The Server PSU and Power Systems Space
Market Structure, Material System Competition, Supply Chain Constraints, and Architectural Integration for AI Rack Power Delivery, 2026 through 2035
Abstract
The server power supply unit and rack-level power conversion market sits at the front of a structural transition driven by artificial intelligence rack densification. Rack power densities that ran 10 to 30 kilowatts in conventional cloud infrastructure now run 60 to 240 kilowatts in mainstream Blackwell-class AI rack designs, and the next deployment generation moves to the 600 kilowatt to 1.4 megawatt envelope. The power conversion stack between the utility transformer and the silicon load is the loadbearing element of this transition, and the suppliers who win at each conversion stage are not the same as the suppliers who won in the prior generation.
The paper analyses the market at three levels. At the device level, it positions silicon, silicon carbide, and gallium nitride against each other by conversion stage, voltage class, switching frequency, and qualifiedsupply posture, and identifies where each material system is winning, where each is losing, and where the contest remains open. At the module level, it positions the power semiconductor suppliers Infineon, Monolithic Power Systems, Renesas Electronics, Texas Instruments, Navitas Semiconductor,
Innoscience, Power Integrations, Vicor, EPC, onsemi, Wolfspeed, and others against rack PSU vendors Delta Electronics, Lite-On, Chicony, Flex, Murata, Artesyn, and emerging entrants. At the architecture level, it positions 415 VAC versus 400 VDC versus 800 VDC distribution choices against the mediumvoltage rectifier and solid-state transformer integrations that pull device content forward into the facility envelope.
Findings. The serviceable addressable market for rack PSU and intermediate-bus power conversion expands from roughly 12.4 billion US dollars in 2025 to a 2030 mid-case of 47 billion US dollars driven by rack densification and conversion stage consolidation. GaN dominates the 48 volt intermediate bus and the 400 volt rack PSU stage at densities above approximately 4 kilowatts per cubic decimeter. SiC dominates the rectifier and front-end conversion stages at the 800 volt level and above. Hyperscaler qualification cadence has compressed for incumbent suppliers and expanded for new entrants. Geopolitical sourcing risk is now a material discriminator at the supplier list level, with China-domiciled GaN suppliers facing restricted access to US federal and hyperscaler-sensitive workloads.
Recommendations. Sourcing organizations should dual-source at the rack PSU SKU level by 2027. Capital plans should treat the 800 VDC transition as a power semiconductor decision rather than a transformer decision. Geopolitical sourcing exposure should be managed explicitly with a quarantine posture against China-fab origin power semiconductor content for US-sensitive deployments. The three recommendations are mutually reinforcing rather than independent and should be implemented as a coordinated sourcing-posture program rather than as discrete procurement actions.
Scope and posture. The scope of the analysis is global with explicit treatment of North American, European, Asian, and China-domestic dynamics. The temporal scope is 2026 through 2035 with primary modeling on 2026 through 2030 and scenario extension through 2035. The analytical posture is thirdperson analytical with practitioner-experience framing where directly relevant. The paper does not contain financial advice or specific procurement recommendations and should not be relied upon as such; it is offered as a structured industry reference for senior infrastructure decision makers.
Executive Summary
The server PSU and power systems space is undergoing the most consequential architectural and supplier realignment of the last twenty years, and the rate of change accelerates through the back half of 2026 as the first 800 volt direct current production racks enter hyperscaler service. The transition is not a single event but a layered restructuring across three independent vectors that move at different speeds: the device material system at the silicon, gallium nitride, and silicon carbide level; the modulelevel power conversion stack from front-end rectifier through point-of-load; and the facility-level distribution choice between legacy 415 volt alternating current, 400 volt direct current, and emerging 800 volt direct current. Each vector reshapes a different supplier population, and the suppliers who emerge dominant at the end of the decade are not the same set who dominated the prior generation.
Thesis. The server PSU and power semiconductor space is in the early innings of a multi-year restructuring driven by AI rack densification. The outcome is determined not by component performance alone but by manufacturing readiness, dual-sourcing discipline, geopolitical sourcing risk, and the speed with which 800 volt HVDC and direct-to-rack distribution displace the legacy 415 VAC and 48 volt intermediate bus topology stack. Incumbent suppliers Infineon, Texas Instruments, Monolithic Power Systems, Renesas, and Vicor each face this transition from different competitive positions, and the resulting share map at the end of the decade is materially different from the map at the start.
Three findings frame the analysis. First, the rack PSU and intermediate bus power conversion total addressable market expands from approximately 12.4 billion US dollars in 2025 to a 2030 range of 34 to 48 billion US dollars driven by AI rack power densification from 30 to 100 kilowatts per rack to 500 to 1,400 kilowatts per rack, and by the consolidation of intermediate conversion stages onto fewer but more capable converter modules. Second, wide-bandgap semiconductors do not win uniformly across the conversion stack. GaN dominates the 400 VAC to 48 VDC intermediate bus stage and the 48 V to 12 V or 48 V to point-of-load stage at densities above approximately 4 kilowatts per cubic decimeter where switching frequencies above 1 megahertz are required. Silicon carbide dominates the rectifier and frontend conversion stages at the 800 V to 400 V level and above where high voltage standoff and reverse recovery behavior dominate selection. Silicon retains dominance at the point-of-load and below 100 watts where cost discipline outweighs the efficiency benefit of wide-bandgap alternatives. Third, hyperscaler dual-sourcing pressure is reshaping the qualification cadence. Where 2018 through 2022 qualification cycles ran 18 to 24 months for a new PSU vendor at hyperscaler tier, the 2024 through 2026 cadence has compressed to 9 to 14 months for incumbent qualified suppliers and expanded to 24 to 36 months for new entrants who lack a prior qualified-vendor record.
Three recommendations follow. For sourcing organizations procuring AI rack power, dual-sourcing at the rack PSU SKU level should be in place by 2027 and single-sourced PSU SKUs should be treated as material supply chain risk; the qualification cycle requires advance planning by approximately 12 to 18 months. For capital planning, the 800 VDC transition should be treated as a power semiconductor decision rather than a transformer-side decision because the inflection in rack architecture pulls SiC and GaN content forward by approximately one full capex generation versus the assumption that power transitions are paced by switchgear refresh. For geopolitical risk management, exposure to China-fab origin power semiconductor and PSU content should be managed explicitly at the supplier list level. Innoscience and other China-domiciled GaN suppliers compete on cost but cannot be qualified for US Government, federal, or hyperscaler-sensitive workloads without specific export-control review under the standards mapping framework documented in Appendix D.
Scenario. The 2030 base case has rack PSU and conversion revenue at approximately 47 billion US dollars globally with GaN content at 28 to 35 percent of new SKU value, SiC content at 12 to 18 percent, and silicon retaining the remainder. The bear case at 34 billion reflects a slower 800 VDC ramp, delayed hyperscaler deployment, and constrained GaN wafer capacity. The bull case at 60 billion reflects acceleration of NVIDIA Rubin and Vera Rubin-class racks, the AWS, Microsoft, Google, and Meta dualsource policy fully implemented, and successful SST commercialization at the facility envelope. The paper develops each scenario in Chapter 9.
Full white paper below
