Server Power Supply Market (2026-2036)

This report covers the conversion equipment that takes facility power and delivers the direct current a server consumes, across the per-server power supply unit, the rack-level power shelf, the external sidecar rectifier and the direct current input converter inside the server. The growth of this market is driven by rising power per rack rather than by server unit shipments, since the kilowatts a rack must be fed have risen far faster than the number of boxes in it.

The structural development in this market is that the server power supply is being removed from the server. It moved first to a rack-level power shelf, and it is now moving again to an external sidecar, with the unit remaining inside the box reduced to a direct current converter. Each step relocates the alternating to direct current conversion further upstream, and each step changes who buys the equipment, through which channel and against which specification. A supplier whose relationships sit with server manufacturers is losing its customer rather than its market. That distinction determines where development and sales investment should go, and it is the question this report is organised around rather than the growth rate itself. It also explains why volume growth and supplier prosperity have diverged in this category. The second is the subject of the cautions set out later in this report.

A server power supply converts facility alternating current into the direct current rails a server consumes. For most of the history of the product this was a component inside the chassis, specified by the server manufacturer, procured in a bill of materials and duplicated for redundancy, and the market behaved accordingly as a high-volume component business competing on cost, efficiency rating and form factor compliance. That description is becoming inaccurate, and the reason is architectural rather than commercial.

The first relocation moved conversion to the rack. The Open Rack Standard V2.0, whose licence section is dated as of 15 July 2016, specifies rack-level 48 V and 12 V power shelves housing rectifiers or batteries with a minimum of N plus one redundancy, front-mounted and placeable anywhere in the rack. The 48 V system operates between plus 40 V and plus 59.5 V direct current at a nominal plus 54.5 V, while the 12 V system is held at 12.2 V plus or minus 0.4 V averaged at any point along the bars, and bus bars sit in the power and cooling zone on the hot aisle side of the rack, running vertically with one or three bars per power zone. The 48 V rectifiers specified take a single-phase alternating current input of 200 to 240 V or 200 to 277 V, provide an output programmable from 42 V to 58 V defaulting to 54.5 V, achieve peak efficiency above 97 percent at 230 V alternating current, and support redundant parallel operation with load sharing and hot insertion and removal. Payload trays use a hot-swap controller for inrush current control and fault isolation so they can be inserted into or removed from a live bus, capacitors cannot be placed before the hot-swap device, the short pin disconnects first with a required 5 millisecond delay before the other pins during removal, and remote restart capability is mandatory. Power reaches equipment through a direct current connector that blind-mates into the bus bars at the rear, with the 48 V connector carrying return, power and short pin contacts in a make sequence of return, then power, then short pin, and each tray fused on the positive 48 V input ahead of its hot-swap circuit and voltage regulators. The standard recommends direct 48 V to point-of-load conversion for processor and memory rails.

Read as a commercial document rather than an engineering one, that specification does three things. It removes the alternating to direct current conversion from the server and places it in a shelf bought with the rack. It standardises the mechanical and sequencing behaviour of the unit in exceptional detail, down to the make order of the connector contacts and a 5 millisecond pin disconnect delay, which leaves little scope for a supplier to differentiate on the interface. And it sets a peak efficiency expectation above 97 percent at 230 V alternating current, which is higher than the regulatory floor discussed below. The consequence is a product defined by the customer, assembled to a published behaviour, and judged on cost and reliability.

The second relocation is now in progress and moves conversion out of the rack altogether. A white paper dated 25 September 2025, authored by Pietro Scalia, Yong Perry Li and Ashish Ekbote of Renesas Electronics and written in compliance with Open Compute Project tenets, records that artificial intelligence driven rack power growth is outpacing conventional 48 V in-rack distribution, that 800 V is expected to be selected in either two-wire or three-wire form, and that the direct current bus is generated in an external sidecar rack from a three-phase 480 Vrms alternating current source using a Vienna-topology front end with bidirectional gallium nitride switches. A high-power-density isolated 16 to 1 direct current converter steps 800 V down to 48 V so that the existing 48 V ecosystem can be reused, sidecar building blocks are trending towards approximately 20 kW or higher, and a validated high-voltage direct current design of 12 kW is built from two 6 kW units. The unit remaining inside the server is a direct current to direct current converter, and the recommendation in the rack standard for direct conversion from the bus to the point of load for processor and memory rails points in the same direction, towards fewer conversion stages inside the box and more outside it.

The regulatory position is where this market contains something genuinely counter-intuitive. Commission Regulation (EU) 2019/424 of 15 March 2019 lays down ecodesign requirements for servers and online data storage products, and its power supply unit provisions in Annex II apply to those products while expressly excluding direct current products. From 1 March 2020 a multi-output unit was required to achieve 88 percent efficiency at 20 percent load, 92 percent at 50 percent load and 88 percent at 100 percent load with a minimum power factor of 0.90 at 50 percent load, and a single-output unit 90 percent, 94 percent and 91 percent at the same load points with a minimum power factor of 0.95. From 1 January 2023 those values were replaced, requiring a multi-output unit to achieve 90 percent at 20 percent load, 94 percent at 50 percent load and 91 percent at 100 percent load with a minimum power factor of 0.95, and a single-output unit to achieve 90 percent at 10 percent load, 94 percent at 20 percent load, 96 percent at 50 percent load and 91 percent at 100 percent load with the same power factor. The requirements differ by unit type rather than by rated wattage, with a single-output unit defined as one delivering most of its rated power to a single primary direct current output, and the only rated-output distinction at 750 watts appears in the indicative best-available-technology benchmarks under Annex V rather than in the binding requirements.

The detail that matters is in the first sentence. The regulation applies to servers and online data storage products while expressly excluding direct current products. As conversion migrates to a power shelf and then to a sidecar, the equipment inside the server becomes a direct current input device, and the efficiency and power factor floors that bind an alternating current server power supply do not follow the function upstream. The architecture migration is therefore also a regulatory exit. That cuts in both directions, since it removes a compliance obligation and simultaneously removes the floor that guaranteed efficiency, leaving the question to voluntary specification. In practice the voluntary specification is more demanding than the regulation, with the rack standard expecting above 97 percent peak against a regulatory requirement of 96 percent at 50 percent load for a single-output unit, but that is a matter of customer insistence rather than law.

There is a second regulatory nudge in the same direction that is easily missed. The same regulation sets idle state power limits from 1 March 2020 for servers with one or two processor sockets, excluding resilient, high performance computing and integrated accelerated processing unit servers, calculated as a base allowance of 25 watts for a single-socket server, 38 watts for a two-socket server and 40 watts for a blade or multi-node server, plus component additions that include 10 watts for each additional power supply unit, 5.0 watts for each hard disk or solid state drive and 0.18 watts per gigabyte of memory above 4 gigabytes. Active state efficiency minimums of 9.0, 9.5 and 8.0 apply to the same three categories, manufacturers must declare an operating condition class from A1 to A4 without a minimum class being set, and from 1 March 2020 joining, fastening or sealing techniques must not prevent disassembly for repair or reuse of listed components including the power supply unit. The idle allowance explicitly adds 10 watts for each additional power supply unit, which prices per-server redundancy into the compliance calculation and makes a one plus one arrangement inside the chassis more costly against the limit than redundancy provided once at shelf level for many servers. Regulation written for the per-server architecture therefore contains an incentive to leave it.

The commercial consequence of all this is a change of customer rather than a change of volume. A per-server unit is sold to a server manufacturer as a line in a bill of materials, against a form factor standard, on cost and efficiency class. A power shelf is sold to a rack integrator or directly to an operator, as part of a rack specification the operator has frequently written itself. A sidecar rectifier is sold to the electrical infrastructure buyer, alongside switchgear and distribution, on terms that resemble industrial power equipment rather than information technology components. The same function passes through three different channels with three different competitive sets, and incumbency in one confers little advantage in the next.

Efficiency has also stopped being a useful differentiator, which is unusual for a power product. The binding requirement for a single-output unit reached 96 percent at 50 percent load from 1 January 2023, and the open rack specification already expected above 97 percent peak in a document whose licence section dates from 2016. When the regulatory floor and the voluntary expectation are both in the mid to high nineties, the remaining spread between suppliers is narrow enough that a buyer cannot meaningfully choose on it, and competition moves to power density, thermal behaviour, controls and telemetry, and manufacturing cost. The detailed hot-swap and sequencing requirements in the rack standard close off the mechanical interface as a differentiator as well.

A caution belongs in any assessment of this market. The migration described is well evidenced in published specifications and in architecture work, and the per-server unit nonetheless remains the largest segment and will remain substantial throughout the forecast period, because the installed base is enormous, because general purpose enterprise and telecom servers are not being rearchitected, and because the sidecar approach is described as expected rather than settled. Forecasts treating the newer architectures as displacement rather than as addition at the high-power end will understate the per-server segment considerably.