The 27 kW Power Supply That Shows Where AI Server Electricity Is Headed

When the Rack Becomes the Bottleneck

For years, the conversation around AI infrastructure has centered on GPUs: how many, how fast, how much memory bandwidth. Increasingly, though, the harder problem is getting enough clean, stable power into the rack in the first place. A newly introduced 27 kW three-phase power supply unit designed for next-generation AI server power architectures is a useful marker of just how far that problem has escalated — and it tells a quieter story about the capacitors that have to absorb, filter, and smooth that power along the way.

What’s Actually Driving This Shift

The headline figure is 27 kW from a single PSU, built to support a move toward 800 VDC or ±400 VDC power distribution architectures inside AI server racks. That shift away from traditional lower-voltage DC distribution is a direct response to two compounding pressures: GPU power consumption has climbed sharply generation over generation, and rack densities keep increasing as operators try to pack more compute into the same data center footprint. Distributing power at a higher DC voltage reduces resistive losses and the amount of copper needed to move the same wattage, which is why the industry’s center of gravity is visibly shifting toward higher-voltage DC busbars at the rack level.

But higher voltage, higher power density architectures do not succeed on semiconductor design alone. Every stage of power conversion — from the three-phase AC input, through rectification, through isolation, down to the point-of-load conversion feeding the GPU itself — depends on capacitors to handle bulk energy storage, ripple filtering, and transient response. As the power level climbs from the kilowatts typical of legacy server PSUs into this 27 kW class, the capacitor bank supporting that PSU has to handle proportionally larger ripple currents and transient energy swings without a proportional increase in physical size, because rack space is, if anything, under more pressure than ever.

The Capacitor’s Role in High-Power PSU Design

Inside a PSU of this class, several capacitor types typically work together. Bulk electrolytic or film capacitors on the DC link smooth the rectified AC input and absorb energy during load transients, acting as a buffer between the utility-facing AC side and the fast-switching DC-DC conversion stages downstream. High-frequency ceramic capacitors, including MLCCs, handle the higher-frequency ripple generated by switching converters, filtering noise that would otherwise propagate downstream toward sensitive GPU power rails. As switching frequencies rise — a common technique for shrinking magnetic component size — capacitor selection has to account for both higher ripple current handling and lower equivalent series resistance (ESR) and equivalent series inductance (ESL), since losses in these components scale directly with how much ripple current they are asked to absorb.

The move to 800 VDC or ±400 VDC distribution also raises the voltage rating requirements across the entire capacitor bank, which is not a trivial re-spec. Film and electrolytic capacitors rated for these voltage classes typically require different dielectric formulations and construction than their lower-voltage counterparts, and the thermal design margin matters more, not less, because a 27 kW PSU operating continuously inside a dense AI server rack has far less thermal headroom than a legacy design running at a fraction of that power.

Why This Matters Across the AI Server Ecosystem

This single PSU announcement is really a proxy for a much broader infrastructure shift. As AI training and inference clusters scale, data center operators are increasingly treating power delivery architecture as a first-order design constraint rather than a downstream afterthought bolted onto a rack design that was finalized around compute considerations alone. The capacitors and passive components supporting that power delivery chain — DC-link capacitors, MLCCs for high-frequency filtering, and increasingly supercapacitors or battery-backup-adjacent components for ride-through during power transients — sit directly on the critical path for whether these higher-density racks can be deployed reliably at scale.

Supply Chain and Design Implications

For power supply designers, the shift toward 800 VDC-class architectures means qualifying capacitor banks against a higher voltage and higher ripple-current specification than most legacy server PSU designs required, which in turn narrows the pool of qualified components and suppliers capable of meeting both the electrical and thermal requirements at this power class. For data center operators and system integrators, this translates into a procurement reality where capacitor and passive component availability at the right voltage and ripple-current rating becomes a gating factor for rack deployment schedules, not merely a line item on a bill of materials.

For passive component suppliers, rising AI server power density represents one of the more durable demand drivers in the industry right now, since each generational jump in GPU power consumption effectively forces a corresponding jump in capacitor voltage rating, ripple-current capability, and thermal robustness across the entire power delivery chain supporting it — a dynamic that is unlikely to reverse as long as AI compute demand keeps climbing.

The Bigger Picture

A 27 kW PSU is, on its surface, a power electronics story about semiconductors and topology. Underneath it, though, is a steady, compounding demand signal for the capacitors and passive components that make higher-voltage, higher-density AI server power architectures physically possible. As rack power budgets keep climbing, the humble capacitor bank inside the PSU is quietly becoming one of the more consequential design bottlenecks in AI infrastructure.

There is also a reliability dimension worth noting. A PSU running continuously at tens of kilowatts inside a densely packed rack rarely gets the luxury of a scheduled shutdown for maintenance, which means the capacitor bank inside it has to deliver consistent ripple filtering and transient response across years of near-continuous operation, not just at the moment the unit first passes qualification testing. Capacitor aging — the slow, predictable drift in capacitance and ESR that every electrolytic and film capacitor experiences over its service life — becomes a first-order reliability concern at this power class, because a capacitor bank that was comfortably oversized on day one can quietly lose margin over years of operation at elevated temperature. Data center operators planning multi-year deployment cycles increasingly factor that aging curve into their maintenance and replacement planning, rather than treating the PSU as a fixed, maintenance-free black box for its entire service life. That shift in mindset, as much as any single component spec, is part of what higher-density AI server power architectures are forcing onto the industry.

Related Listed Companies to Watch

Directly Related Companies

Company / 公司	Ticker / 股票代碼	Market / 市場	Relation / 關聯角色	Strength / 關聯強度
Vishay	VSH	NYSE	DC-link and film capacitor manufacturer	High
TDK	6762.T / TTDKY	TSE/OTC	MLCC and film capacitor manufacturer	Medium
Yageo (國巨)	2327	TW	MLCC and capacitor manufacturer (KEMET)	Medium

Extended Supply-Chain Watch

Company / 公司	Ticker / 股票代碼	Market / 市場	Relation / 關聯角色	Strength / 關聯強度
Delta Electronics (台達電)	2308	TW	Data center power supply, extended demand side	Medium
Lite-On (光寶科)	2301	TW	Power supply module, extended demand side	Medium
Vertiv	VRT	NYSE	Data center power infrastructure, extended demand side	Medium
Wiwynn (緯穎)	6669	TW	AI server systems, extended demand side	Medium

This section is for industry-chain reference only and does not constitute investment advice.