Overview

Envision Group said on November 2, 2026 that an 800V DC power architecture is running live AI workloads at its Galaxy campus in Ulanqab, Inner Mongolia, describing it as a world first. Solid-state transformers convert 10 kV grid AC directly to an 800V DC bus, removing the low-voltage transformer, UPS and rectifier stages used in conventional data centers. SemiAnalysis, which visited the site, called it 'China's Stargate' and reported a plan for up to 800 MW of IT capacity, with the first 120 MW built within a year at about a fifth of US costs.

How the 800V DC Architecture Works

The key device is the solid-state transformer, which uses power semiconductors rather than copper coils and iron cores. It accepts 10 kV AC from the grid and outputs an 800V DC bus directly, combined with battery storage and DC-DC converters for the IT load. A conventional AC facility instead passes power through a low-voltage transformer, a UPS and a rectifier before the rack. Envision's design removes all three stages.

  • Grid input: 10 kV AC, followed by direct conversion to an 800V DC bus.
  • Low-voltage transformer: required conventionally; removed through the solid-state transformer.
  • UPS: required conventionally; replaced by integrated battery storage.
  • Rectifier: required conventionally; removed because the SST outputs DC directly.

Prefabricated substations can be built in under three months, compared with about six months conventionally. Solid-state transformers are compact and reportedly more than 98.5% efficient, though no official single figure has been published.

The Galaxy Campus: Scale, Power and Cost

Galaxy is planned for up to 800 MW of IT capacity across four phases. Phase one is 120 MW, delivered in August 2026 and built within a year; the full build-out target is the end of 2027. Power comes mostly from Envision's own wind farms plus four-hour batteries. Those wind farms total 420 MW, about 1.75 times the 240 MW load, linked by a dedicated line, and Envision aims to supply more than 60% of the first two phases from them. Liquid- and air-cooled servers share one water loop with no central chiller. Power costs about $0.055 per kWh all-in, and Chinese data center rents have roughly halved in five years to about $40 per kW per month. Campus cost is about a fifth of US costs, or under $3 million per MW.

Market Context: China vs. US AI Infrastructure

The contrast with US projects is stark. Stargate Abilene, associated with Crusoe, Oracle and OpenAI, is planned for 1.2 GW of power and about 820 MW of IT capacity, with construction beginning in mid-2024; it relies on grid power plus on-site gas turbines. US AI labs have increasingly bought gas generators to avoid grid queues, as shown by Anthropic's $3.4 billion generator deal. Galaxy, by contrast, is built around wind generation, battery storage and DC distribution. The comparison suggests China is building comparable AI campuses faster and cheaper, even if the long-term role of DC distribution remains uncertain.

Industry Impact: Efficiency, Footprint and Supply Chain

Removing three conversion stages has direct consequences. Each conversion step wastes energy and occupies floor space; eliminating the low-voltage transformer, UPS and rectifier can reduce both losses and equipment footprint. Higher voltage also lowers current for a given power level, reducing resistive losses and copper requirements. That matters as AI racks grow from tens of kilowatts to hundreds of kilowatts. An Nvidia Vera Rubin rack is in the 250 kW class, and delivering that at low voltage would require very large currents and thick copper. Nvidia has signaled plans for 800V DC in future racks. Galaxy therefore tests not only solid-state transformer reliability but also the surrounding ecosystem of DC-DC converters, batteries and protection equipment.

Implications for AI Data Centers

For AI data centers, the 800V DC approach addresses a growing bottleneck: low-voltage distribution is straining under high-density racks. An 800V DC bus can serve racks with less conversion hardware and potentially better efficiency. Integrated battery storage replaces the UPS, and prefabricated substations shorten construction timelines. The design may also simplify coupling with renewable generation and on-site storage, as Galaxy demonstrates with wind and four-hour batteries. However, the article notes that the total size of the 800V DC system actually in service was not disclosed; SemiAnalysis describes a live 5 MW pilot hall. Which AI accelerators are running at the site is also unknown. Those gaps matter for judging whether the architecture is ready for broad deployment or remains a large-scale pilot.

What Remains Unclear and Future Outlook

Several claims still require independent confirmation. Envision's 'world first' label is the company's claim, and the in-service capacity of the 800V DC system has not been published. The 800 MW campus figure comes from SemiAnalysis; another report mentions a planned 3 GW AI campus in the area, but the relationship was not confirmed. Efficiency above 98.5% appears in only one report. Even so, the direction is clear. If Galaxy scales from its first 120 MW phase toward the planned 800 MW by the end of 2027, it could become a reference design for high-density AI campuses combining renewable power, battery storage and DC distribution. The next milestones are official efficiency data, the identity of the AI chips deployed, and whether solid-state transformers and 800V DC racks move from pilot halls into standard data center supply chains.