AI compute expansion is turning power into the core bottleneck, and 800VDC will reshape data center power semiconductors
AI summary card
AI compute expansion is turning power into the core bottleneck, and 800VDC will reshape data center power semiconductors
BofA expects AI rack power to rise from the traditional 10-15kW to the 600kW-to-1MW range, driving AI analog semiconductor TAM to about $27 billion by 2030 and benefiting vendors with broad power portfolios and system-level design capabilities.
- AI training and inference demand are increasing rack power density, with Nvidia's Feynman-era rack power potentially reaching the 1MW level.
- Traditional 48V/54V power architectures face bottlenecks in space, copper usage, and conversion efficiency, while 800VDC could reduce losses, copper usage, and system complexity.
- The report estimates from the bottom up that CY25-CY30 will require about 233GW of additional AI compute-related power capacity cumulatively.
- The AI analog semiconductor market is expected to expand from about $7.9 billion to $27 billion by CY30, with rack-to-core TAM of about $25 billion and a grid-to-data hall opportunity of about $1.8 billion.
Report interpretation
Overview
This report discusses the industry inflection point in AI data centers shifting from 'compute expansion' to 'power constraints.' As GPU/XPU, networking, and rack-level integration improve, traditional cloud rack power consumption of 10-15kW is evolving toward Blackwell's 100-120kW, Rubin Ultra's 600kW+, and the Feynman era's 1MW class. The existing 48V/54V power supply system cannot support this level of power density, prompting a full redesign of data center power architecture from the grid to the rack and then to the chip core.
Core views
The core view is that power will become the key constraint on AI scaling, and 800VDC architecture, energy storage, SST/SSCB, and microgrids will gradually enter data center power systems. Analog ICs, power devices, VRMs, IBCs, PSUs, protection, sensing, analog content related to optical modules, and SiC/GaN materials will benefit. Suppliers such as TXN, Infineon, ADI, and ON, with broad product portfolios, high reliability, and system-level design capabilities, are best positioned to gain share.
Analysis framework
The report uses a bottom-up AI analog semiconductor industry model, converting GPU/XPU and custom ASIC shipments plus rack demand into GW demand, then breaking down TAM and supplier revenue opportunities by low-power and high-power racks, power-supply layers, component types, and device materials.
Methodology notes
Deriving power semiconductor content volume from accelerator and rack demand
The report maps AI accelerators, rack power density, and deployment pace into content pools such as VRM, IBC, PSU, SST, SiC, and GaN, and estimates supplier revenue share.
Evolution from traditional 415VAC/48VDC toward 800VDC and microgrids
The report compares the space, copper, and efficiency bottlenecks of traditional multi-stage AC/DC and DC/DC conversion architectures, and analyzes the phased migration toward sidecar power, MV rectifiers, SST, and ultimately DC microgrids.
Assessing share opportunity based on product breadth, high-voltage reliability, and system optimization capability
The report argues that the best suppliers need to cover multiple device types from grid-to-core and be able to meet high-voltage, high-reliability, and system-level design requirements.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Texas Instruments (TXN)One of the main beneficiaries in AI power analog semiconductors
- Strengths
- It has a leading power semiconductor business and a broad analog product portfolio, providing a share base in multi-phase VRMs, power management, and system design.
- Weaknesses
- The report also notes that competition will come from rivals with more complete SiC/GaN and high-voltage portfolios.
- Comparison
- The report views TXN as having the highest market share in this market.
- Risks
- AI data center power architecture migration is slower than expected, or customers adopt more vertically integrated power solutions.
- InfineonOne of the broadest suppliers covering from grid-to-core
- Strengths
- Its portfolio spans Si, SiC, and GaN, and it has automotive/industrial experience in high-voltage, high-reliability power devices.
- Weaknesses
- SST, MV rectification, and the high-voltage DC ecosystem are still evolving, so revenue realization depends on the pace of architecture deployment.
- Comparison
- The report believes Infineon may achieve the largest share gain during CY25-CY30.
- Risks
- SiC/GaN ramp-up, cost curves, and customer qualification progress may fall short of expectations.
- Analog Devices (ADI)Beneficiary in analog ICs, sensing, signal chain, and power management
- Strengths
- It has high-value content in signal chain, sensing, and power management, and the Empower acquisition strengthens related opportunities.
- Weaknesses
- Compared with vendors that have broader power-device material portfolios, its exposure to some high-power discrete devices may be lower.
- Comparison
- The report says ADI has the third-largest revenue opportunity and is positioned to gain share.
- Risks
- If value shifts more toward discrete power devices or customer in-house modules, revenue leverage may be constrained.
- ON Semiconductor (ON)Beneficiary in high-power segments such as SiC/GaN and intermediate bus conversion
- Strengths
- It has high leverage to emerging SiC and GaN technologies and may increase wallet share with customers.
- Weaknesses
- The opportunity depends on accelerated adoption of high-power racks and 800VDC architecture.
- Comparison
- The report lists ON as a supplier with strong leverage in new-material power devices.
- Risks
- Intensifying SiC/GaN competition, pricing pressure, or volatility in the AI demand cadence.
- STMicroelectronics, Renesas, and other analog/power suppliersShare incremental content in optics, protection, sensing, and the power chain
- Strengths
- They benefit from increasing complexity and higher reliability requirements in data center power architecture.
- Weaknesses
- The report gives them less prominent relative positioning than TXN, Infineon, ADI, and ON.
- Comparison
- More skewed toward opportunities in niche components and localized content.
- Risks
- Customer platform choices and changes in architectural standards may affect penetration of individual components.
Key data
- AI analog semiconductor TAMAbout $7.9 billion to $27 billionThe report estimates growth from about $7.9 billion currently to about $27 billion by CY30, implying a CAGR of about 28%.
- rack-to-core TAMAbout $7.6 billion to $25 billionTAM related to the data center rack-to-chip-core segment is expected to reach about $25 billion by CY30, implying a CAGR of about 27%.
- grid-to-data hall TAMAbout $245 million to $1.8 billionThe grid-to-data hall opportunity is expected to reach about $1.8 billion by CY30, implying a CAGR of about 49%.
- Cumulative AI power demandAbout 233GWThe report estimates from the bottom up that CY25-CY30 AI accelerator demand corresponds to about 233GW of cumulative incremental capacity.
- Annual GW additionsAbout 17GW to 60GWAI compute-related annual GW additions could rise from about 17GW in CY25 to about 60GW in CY30.
- Rack power progression10-15kW to the 1MW classTraditional cloud racks are about 10-15kW, Blackwell is about 100-120kW, Rubin Ultra exceeds 600kW, and the Feynman era could enter the 1MW class.
- Rack content valueAbout $36,000 to nearly $1 millionThe report estimates semiconductor content per rack rising from about $36,000 currently to about $300,000 for a 600kW rack, and close to $1 million for an MW-class rack.
- 800VDC efficiency and material benefitsEfficiency improvement of up to about 5%, copper reduction of about 45%Citing Nvidia metrics, the report says 800VDC can reduce conversion losses and lower copper demand.
- TCO improvementUp to about 30%Centralized and simplified power conversion could reduce maintenance and system complexity, with maintenance costs potentially falling by up to about 70%.
- Data center power demandAbout 100GW to nearly 300GWChart data indicate that cumulative global data center power demand could rise from about 100GW currently to nearly 300GW by CY30.
Impact & implications
The industry implication is that the bottleneck in AI infrastructure is expanding from pure chip compute power to power acquisition, conversion, distribution, and stability. Data center operators need higher voltage, fewer conversion stages, stronger energy-storage buffering, and more reliable power devices; semiconductor vendors, meanwhile, have an opportunity to transfer their historically cyclical automotive/industrial power experience into the longer-duration, higher-growth AI data center market.
Risks
- 800VDC equipment, SST, SSCB, and the high-voltage DC ecosystem mature more slowly than expected.
- Data center grid interconnection, transformers, and gas turbine delivery timelines could delay the rollout of AI power capacity.
- AI training load volatility could create pressure on grid stability and energy storage costs.
- If AI capex or GPU/XPU demand comes in below expectations, the 233GW cumulative demand and $27 billion TAM estimates may be revised down.
- Customers may compress external semiconductor supplier share through in-house power modules, system integration, or alternative architectures.
What to watch
- Rack power and production ramps for Nvidia Rubin Ultra, Feynman, as well as AMD, Intel, and custom ASIC platforms.
- The actual adoption path of 800VDC in new data centers and retrofits of existing server rooms.
- Commercialization progress of MV rectifiers, SST, SSCB, ESS/UPS, and DC microgrids.
- Qualification, cost, and supply capability of SiC and GaN in high-voltage power conversion for data centers.
- Revenue disclosure and order visibility for suppliers such as TXN, Infineon, ADI, and ON in AI power-related businesses.