Envision Turns On World's First Commercial SST-Based 800V DC Power at Ulanqab AI Campus
Galaxy Campus converts 10 kV AC directly to 800V DC in one stage as competitors remain in prototype phases
Envision has put what it describes as the world's first commercial solid-state-transformer-based 800-volt DC power system into live operation at its Galaxy AI campus in Ulanqab, Inner Mongolia, the company announced on November 2. The deployment moves an SST-plus-battery architecture from laboratory development to commercially operating AI workloads — a transition the rest of the industry has promised but has not yet delivered.
The significance lies in what the system replaces. Conventional AI data center power chains route medium-voltage AC through a step-down transformer, then through an uninterruptible power supply, then through AC-to-DC rectification, before reaching a 380–400-volt DC bus at the server. Each stage bleeds efficiency. Envision's design eliminates the intermediate steps, converting 10-kilovolt medium-voltage AC directly to an 800-volt DC bus in a single integrated unit.
Inside the Conversion: How SST Collapses Three Stages Into One
A solid-state transformer uses silicon carbide MOSFET switches operating at tens of kilohertz rather than the 50 Hz iron core of a conventional transformer. High-frequency switching allows galvanic isolation and voltage conversion in a unit far smaller than its iron equivalent. Envision's production module is a containerized 2.5-megawatt unit that accepts 10-kilovolt or 13.8-kilovolt AC and delivers 800-volt DC output directly — completing what a conventional facility would require three separate pieces of equipment to accomplish.
Coordinating the SST with battery energy storage is Envision's Dubhe Controller, which runs the company's Dubhe Energy Foundation Model — an AI system trained on power telemetry and designed to manage the interaction between transformer output and battery charge and discharge at millisecond timescales. The coordination layer addresses a structural problem in large-scale GPU clusters: rack power demand can swing from partial load to full draw in under a hundred milliseconds as training jobs ramp up or inference tasks arrive in bursts. Without fast coordination, the SST must be over-provisioned for peak demand, wasting capital. With battery absorbing instantaneous spikes while the SST tracks sustained load, the system sizes closer to average draw. The battery also provides second-scale support and minute-scale backup. Notably, the Dubhe Energy Foundation Model makes Envision's installation the first known deployment of a trained AI model managing AI data center power in real time — a feedback loop where AI infrastructure optimizes itself.
Envision claims the architecture achieves up to 98.5% conversion efficiency and reduces copper cabling requirements by up to 80% compared with conventional chains. Those figures are company-reported and have not been independently verified. Conventional multi-stage chains typically achieve 93–96% total efficiency; a genuine 2.5-point improvement at gigawatt scale reduces waste heat and cooling load meaningfully, but independent benchmark testing of SST systems at this power level does not yet appear in publicly available sources.
Why 800V DC Is Becoming a Mandatory Requirement for GPU Rack Deployments
The move to 800-volt DC is not arbitrary. NVIDIA's Kyber rack system, designed for the upcoming Vera Rubin Ultra generation of accelerators, requires 800-volt DC input, meaning any data center deploying those racks at scale must eventually deliver 800 volts to the row. SemiAnalysis projects 800-volt DC adoption will accelerate sharply, potentially approaching 80% of new global data center capacity by 2030 as GPU rack design requirements increasingly mandate the architecture.
All major competitors have announcements. Enphase is targeting full system demonstrations in late 2026, with customer pilots in 2027 and volume shipments in 2028. Sungrow launched its EnerNeo SST system in July 2026. Siemens and Reinhausen announced an SST development partnership in August 2026. SolarEdge and Infineon jointly announced an SST initiative in November 2025 targeting more than 99% efficiency. None of these constitutes a commercial deployment on live AI workloads. Envision's Galaxy campus does — and that gap between announced intention and operating reality is the commercial position Envision now holds.
Galaxy Campus and the Mission Gobi Replication Blueprint
The Galaxy campus was commissioned on August 6, 2026, anchored by a 120,000-square-meter facility. Envision plans 3 gigawatts of total capacity by 2028, supporting up to one million AI accelerators simultaneously, with more than 80% of power from direct renewable connections. Ricky Zheng, Envision's General Manager of AI Data Center, confirmed the 3-gigawatt target and 2028 timeline in a Bloomberg interview on October 30.
The campus is also the first deployment under Mission Gobi, an initiative unveiled at VivaTech Paris in June 2026 targeting 5 gigawatts of green AI data center capacity in desert regions globally by 2030. Envision markets Galaxy at electricity prices 20% below grid rates under a 10-year fixed contract — an economics made possible by its vertical integration across wind turbines, battery storage, and now data center power systems.
What the Company Claims and Where Independent Verification Is Still Missing
The "world's first commercial deployment" designation is Envision's own. No independent body has certified the claim, and the efficiency figures carry the same limitation. A 98.5% SST efficiency across a 2.5-megawatt containerized unit is also a figure measured under specified conditions; real-world performance under load cycling, thermal stress, and multi-year operation remains to be demonstrated. The medium-voltage input requirement is a further constraint: not every candidate site has direct 10-kilovolt distribution, and extending it adds cost and permitting complexity that varies by location.
Enphase's 2027 customer pilot program will provide the next independent data point on whether SST-based 800V DC can be replicated at commercial scale. NVIDIA's Kyber ramp toward its second-half 2027 target will make the architecture a near-mandatory requirement for competitive GPU cluster builds regardless of vendor. The more durable test for Envision is whether the Dubhe Energy Foundation Model's millisecond-scale power coordination sustains the efficiency numbers in production — a question that only operational data from the Galaxy campus can now answer.