Solid-State Transformers and the AI Factory
Advisory Reference: Architecture, Standards, and Capital Posture for SST Deployment
Abstract
This white paper provides an executive technical reference on solid-state transformers (SSTs) and their role in the power architecture of AI factories. The central position is that SSTs are no longer a research curiosity. Multiple original equipment manufacturers, several hyperscale operators, and at least two well-funded specialist firms have moved SSTs from the laboratory into engineered product programs targeted at AI data center deployment between 2026 and 2029. The architecture is unfamiliar to many practitioners and authorities, but unfamiliarity is not the same as immaturity, and immaturity is not the same as unsuitability. This paper develops those distinctions in detail.
The paper is organized in eight parts. Part I establishes the analytical framing: what an SST is, why AI factories compress the power chain, and how the conversation differs from prior generations of converter technology. Part II develops the technology fundamentals: silicon carbide and gallium nitride device physics, high-frequency magnetic design, dual-active-bridge and modular-multilevel-converter topology selection, and the thermal and reliability envelope of a working SST. Part III places SSTs in the AI factory context: rack-scale densities at the gigabit-per-second AI inference and training boundary, NVIDIA Mt. Diablo and 800 VDC sidecar architecture, the disaggregated power rack, and the grid-to-chip voltage stack. Part IV addresses architectures and topologies in commercial development. Part V presents the manufacturing, supply-chain, and OEM landscape with hard product, capacity, and timeline data. Part VI addresses standards, codes, and governance: National Electrical Code, Underwriters Laboratories, NFPA, IEEE, IEC, and OCP / NVIDIA reference specifications. Part VII addresses operations, telemetry, lifecycle, spares, commissioning, and cybersecurity. Part VIII addresses capital deployment, total cost of ownership, decision frameworks, and the architecture authority required to execute SST programs at hyperscale.
The paper distinguishes verified facts, considered analysis, structured inference, and explicit assumption throughout. It is calibrated to programs operating at densities at or projected to exceed two megawatts per rack during the 2026 through 2029 deployment window. It includes more than forty figures, more than twenty tables, and a comprehensive reference list of standards, listings, products, and program sources discussed.
Executive Summary
Position
The conventional power chain in a hyperscale data center has six conversion or transformation stages between the transmission grid and the central processor. Each stage costs efficiency, footprint, copper, and lead time. AI factories raise rack power between five and forty times above the densities for which that chain was designed. The result is the architectural equivalent of running a Formula One engine through a domestic transmission. The transmission has not yet failed, but it is the wrong gearbox for the engine that is now fitted.
The solid-state transformer is the gearbox redesign. It replaces the iron-core medium-voltage transformer and several downstream conversion stages with a single power-electronic converter that takes medium-voltage alternating current directly to 800 volt direct current at the rack. It is a single piece of equipment with three functional domains: primary rectification at medium voltage, isolated direct-current to direct-current conversion through a high-frequency transformer, and low-voltage direct-current output regulation and protection. Modern SSTs use silicon carbide power devices on the medium-voltage side and silicon carbide or gallium nitride on the low-voltage side.
Why now
Three forces have collapsed onto each other to make 2026 through 2029 the decisive deployment window for SSTs in AI factories. First, NVIDIA has published an 800 VDC architecture and demonstrated an 800-volt sidecar for the Rubin Ultra Kyber rack. Second, Microsoft, Meta, and Google jointly published the Mt. Diablo disaggregated power specification through the Open Compute Project, defining a sidecar rack architecture that operates at plus-or-minus 400 VDC or 800 VDC and supports racks from 100 kilowatts to 1 megawatt. Third, multiple commercial SST programs have moved into production engineering: Hitachi Energy has committed an additional 1.5 billion United States dollars to expand transformer and SST production; Amperesand raised an 80-million-dollar Series A and announced 30 megawatts of commercial systems for delivery in 2026; Enphase announced a 1.25-megawatt distributed SST architecture targeting volume shipments in 2028; Vertiv has published a dual-path roadmap with a medium-voltage direct-current uninterruptible power supply at 34.5 kilovolts in 2027 and a 13.8-kilovolt SST in 2028; Eaton, LS Electric, Sieyuan, and Siemens have all disclosed product or research programs aligned to the same window.
What changes
Four things change for the operator who adopts SSTs. The first is footprint: a five-megawatt SST block occupies roughly seventy percent less floor area than the equivalent conventional medium-voltage to low-voltage transformer and uninterruptible power supply gallery. The second is conversion stage count: the chain from utility to chip drops from seven stages to four, and chain efficiency rises from approximately ninety-three percent to a target of ninety-eight to ninety-nine percent. The third is grid services: the SST is bidirectional, supports four-quadrant reactive power, can ride through grid faults, and integrates direct-current energy storage natively. The fourth is telemetry: the SST is a software-defined asset whose internal state is fully instrumentable, where the iron-core transformer is a black box that reports temperature and oil chemistry.
What does not change
Three things do not change, and pretending they do invites avoidable cost. SSTs are still subject to the National Electrical Code, to Underwriters Laboratories listing requirements where they apply, to NFPA fire and life safety standards, and to the authority having jurisdiction. The architecture is permitted by the existing code framework. It is not unfamiliar to the framework; it is unfamiliar to many of the inspectors and engineers who operate within the framework, and the gap between those two facts is exactly where AHJ engagement effort goes. SSTs also do not remove the operator’s obligation to maintain spares depth, factory acceptance test discipline, commissioning rigor, and cybersecurity posture; they relocate those obligations onto power-electronics modules and control firmware rather than onto oil samples and bushing inspections. Finally, SSTs do not yet remove the cost premium. In 2026, an SST is still roughly two-and-a-half to three times the bill-of-materials cost of a conventional medium-voltage transformer of equivalent rating.
Recommendation
FCG advises operator clients to prepare for SST adoption now, on a risk-managed pilot footing, rather than waiting for the listing and code landscape to fully mature. The recommended posture has four elements: first, run an architecture authority review on the existing power chain to identify which stages are SST-replaceable and which are not; second, engage the authority having jurisdiction during pre-design rather than during permit; third, structure a pilot procurement that ties payment milestones to factory acceptance, site acceptance, and post-soak performance rather than to delivery; fourth, place the SST under the operator’s existing cybersecurity governance, not under the original equipment manufacturer’s. The capital exposure of a 2026-2027 pilot is modest, and the cost of being unprepared in 2028 when peers begin deploying at scale is significant.
Full white paper below
