Envision deploys 800V DC AI data centre system

Envision has put an integrated 800V direct-current power system into commercial operation at its Galaxy AI data centre campus in Ulanqab, Inner Mongolia. The installation is already supplying live computing workloads through a combination of solid-state transformers, battery energy storage, and coordinated power controls. Electricity passes from the medium-voltage incoming supply to the computing racks through a common DC architecture intended to accommodate the high and rapidly varying loads associated with AI accelerators.

Incoming electricity is supplied at 10kV alternating current and converted to an 800V DC distribution bus by solid-state transformers. These use power electronic conversion rather than a conventional transformer followed by a separate series of large conversion stages. Additional DC conversion equipment then supplies the voltages required by computing hardware. The arrangement moves voltage regulation and energy storage into a coordinated electrical system instead of treating the racks, standby supply, and incoming network connection as independent installations.

In a conventional data centre, the electrical route can include AC distribution, rectification into a battery-backed uninterruptible power supply, inversion back to AC, and further rectification within server equipment. Supplying racks over a high-voltage DC bus can remove some intermediate conversions, although semiconductor loads still require precisely regulated power. The practical savings depend on the selected converter topology, rack interfaces, fault protection, efficiency at different loading levels, and the installation’s existing electrical layout.

Distribution voltage also affects conductor current. For a given power level, an 800V bus can carry electricity at a lower current than a lower-voltage DC arrangement, reducing some cable and copper requirements. Actual conductor size and losses depend on route length, permissible temperature rise, protective devices, and how current is shared among loads. DC fault interruption and safe isolation require dedicated engineering because the absence of regular current zero crossings changes how arcs and protective equipment behave.

Envision has integrated battery storage into normal power-system coordination as well as backup provision. The company describes second-scale support for changes in demand and minute-scale reserve during interruptions, while communication between the storage system and solid-state transformers operates at millisecond timescales. Storage can absorb or supply energy while converter controls adjust, limiting the size and speed of changes seen by upstream electrical equipment. Those functions have different requirements for battery power rating, available energy, converter response, and state-of-charge management.

Large groups of graphics processing units can move abruptly between computational states, producing sharp changes in the demand at a rack and at the wider distribution bus. Envision’s Dubhe Controller, supported by its Dubhe Energy Foundation Model, coordinates transformer operation and the battery system to respond to these changes. The control equipment manages the interface between the grid, locally stored electricity, and computing loads, where voltage stability and reserve availability have to be maintained simultaneously.

During a rapid rise in computing demand, the battery and power converters can be coordinated so that the incoming supply does not have to follow the entire step change instantaneously. When load falls, the same controls can adjust conversion and storage operation to handle the reduced demand. A DC architecture still needs protection selectivity, fault containment, converter ride-through, and maintenance isolation. Those measures have to be designed for the actual configuration and loading of the campus rather than inferred from the nominal 800V rating alone.

The Galaxy campus began operating in August 2026 and has a planned full-build capacity of 3GW. Envision’s wider plans provide for as many as one million AI accelerators and target more than 80% of the campus’s electricity from direct renewable power connections. These figures describe a planned campus configuration. The company has not specified the electrical capacity of the live 800V installation, its battery energy rating, or the current renewable electricity share attributable specifically to that power system.

Kevin Huang, senior vice-president of Envision Energy and president of its energy storage and AI data centre power systems business, has linked the deployment to coordinating renewable generation, grid supply, and batteries with computing demand. Such coordination requires both electrical response at the rack and energy management across the site, especially when renewable output and processing workloads vary on different timescales. The operating campus gives the company an installation on which to test those controls under real computing demand.

Envision plans to expand the Galaxy campus and has set a wider Mission Gobi target of 5GW of green AI data centre capacity in desert regions by 2030. As additional racks and power equipment are installed, distribution capacity, isolation arrangements, converter redundancy, battery reserve, and incoming grid connections will need to be matched to the growing load. The current installation establishes the use of the integrated 800V arrangement at an operating site, with its larger rollout dependent on the subsequent campus build-out.