Solid-state transformers for 800V DC AI data-center power architecture Report Interpretation

Overview

This thematic power-infrastructure report explains why high-density AI computing may accelerate the shift to 800V DC data-center power systems and create an emerging market for solid-state transformers. Bernstein considers SSTs technically compelling and potentially much larger than published market estimates, while emphasizing that commercial adoption is still early.

Core views

Bernstein argues that rack power density is the forcing function behind a transition in data-center electrical architecture. Current AI racks draw roughly 60–100 kW, while next-generation designs are being discussed at up to about 1 MW per rack in a similar footprint. A conventional AC chain of transformer, switchgear, UPS, and rack power supply was not designed for such loads. Because servers ultimately consume DC power, repeated AC/DC and DC/AC conversions waste energy; the report therefore sees 800V DC (±400V DC) emerging as the next-generation distribution standard. An SST replaces the copper-and-iron, line-frequency transformer with high-frequency SiC or GaN power electronics, magnetics, and digital controls. It can convert medium-voltage AC, commonly 10–35 kV in relevant designs, directly to 400V or 800V DC, eliminating separate step-down-transformer, rectifier, and UPS stages. Bernstein highlights bidirectional power flow, sub-millisecond control of power quality and faults, and modular designs of roughly 1.2 MW to 4–5 MW per unit. Compared with conventional transformers, SSTs are described as faster, 50–70% smaller in size and weight, programmable, and compatible with solar, batteries, EV charging, and grid services. The practical case for 800V is current reduction. A 15 kW rack at 48V requires about 312 amperes, while a 1 MW rack at 48V would require about 20,833 amperes—more than 66 times as much current and an impractical amount of copper. Bernstein compares a traditional 15 kW enterprise architecture with 85% efficiency, a current 30 kW hyperscaler architecture with 93% efficiency, and a future 1 MW 800V DC architecture with 96% efficiency. More broadly, industry commentary cited by the report points to roughly 96–98% efficiency for a fully DC-native SST-based design versus an 84–91% AC baseline. SST compactness could also free data-center white space for additional compute racks. The report notes potential resource benefits as well as power-system benefits. An illustrative comparison shows 0.15 kg of copper per kVA for a medium-voltage SST and 0.32 kg per kVA for a low-voltage SST, versus an estimated 0.6 kg per kVA for a low-frequency transformer. Bernstein links lower copper intensity to reduced trade risk, noting that the U.S. imported about 45% of total copper consumption in 2024. SSTs may also displace or consolidate adjacent equipment including voltage regulators, rectifiers, inverters, and reactive-power compensation equipment. Bernstein frames its market sizing as a scenario exercise rather than a forecast. It calculates TAM by multiplying annual AI capacity additions in GW by SST penetration of that year's new capacity and installed SST cost per MW, then cumulating through 2030. Its first capacity scenario, closest to its power-model base case, assumes about 3% CAGR in overall power demand through 2030; the report also cites about 80 GW of incremental power additions. Annual data-center capacity additions range from 9 GW in 2026 to 8–20 GW in 2030 across scenarios. Installed SST-layer cost assumptions are $0.55M, $1.00M, and $1.65M per MW for low, base, and high cases, respectively. On those assumptions, annual SST TAM rises from $0.05B, $0.09B, and $0.29B in 2026 to $1.32B, $4.80B, and $11.55B in 2030 in the low, base, and high cases. Cumulative 2026–2030 TAM is $2.80B, $9.69B, and $24.05B, leading Bernstein to characterize the U.S. opportunity as roughly $3–20B, with further upside if data-center demand exceeds expectations or adoption is faster. This contrasts with published third-party estimates of around $1B for the data-center-specific SST market through 2030, which Bernstein believes may reflect a later commercial start and revenue-recognition framing. The report's high case assumes 35% adoption by 2030; its longer-run low, base, and high penetration curves reach about 35%, 55%, and 75% by 2035. Commercial timing is central to the thesis. Vendor roadmaps indicate full-system demonstrations in late 2026, customer pilots in 2027, and volume shipments beginning in 2028; broader adoption is presented as a 2028–2030 development, with SSTs potentially becoming a dominant architecture choice through the 2030s as component costs fall and the SiC/GaN, magnetics, and systems-integration supply chain matures. Bernstein identifies GE Vernova as furthest along, with 5 MW indoor and 6 MW outdoor SST applications being tested with two hyperscalers. The report sees the next critical milestone as 2027 order booking after a six-month hyperscaler test. The competitive field includes venture-backed pure plays, public companies extending inverter or DC-DC capabilities, and diversified incumbents with transformer and switchgear installed bases. Three startups—Heron Power, DG Matrix, and Amperesand—raised a combined roughly $280M in the twelve months to mid-2026, while Eaton's acquisition of Resilient Power Systems is described as the first meaningful incumbent M&A entry. Enphase has announced a 1.25 MW IQ SST rack, scalable to 5 MW, supporting NVIDIA 800V DC and OCP ±400V DC standards; it targets late-2026 demonstrations, 2027 pilots, and 2028 volume shipments. The report describes Infineon as offering a complete SST semiconductor stack, while Renesas has a mid-term SST front-end roadmap and more established exposure in rack and core power conversion. Bernstein also details the adoption constraints. Semiconductor-system cost, especially SiC or GaN, is the primary hurdle and must decline with scale. Existing electrical systems require costly reconfiguration to support high-frequency SST architecture. Long-term grid reliability remains unproven because installed grid equipment must operate for decades under semiconductor stress. Finally, high-frequency designs introduce hysteresis and eddy-current losses that need further innovation. These constraints explain why the report expects a gradual adoption curve rather than immediate replacement of conventional infrastructure.

Analysis framework

The report first explains the technical power-conversion problem created by higher AI rack density, then compares conventional and 800V DC architectures. It uses operating characteristics and equipment comparisons to explain SST benefits, builds a bottom-up TAM from data-center capacity, penetration, and cost-per-MW scenarios, and reviews pilots, vendor roadmaps, competitive positioning, and commercialization barriers.

Methodology notes

• Industry AnalysisSupply-demand framework

  Bottom-up SST TAM scenario analysis

  Bernstein estimates market potential from annual AI data-center capacity additions, assumed SST penetration, and installed cost per MW, rather than relying solely on published market-size estimates.

• Industry AnalysisPenetration-Rate S-Curve

  Adoption S-curve

  The report models low, base, and high SST penetration curves following pilots and volume shipments, reaching about 35%, 55%, and 75% adoption by 2035.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Power-electronics supply-chain transmission

  The report links data-center power demand and 800V architecture to SST vendors, SiC/GaN suppliers, magnetics, controls, inverters, and incumbent electrical-equipment providers.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• GE Vernova (GEV)

  Covered electrical-equipment company described as furthest along in SST testing and hyperscaler pilots.

  Strengths

  Two SST configurations—a 5 MW indoor system and a 6 MW outdoor system—are being tested with two hyperscalers; its Electrification segment provides a relevant installed-base and partner path.

  Comparison

  The report places GE Vernova ahead of other public companies in SST development and piloting.

  Risks

  Pilot execution and subsequent 2027 order conversion remain key.

• Enphase Energy (ENPH)

  Covered solar-power-electronics entrant developing the IQ Solid-State Transformer.

  Strengths

  Its planned 1.25 MW rack supports NVIDIA 800V DC and OCP ±400V DC standards and can scale to 5 MW.

  Weaknesses

  Commercial demonstrations, pilots, and volume shipments are still prospective.

  Comparison

  Bernstein links Enphase's solar power-electronics heritage to SST development.

  Risks

  Late-2026 demonstrations, 2027 pilots, and 2028 volume-shipment targets must be met.

• Infineon

  Covered semiconductor supplier positioned as an SST component beneficiary.

  Strengths

  The report describes a complete portfolio spanning power switches, gate drivers, microcontrollers, security ICs, and sensors.

  Weaknesses

  Its disclosed global SST semiconductor SAM of USD1B+ within five years is characterized by the report as conservative.

  Comparison

  Described as the only company offering a complete SST semiconductor stack in one portfolio.

  Risks

  Demand depends on SST commercialization and the timing of named customer programs.

• Renesas

  Covered power-semiconductor company with existing rack-and-core exposure and a developing SST front-end roadmap.

  Strengths

  Its products are already designed into AI-server platforms at rack and core layers; its GaN reference design achieved 98% full-load efficiency.

  Weaknesses

  The report says its SST front-end offering is still maturing and requires two further components, expected in 2027–2028.

  Comparison

  More commercially established in rack and core conversion than at the grid-interface SST layer.

  Risks

  SST-front-end roadmap completion and adoption timing are uncertain.

• Eaton (ETN)

  Covered incumbent that entered SSTs through the acquisition of Resilient Power Systems.

  Strengths

  The acquisition brings medium-voltage SST technology in-house.

  Comparison

  The report identifies the acquisition as the first meaningful incumbent M&A entry into the nascent category.

  Risks

  Commercial scale remains dependent on technology cost, customer adoption, and system integration.

• Vertiv (VRT)

  Covered data-center infrastructure company developing SST products.

  Strengths

  Relevant exposure to data-center electrical infrastructure.

  Weaknesses

  The report says it appears to lag Tier 1 players and remains under product development.

  Comparison

  Positioned behind leading SST pilot participants.

  Risks

  Competitive execution and product-development timing.

Key data

• AI rack powerRoughly 60–100 kW today; up to about 1 MW in next-generation designsThe report identifies higher rack density as the key driver of architectural change.
• End-to-end architecture efficiency84–91% AC baseline versus roughly 96–98% for a fully DC-native SST architectureFewer conversion stages are expected to reduce energy losses.
• U.S. SST TAM, cumulative 2026–2030$2.80B low / $9.69B base / $24.05B highBottom-up scenario exercise; Bernstein summarizes the opportunity as roughly $3–20B.
• SST adoption in 203012% low / 24% base / 35% highThe high case assumes fast adoption following commercialization.
• Commercialization timelineLate-2026 demonstrations; 2027 pilots; 2028 volume shipmentsVendor-roadmap timing underpinning the adoption curve.
• GE Vernova SST pilots5 MW and 6 MW configurationsTesting with two different hyperscalers.

Impact & implications

Bernstein sees SSTs as a potentially important layer in 800V DC data-center power systems, with implications for grid equipment, solar-derived power electronics, and SiC/GaN semiconductor suppliers. The report considers the potential market materially larger than commonly cited through-2030 estimates if hyperscalers adopt the architecture, but stresses that the opportunity remains contingent on successful commercialization.

Risks

• SiC and GaN semiconductor costs remain the primary obstacle to large-scale SST adoption.
• Existing electrical systems require costly upfront reconfiguration for high-frequency SST architecture.
• Long-term reliability under grid conditions has not yet been proven for equipment expected to operate for decades.
• High-frequency operation creates hysteresis and eddy-current losses that require further technical improvement.
• The TAM is a scenario exercise and depends materially on data-center capacity, penetration, and cost-per-MW assumptions.

What to watch

• Whether GE Vernova's 5 MW and 6 MW SST pilots with hyperscalers occur on schedule in late 2026.
• Which hyperscalers announce named 2027 pilot deployments or orders.
• Whether 2028 volume-shipment guidance holds or slips.
• The pace of SiC/GaN cost declines and supply-chain maturation.
• Progress in long-term reliability testing and system-level integration.