how to lock in supply before allocation. This one deals with the part you cannot get at any price you are willing to pay.

Substitution Has Three Levels, and Only One of Them Is Cheap

Buyers usually talk about substitution as a single action. In practice it splits into three, and the effort climbs sharply as you move down the list.

Most shortage-driven substitution projects start at level one because it looks free. That is also where the quiet failures happen.

Level One: A Different Maker, Same Specification

Two capacitors can carry identical capacitance, voltage and temperature ratings and still behave differently on the board. The numbers that govern performance are quoted under conditions that rarely match yours. A ripple current rating measured at 120 Hz tells you little about a part running at 60 kHz. An ESR figure quoted at 100 kHz is not the ESR you will see at 400 Hz.

Before signing off a same-type swap, pull these six parameters side by side and compare them at your operating point, not at the datasheet’s reference point.

The DC bias column deserves a second look if you are substituting multilayer ceramic capacitors. A 10 µF X7R capacitor measured at 1 kHz and zero bias is not a 10 µF capacitor at its rated 25 V. Depending on the case size and dielectric, the effective capacitance can fall by a third or more, and the curve is specific to each maker’s formulation. Two parts with the same part number suffix and the same nominal value can deliver different effective capacitance in the same socket.

Level Two: Crossing to a Different Capacitor Type

When no supplier can ship the right type, the next question is whether a different type can do the job. Some crossings are routine. Others look workable on paper and then fail on the bench. A few are prohibited outright.

Ceramic to film

This crossing makes sense in high-voltage and high-capacitance positions, where a film capacitor offers stable capacitance across voltage and temperature and a much higher ripple current capability. What you pay for it is board area and cost per microfarad. Film capacitors also come in a different footprint family, so a drop-in replacement is rare outside through-hole positions that were designed with spare room.

Aluminum electrolytic to film

In DC-link and snubber duty this is a well-established route, and it is the crossing most often proposed during a shortage of aluminum electrolytic capacitors. Film parts tolerate high ripple current, hold their capacitance for decades, and do not dry out. The trade is physical: matching the capacitance of a large electrolytic with film technology can take several times the volume, and the cost per microfarad rises accordingly. Where the design has the space and the service life matters more than the bill of materials, it works. Where the enclosure is already closed, it usually does not.

Aluminum electrolytic to polymer or solid

Moving to a conductive polymer part improves ESR by roughly an order of magnitude and removes the electrolyte drying mechanism that limits electrolytic life. The ceiling is capacitance and voltage. Polymer parts cover low-voltage rails well and become scarce above a few hundred microfarads or above roughly 100 V. For a low-voltage, ripple-heavy rail, this is one of the better swaps available. For a 450 V bulk position, it is not an option.

The parts that never cross over

X and Y safety capacitors sit in a different category. Their ratings come from safety agency approval tied to the specific part and construction, not from a datasheet parameter you can compare. Substituting one changes the certification basis of the product, and no amount of electrical similarity substitutes for the paperwork. The same caution applies to any capacitor that appears in a certified safety or automotive function.

Level Three: Change the Design, Not the Part

Sometimes the fastest route to a shippable board is to stop looking for a substitute and change what the circuit demands. Splitting one large bulk capacitor into several smaller ones in parallel spreads the ripple current across more parts, which lets you build the same capacitance from components that are actually in stock. The trade is board area and assembly cost, and it changes the ESR and ESL network, so the loop response needs a look.

A second option is to derate. Running a part well below its rated voltage and temperature extends life and widens the pool of acceptable substitutes, because a 63 V part used on a 24 V rail has more margin than the same part used at 48 V. This costs nothing in components and is often the simplest way to make a constrained design buildable.

The third option is topological. Designs that use a single large capacitor for hold-up or filtering can sometimes be rearranged so that the energy storage sits on a lower-voltage rail where parts are plentiful. That is a genuine redesign with a real schedule cost, and it belongs in the discussion only when the first two levels have failed.

Why Waiting Is Not a Strategy in 2026

The supply picture behind this series has not improved. High-capacitance MLCC production carries a structural penalty that does not respond to price signals quickly. A standard 0402 part at 1 µF needs roughly 50 dielectric layers. A 10 µF part in the same case size needs around 500 layers, and a 20 µF part needs over a thousand. Stacking that many layers multiplies the labour and material consumed per unit, so a line that produced 50,000 standard parts a day produces fewer than 10,000 high-capacitance parts a day after conversion. Industry analysis of MLCC capacity published in 2026 put the conversion loss at roughly five to one.

Yield compounds the problem. High-capacitance parts for AI server duty have proven difficult to ramp, with reported yields on those lines running below 60% during early shipments to a GPU maker and taking the better part of a year to reach 80%. The equipment that makes the thinnest dielectric layers is scarce as well. High-end casting machines come from a small number of suppliers and carry lead times measured in many months, which is why new high-capacitance capacity is planned in years rather than quarters. Capacity announced by a leading manufacturer in 2026 will not begin producing until late 2027 at the earliest, with a second phase running into 2029.

The practical read: high-capacitance parts remain tight through at least the first half of 2027, and demand projections for AI server capacitors keep rising year over year. Treating substitution as a temporary workaround misreads the timeline. Many approved vendor lists will stay wider than they were before 2026.

How to Qualify a Substitute Before You Commit

Skipping qualification converts a supply problem into a field failure problem, which costs more in every dimension. The checks below are the ones that catch the differences a datasheet comparison misses.

Start with electrical verification at the operating point rather than the datasheet point. Measure capacitance, ESR and impedance across the frequency range the circuit actually uses, and confirm the part meets the ripple current requirement with the temperature rise you measured, not the one the catalogue assumes. For ceramic parts, measure effective capacitance at the DC bias the rail applies.

Then environmental. Run the candidate through the temperature range the product will see, including the extremes the datasheet ratings imply rather than the comfortable middle. Where the application is safety-related or long-service, ask for load life data at the actual operating temperature and check whether the endurance curve supports the product’s service window.

System-level checks come next, and they catch what bench testing misses. Switching regulators that were stable with one capacitor can oscillate with another, because ESR and ESL set the compensation network’s behaviour. Conducted and radiated emissions can shift. Inrush behaviour on a bulk capacitor changes with ESR, which can trip protection circuits on startup.

Finish with a pilot build rather than a full release. A single production lot exposes handling, placement and process issues that no bench test will reveal, and it keeps the exposure small if something surfaces.

What a Wider Approved List Costs You

Every substitute added to the approved vendor list carries a maintenance cost. More part numbers mean more incoming inspection, more inventory lines, more supplier relationships to manage, and more chances that a purchasing decision made under pressure quietly changes the electrical behaviour of a product. The saving is real. So is the overhead.

The sensible split is between parts worth dual-sourcing permanently and parts where substitution is a one-time bridge. Components with long lead times, single-source risk or high value per unit earn a permanent second source and a documented qualification file. Commodity parts that happen to be temporarily constrained are better handled by a short-term deviation with an expiry date, so the approved list does not accumulate substitutes nobody revisits.

Where the Series Goes Next

Every substitution is a repair. The cure sits upstream at the schematic: how much ripple margin to leave, how far to derate the voltage, and whether to name an approved alternate before a shortage forces the question. That is the subject of the next article.

For a closer look at one specific crossing, the comparison of MLCC and film capacitors covers where film genuinely substitutes for ceramic and where it does not, with the DC bias behaviour that drives most of the difference.

Common Questions About Capacitor Substitution

Can I replace an aluminum electrolytic capacitor with a film capacitor?

In DC-link, snubber and high-ripple positions, yes, and the result is often a longer-lived circuit. The constraint is volume and cost. Film capacitors store less energy per unit of volume than aluminum electrolytics, so matching a large bulk capacitance can take several times the space. If the enclosure is fixed, the swap usually fails on physical grounds before it fails electrically.

Is a higher voltage rating always a safe substitute?

Higher voltage rating is safe from a dielectric stress standpoint, but it is not free. A higher-rated part of the same capacitance is physically larger, which affects layout and can change ESL. In ceramic parts, moving to a higher voltage rating within the same case size usually means fewer active layers and lower effective capacitance. Always compare effective capacitance at the operating voltage rather than nominal capacitance.

Does substituting a capacitor require re-certification?

It depends on where the part sits. Components in a certified safety function, including X and Y safety capacitors and parts listed in an automotive or medical approval, generally require the certification to be revisited. Parts in ordinary filtering or bulk positions typically do not, provided the substitution stays within the approved specification and the change is documented. When in doubt, treat anything with a safety agency listing as a re-certification item.

How long does it take to qualify a second source?

For a same-type part with a comparable datasheet, a focused qualification can take a few weeks: electrical verification, a temperature run and a pilot lot. Cross-type substitutions run longer because the design margins have to be re-established, and a design change takes longer still. The qualification time is usually shorter than the lead time it is meant to solve, which is the argument for starting before the shortage reaches your part number.

Xuansn manufactures aluminum electrolytic, film, polymer and supercapacitors, and has supplied through previous shortage cycles. Because we build across several capacitor families, we can tell you whether a proposed substitution is feasible on our own lines rather than sending you to a third party for the answer. Our complete guide to capacitor types covers where each family fits, and our aluminum electrolytic selection guide covers the parameters that matter most when comparing parts across suppliers.

Evaluating a substitute for a part you cannot source? Send us the original specification and the position it sits in. Our engineers will confirm feasibility and reply with a quote within 24 hours.

✉ coco@xuanxcapacitors.com | ☎ +86-769-8166 8821 | Send your specification

The post Capacitor Substitution in 2026: What You Can Swap and What You Cannot appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.

We covered the price side of this cycle in our article on the 2026 capacitor price increase. This one works through the other half, in the order a buyer actually has to deal with it.

Where Lead Time Pressure Actually Sits

Two parts can sit in the same spreadsheet at the same price and carry completely different risk. Separating price from availability is what makes the ranking work, because the lines that deserve a commitment are rarely the ones that cost the most.

MLCC carries the longest waits. It sits closest to the AI server demand shock, and commodity grades are the first to be cut when capacity is redirected toward higher-margin parts. Aluminum electrolytic capacitors come next, with long-life and high-voltage series hardest to find while general-purpose values stay comparatively easy. Film is firmer on price but shorter on lead time. Supercapacitors behave differently again: the delay usually sits in qualification rather than in production, because fewer suppliers can meet the cell-level requirements for a UPS or hold-up module.

Nobody’s lead time quote survives a demand shift

A supplier quoting sixteen weeks in a tightening market is describing a plan. When demand moves again, that plan moves with it, and the purchase order you are holding does not travel with it. So track what the last three shipments on that part actually took, rather than what the system said they would take.

Which Parts Deserve a Commitment

Most buyers treat “should we lock in?” as one decision for the whole bill of materials. It splits into a per-part decision once you ask three questions.

How much of the part do you consume in a year, and how hard would it be to stop consuming it? How far out is the supplier quoting, and is the part already on allocation? If it goes short, can the design take a qualified substitute without a full requalification cycle?

If a part scores badly on all three, negotiation will not rescue it. The fix belongs in the design, and it is worth telling engineering that early.

The expensive parts are often the wrong ones to lock

It is tempting to start with the highest-value lines. That instinct usually produces the wrong list. A cheap capacitor appearing a thousand times on one board stops a line just as completely as an expensive one, and it is more likely to be allocated, because commodity grades are exactly what gets squeezed when capacity moves. Sort by exposure rather than unit cost.

What Allocation Actually Means for Your Order

Allocation works off purchase history. When a supplier cannot meet demand, they divide what they have in proportion to what you have bought in the past rather than this quarter’s requirement. A buyer with three years of steady orders keeps a larger share than one who placed the same total volume in a single urgent order.

That changes what good procurement looks like in a shortage. Predictable ordering is worth more than it appears, and it is one of the few levers a buyer still controls once a part goes short.

Why doubling your order works against you

When word of a shortage spreads, the instinct is to order more than you need and cancel later. Everyone does it at once. The shortage then looks worse than it is, the supplier allocates harder, and your purchase history stops reflecting your real usage. If the market softens, you are also sitting on volume you cannot consume. Ordering to forecast, and being visibly disciplined about it, keeps your allocation position defensible.

The Clauses That Make a Lock-In Worth Signing

A long-term agreement is only as useful as its terms. Committing to a volume at a price, with nothing else attached, moves all the risk onto the buyer.

A cap beats a fixed price

Ask a supplier for a fixed twelve-month price in a rising market and they will protect themselves somewhere else, usually in the base price or in delivery priority. A capped price with a stated trigger for review tends to produce a workable number, and it keeps the relationship intact when the market moves again.

When to Change the Design Instead

For the parts that score worst, a contract may be the wrong tool. Revisiting the specification can be quicker than negotiating for something nobody can supply.

Start with over-specification. A part carrying an automotive-grade rating on an industrial board, or a 125 °C rating where 105 °C will do, is paying for margin the design never uses, and it competes for capacity against buyers who genuinely need that grade. Then look at the type itself. Some designs use ceramic where a polymer or aluminum electrolytic part would work, which puts the BOM on the tightest part of the market for no functional gain. Matching the part to the rail rather than to habit is the exercise, and our selection guide for AI server power stages walks through it.

How long a second source really takes

Longer than most buyers expect. There are samples to run, bench verification at the actual operating conditions, a lifetime check at temperature, and for regulated or automotive programs a full qualification file. A process like that runs across quarters. It also cannot be compressed once a shortage is already underway, which is the argument for starting on the Tier 1 parts while there is still time to do it properly.

If your BOM leans on aluminum electrolytic capacitors for bulk and hold-up, the selection notes for 48 V and 800 V data center rails cover where a second source is realistic and where the electrical requirements leave no room.

A 90-Day Plan to Take Into the Next Negotiation

None of this has to happen in a single push, but the order matters more than the speed.

In the first month, build the ranked list. Sort the BOM into the three tiers, check current lead times against recent shipments rather than quoted figures, and flag every part with no workable substitute. In the second month, open the conversations. Share a rolling forecast that reaches beyond one quarter, ask for a capacity reservation on Tier 1 parts, and put a second source into qualification for anything that scored badly and cannot be re-specified. In the third month, close the agreements with the clause list above in hand, then set a review cadence so the ranking gets revisited as the market moves.

FAQ: Lead Times, Allocation, and Locking In Supply

How long are capacitor lead times in 2026?

It depends on family and grade. AI-grade MLCC is the tightest, with waits that have stretched toward ten months on server demand. Aluminum electrolytic runs longer on long-life and high-voltage series than on general-purpose values, and film is comparatively shorter. Confirm against recent shipments rather than the number in the system, because quotes move faster than they update.

Is the spot market worth the premium?

Sometimes, for a genuine stopgap on a Tier 1 part, but treat it as a last resort. The broker market is where counterfeit and re-marked parts circulate, and the saving on a short delivery can disappear in a failed batch or a recall. If you buy outside the authorized channel, budget for incoming inspection and keep broker stock out of regulated programs unless it comes with full traceability.

Should I lock the whole BOM or just part of it?

Part of it. Committing the entire bill of materials for a year turns a supply problem into an inventory problem, and most designs change enough over twelve months that some of that stock becomes unusable. Commit where lead-time risk and exposure are both real, and stay flexible on the rest.

Can a smaller supplier hold buffer stock for me?

Often yes, and it is worth asking. A supplier with a direct manufacturing line can sometimes hold agreed buffer stock or run scheduled releases against a forecast, which uses your working capital better than carrying the stock yourself. Put the arrangement in writing, including how the buffer is topped up and what happens if demand falls away.

Xuansn has supplied capacitors through previous shortage cycles, so we quote against the specification and volume you actually need rather than a catalog headline.

Locking in capacitor supply ahead of allocation? Xuansn Capacitor manufactures aluminum electrolytic, film and supercapacitor parts – our engineers will reply within 24 hours with feasibility and a quote.

✉ coco@xuanxcapacitors.com | ☎ +86-769-8166 8821 | Send your specification

The post Capacitor Lead Times Are Stretching: How to Lock In Supply Before Allocation appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.

In July 2026 Yageo raised prices across its capacitor lines by roughly 50 percent — the broadest single increase the passive components market had seen in years. It was not an isolated move. MLCC lead times have stretched toward ten months on AI server demand, aluminium electrolytic prices are climbing on data-center consumption, and market trackers now put the average capacitor price increase for 2026 at around 30 percent. If you buy capacitors in volume, the question is no longer whether prices will rise, but how to keep your bill of materials and your margins intact.

This is the first article in our series on the AI-driven capacitor price cycle. The next two cover how to select and lock in orders during a rising market, and how to weigh substitute and domestic alternatives.

What Changed in the Capacitor Market in 2026

For most of the past decade, capacitors were a buyer’s market. Distributors held stock, prices drifted down, and a design engineer could swap a part without a second thought. That has reversed in a single cycle.

The pattern is the one that hit semiconductors in 2021: a demand shock from a single end market, a supply base that cannot add capacity quickly, and a spot market that amplifies the gap. Passive components take longer to expand than most buyers assume. Ceramic powder, aluminium foil and film lines are capital-intensive, and a new line has to be qualified before it ships a production part.

The Three Forces Driving the Capacitor Price Increase

AI servers rewrote the demand curve

A single AI server board carries far more capacitors than a conventional server board. High-power GPU rails need dense MLCC arrays to hold voltage ripple at hundreds of amps, the 48 V intermediate bus needs bulk electrolytics and polymer parts, and the hold-up circuit adds more on top. Multiply that by the build-out of AI data centers and you get a demand step the passive components industry did not size for. AI server power demand is the single biggest driver behind the current increase.

Capacity moved to high-value parts

When demand outruns supply, manufacturers do the rational thing: they prioritise the parts with the best margin. Automotive-grade and high-reliability MLCCs, long-life aluminium electrolytics, and specialty film parts get the capacity, while commodity grades are squeezed. Buyers of standard parts then face the worst of both worlds — higher prices and longer waits, with no design change to show for it.

Lead times, allocation, and the spot market

As lead times extend, allocation appears. Once a supplier starts allocating, buyers who did not forecast early get partial shipments, and the shortfall is filled on the spot market at a premium. That premium then feeds back into contract negotiations. MLCC lead times approaching ten months mean a part ordered today may not arrive until next year, which pushes buyers to order earlier than they need — and early ordering itself tightens supply further.

Which Capacitor Types Are Rising Fastest

Not every type moves together. MLCC is where the shortage is sharpest. Aluminium electrolytic is being pulled up by the same AI data-center demand. Film is firmer but less acute. Supercapacitors are still a smaller market with their own supply constraints. If your BOM is heavy in MLCC, you are exposed to the fastest-moving part of the increase, and that is where procurement attention pays off first.

How to Protect Your BOM in a Rising Market

A price increase is a procurement problem before it becomes an engineering problem. The buyers who come through a rising market well are the ones who act before allocation starts.

Forecast further out and commit earlier. Extend your demand forecast from a quarter to two or three quarters and share it with your supplier. Suppliers allocate to the customers who give them visibility. A forecast is not a purchase order, but in a shortage it is the currency of allocation.

Qualify a second source before you need it. A second source is worth little if you start qualifying it during a shortage. Begin now, on the parts with the longest lead times, and keep the qualification live.

Re-examine your specifications. Over-specified parts are the easiest place to find cost. If a design carries an automotive-grade MLCC where a standard grade would do, or a 125°C part where 105°C is enough, that specification is paying the shortage premium for no benefit. Our guide to selecting capacitors for AI server power stages walks through matching the part to the rail rather than to habit.

Use the right type for the rail. Part of the pressure on MLCC comes from designs that use ceramics where an electrolytic or a polymer part would work. Revisiting the choice between multilayer ceramic capacitors and aluminum electrolytic capacitors can move a BOM off the tightest part of the market.

Build a small buffer deliberately. A modest safety stock on your longest-lead parts is cheaper than a line stop. Size it from real lead-time risk rather than hoarding, and review it monthly while allocation is in force.

What This Means for Long-Term Sourcing

The 2026 increase is unlikely to be a one-quarter event. AI data-center build-out is a multi-year trend, and the passive components capacity needed to serve it takes years to come online. Even as new capacity arrives, prices rarely return to where they started; they settle at a new level.

That changes what a good supplier looks like. Price still matters, but so does allocation history, willingness to hold buffer stock on your behalf, and the ability to offer a genuine substitute when a part is short. A supplier who can move you from a constrained MLCC to a qualified alternative, or from a premium part to a correctly specified one, is worth more in a rising market than a supplier with a lower list price and no stock.

Xuansn has supplied capacitors through previous shortage cycles, and we quote against the actual specification and volume you need rather than a catalogue headline. Send us your BOM and target price and we will confirm availability, lead time, and a substitute where one exists.

FAQ: Common Questions About the 2026 Capacitor Price Increase

Will capacitor prices fall again in 2026 or 2027?

Some grades may ease as new capacity comes online, but a return to 2024 pricing is unlikely while AI data-center demand keeps growing. Expect prices to plateau at a higher level rather than fall back to where they were.

Why are MLCC prices rising faster than aluminum electrolytic?

MLCC sits closest to the AI server demand shock and has the longest lead times, so it moved first and hardest. Aluminium electrolytics are being pulled up by the same data-center demand and by aluminium foil supply, but the shortage is less acute.

Should I lock in a full year of supply now?

Locking volume on your longest-lead and highest-exposure parts usually pays for itself. Locking everything for a year can leave you holding stock you cannot use if demand softens. Commit where the lead-time risk is real, and stay flexible elsewhere.

Can I substitute one capacitor type for another to control cost?

Often yes, but only after checking the electrical and thermal requirements. Moving from MLCC to a polymer or aluminium electrolytic part is sometimes viable on bulk rails, while a timing or RF circuit may not tolerate the change. Qualify any substitute on the bench before it reaches production.

Sourcing capacitors through a shortage cycle? Xuansn Capacitor manufactures aluminum electrolytic, film and supercapacitor parts – our engineers will reply within 24 hours with feasibility and a quote.

✉ coco@xuanxcapacitors.com | ☎ +86-769-8166 8821 | Send your specification

The post Capacitor Price Increase 2026: Why Prices Are Rising and What Buyers Should Do appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.

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.

The post How to Select Capacitors for AI Server Power Stages appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.

This is the third article in our AI server series. In the first article we looked at how AI server power demand is exploding, and in the second at how capacitors work in a switching power supply. Here we map the capacitors onto each power stage of a server, so you can see which type does what – and which one your design actually needs.

Riding the Rail Down From 380 V to 1 V

Think of the server power delivery path as a train running down a voltage line. The alternating input is rectified to a high-voltage DC bus near 380 V. That bus feeds a converter that steps down to a 48 V intermediate rail – the distribution voltage adopted by most modern AI racks. A second stage of point-of-load converters takes that 48 V down to the sub-1 V cores and the 1.8 V memory rails that GPUs and CPUs actually run on.

Each of these steps needs a different capacitor, because the job is different at each voltage and current level.

The pattern is simple: the further down the rail you go, the lower the voltage and the faster the transients – and the more the capacitor choice shifts from big capacitors doing bulk storage to many small ones doing decoupling.

Four Capacitors, Four Different Jobs

Aluminum electrolytic – the hold-up workhorse

Right after the rectifier, an aluminum electrolytic capacitor smooths the high-voltage bus and stores the energy that keeps the server alive during a few milliseconds of AC loss. Its job is defined by hold-up time – the window after input power drops during which the output still stays in regulation. Most server and ATX power specs ask for 16 to 20 ms of hold-up, with some enterprise requirements settling at a 10 ms minimum. That is why bulk electrolytic capacitors with high ripple-current ratings and long 105 °C lifetime dominate the primary side.

Polymer and solid capacitors – response where the load steps

On the 48 V intermediate bus, the load changes fast as GPUs boost. A polymer or solid aluminum capacitor offers much lower ESR than a wet electrolytic, so it can answer a load step without letting the rail sag. That lower ESR also means less self-heating for the same ripple current, which is exactly what a power-hungry AI rail demands.

MLCC – decoupling at the edge

At the point-of-load, tens or even hundreds of small MLCCs sit right next to the GPU core. At sub-1 V, the capacitance needed is small but the switching frequency is high, so what matters is low ESL, high-frequency response and placement as close to the die as possible. This is the decoupling role, distinct from the bulk role upstream. If you want the full argument for why this matters, our pillar on all capacitor types walks through the trade-offs in depth.

Supercapacitors – the ride-through option

When 16-20 ms of hold-up is not enough – say, a longer ride-through window or a battery-less BBU design – a supercapacitor bank can take over. A bank built from 2.7 V-class cells handles seconds of backup where an electrolytic would need an impractically large can. It is a different trade: lower voltage per cell, balancing circuits required, but a far longer ride-through for the same footprint trade-off.

Choosing by Rail: A Practical Comparison

The pattern here is equally practical. Near the inlet, you want big capacitance and rated ripple – the aluminum electrolytic territory we build at aluminum electrolytic capacitors. On the 48 V bus, solid capacitors answer fast load steps. At the core, ceramic capacitors (MLCC) do the high-frequency decoupling. And when your backup window stretches past 20 ms, a supercapacitor bank is the option to compare.

Where Xuansn Fits

You rarely need a single capacitor – you need the right one at every rail. That is why Xuansn builds the full set: aluminum electrolytic, polymer and solid, MLCC, and supercapacitors. Whether you are upgrading an existing AI rack or sourcing a new PSU, the same question applies: which rail is your weak point?

Deeper dives live on our specialist sites: the aluminum electrolytic site covers 48V/800V architectures, and the supercapacitor site walks through BBU and UPS designs. Start here, then follow the rail to the detail.

What’s Next in the AI Server Series

This article mapped the capacitors onto the rails, building on the principles of how capacitors work in a switching power supply. The next piece turns the map into a concrete selection and sourcing plan – how to compare offerings, what to ask a manufacturer, and where to start when you are sourcing AI power capacitors.

Common Questions

Do AI servers really use supercapacitors?

They can, but not in every PSU. In a standard AI server PSU, the bulk and hold-up job belongs to aluminum electrolytics. Supercapacitors appear where you need sec-scale ride-through or a battery-less BBU – a separate power-path decision, not a drop-in swap on the same rail.

What size capacitor do I need for an AI server?

It depends entirely on the rail. For the input bus you size for hold-up energy at the bus voltage window; for the core you size for the load step and switching frequency. Give us the rail, load and hold-up target and we will help you pick – rather than guess from a single number.

Can one capacitor type handle the whole AI power chain?

No – and that is the point of this article. A high-voltage bus calls for bulk, a 48 V rail for low-ESR response, and a sub-1 V core for high-frequency decoupling. Each stage rewards a different capacitor, which is why a mixed set is the norm in serious AI power stages.

The post Which Capacitors Fit AI Server Power Stages appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.

AI-optimized servers are forecast to consume 175 TWh in 2026, up 84% in a year, as we covered in our look at AI server power demand. This article is about the component that quietly holds the power stage together: the capacitor.

Most engineers can name what a capacitor does. Fewer can say why it becomes the first part to fail in a high-density power supply. Let’s fix that.

Three Jobs Every Capacitor Does in a PSU

In a switching power supply, the capacitor is doing three things at once:

• Smoothing the rail. After the rectifier, the capacitor holds the voltage up between switching pulses. The larger the capacitance, the less the rail sags.
• Decoupling (bypassing). Every time a chip switches, it demands a sudden burst of current. The capacitor near the load supplies that burst locally, keeping the rail from dipping.
• Energy storage across the switching gap. In the converter, energy is chopped and transferred in pulses. Capacitors store the energy between those pulses so the output stays steady.

A capacitor is not a passive part in a modern PSU – it is doing active work every microsecond.

Why AI Power Stages Push Capacitors to the Edge

The numbers translate directly into harder conditions for every capacitor in the chain:

• Higher ripple current – ripple current heats the capacitor from the inside. More current means more heat, and heat is what kills lifetime.
• Higher switching frequency – as converters shrink, frequency rises, and the capacitor needs lower ESR to stay cool at that frequency.
• Tighter rails – AI accelerators tolerate less voltage deviation, so the capacitor must hold the rail under bigger, faster load steps.
• Temperature – capacitor lifetime follows the Arrhenius rule: roughly halving for every 10 °C rise in core temperature. In a dense rack, ambient heat and self-heating add up fast.

This is why a capacitor that was fine in the previous generation can become the failure point in the next. That same pressure reaches the capacitors across the AI power chain – supercapacitors for backup bursts, film capacitors in the DC-link, and electrolytic capacitors on the bulk bus.

What to Look For When You Are Sourcing

If you are specifying capacitors for a power design, the parameters that matter most are:

• Ripple current rating – can the part handle the current it will actually see?
• ESR at the switching frequency – lower ESR means less self-heating and less ripple voltage.
• Lifetime at your operating temperature – datasheet lifetime (e.g. 2,000 h) is quoted at a rated temperature; derate it to your real ambient.
• Manufacturer data – the part is only as good as the test data behind it. Xuansn publishes ripple and lifetime data for every capacitor series. Our capacitor types guide walks through choosing the right type.

A capacitor that meets the right capacitance but ignores these will be the component that fails in the field.

Common Questions About Capacitors in Switching Power Supplies

Why does a capacitor fail first in a power supply?

Capacitors carry ripple current and run hot; heat accelerates electrolyte loss and wear. In a dense supply they often hit their thermal limit before other components, so lifetime is the parameter to check first.

What is ESR and why does it matter?

ESR (equivalent series resistance) converts ripple current into heat. Lower ESR at the switching frequency means less self-heating, less ripple voltage, and a longer-lived part.

How do I choose between electrolytic, film and ceramic?

Use aluminum electrolytic for bulk energy and hold-up, film for DC-link and ripple-heavy positions, and ceramic for high-frequency decoupling. Feedthrough capacitors handle EMI at the power entry. The choice follows the job each stage needs.

What to Read Next

We will cover which capacitor types fit AI server power stages – electrolytic, film, ceramic and super. If you are sourcing power components right now, you can send us your requirement.

Need capacitors for high-power switching supplies? Xuansn Capacitor manufactures aluminum electrolytic, film and supercapacitor parts – our engineers will reply within 24 hours with feasibility and a quote.

✉ coco@xuanxcapacitors.com | ☎ +86-769-8166 8821 | Send your specification

Sources: Gartner data center electricity forecast, June 2026.

The post How Capacitors Work in a Switching Power Supply appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.

Global data centres are on track to consume 565 TWh of electricity in 2026 – a 26% jump in a single year, according to Gartner’s June 2026 forecast. The driver is not general internet traffic. It is AI-optimized servers, whose electricity use is growing faster than the power industry has ever seen.

If you design or buy power components, this is not a distant trend. It is already changing what gets specified today.

How Much Power Do AI Servers Really Consume?

• • AI-optimized servers consumed about 95 TWh in 2025 and are forecast to reach 175 TWh in 2026 – roughly 84% growth in one year.
  • They will account for about 31% of total data-centre power consumption in 2026, up from ~20% in 2025.
  • By 2027, AI-optimized server consumption is projected to hit 258 TWh, overtaking conventional servers for the first time.
  • Gartner warns that grid supply will be insufficient once consumption passes 1,200 TWh by 2030 – power availability is now a binding constraint on AI expansion.

What This Means for Power-Supply Design

More power means higher current, tighter efficiency targets and denser racks. The power stages that feed AI accelerators – from the PSU down to the VRM on the motherboard – are pushing every component to its limit:

• Higher ripple current heats the capacitor – lifetime shrinks fast.
• Higher switching frequency calls for lower ESR.
• Tighter rails allow less voltage deviation under load steps.

The capacitor, often the quietest part in a power supply, is increasingly the bottleneck.

What Should You Specify for AI-Era Power Supplies?

For procurement and design teams, the practical consequences are clear:

• Demand for high-reliability, high-ripple-rated parts is growing with the market.
• The spec sheet is getting stricter – ripple current, ESR and lifetime at temperature now matter more than raw capacitance.
• Component supply security matters as much as spec – AI programs move fast, and a manufacturer that can supply at volume and customize where needed is a real advantage. Xuansn Capacitor is one such producer, manufacturing aluminum electrolytic and supercapacitor parts at scale.

Common Questions About AI Server Power and Capacitors

What capacitors are used in AI server power supplies?

AI server power stages typically combine aluminum electrolytic capacitors for bulk energy storage and hold-up at 48 V and 800 V, polymer capacitors for low ESR on tight rails, film capacitors for DC-link duty, and MLCCs for high-frequency decoupling. Supercapacitors also appear in backup and peak-shaving roles at the rack level, and feedthrough capacitors keep EMI off the power entry.

Why does ripple current matter for AI power supplies?

Ripple current heats a capacitor from the inside. AI accelerators draw more current than previous generations, so ripple current rises and the capacitor must be rated to handle it without overheating, or its lifetime collapses.

How do I choose capacitors for AI server power stages?

Check ripple current rating at your switching frequency, ESR at that frequency, and lifetime derated to your real ambient temperature. Our capacitor types guide walks through the comparison. Raw capacitance alone will not tell you whether the part survives in a dense AI rack.

What to read next:

1. How capacitors behave in a switching power supply (next up).
2. Which capacitor types fit AI server power stages – electrolytic, film, ceramic, super.
3. How to choose and source them – the parameters that matter.

If you are specifying power components right now, the next article is worth your time.

Sources: Gartner Press Release, June 2026 – Data Center Electricity Consumption to Grow 26% in 2026. Gartner – AI servers to consume more power than all conventional data center hardware combined by 2027.

Sourcing capacitors for AI-era power designs? Our engineers will reply within 24 hours with feasibility and a quote.

✉ coco@xuanxcapacitors.com | ☎ +86-769-8166 8821 | Send your specification

The post AI Server Power Demand Is Exploding: What It Means for Power Supplies and Capacitors appeared first on Dongguan Xuanxuan Electrolytic Technology Co,.ltd.