Envision deploys 800-volt power system for live AI workloads in China

Envision says a new power system is now supplying live AI computing workloads at its Galaxy campus in Ulanqab, Inner Mongolia. Announced November 2, 2026, the deployment combines electricity conversion, batteries and automated controls in an integrated 800-volt direct-current design. The company’s pitch: deliver power to denser computing equipment with less supporting infrastructure.

Envision calls this the world’s first commercial deployment of an integrated 800V DC power system in a large-scale AI data center.

From grid electricity to the server rack

The power path starts with a 10-kilovolt medium-voltage input. Envision says its solid-state transformers convert that supply directly to an 800-volt DC bus—the shared electrical connection that distributes direct-current power. The integrated system also includes equipment that adjusts the DC voltage for computing hardware, extending the power path from the grid input to the server rack.

Conversion is only one part of the installation. Battery energy storage sits alongside the transformers, while a controller coordinates the two.

The business case rests partly on the space and materials used to deliver electricity. Envision says the architecture improves conversion efficiency, makes power equipment more compact, and reduces cabling and copper requirements. Those changes can lower infrastructure costs and accommodate more computing power within the same space, according to the company.

Envision did not publish measured efficiency gains or cost savings in its announcement.

AI data centers need power systems that coordinate renewable energy, energy storage, and the grid to meet computing demand.

Kevin Huang, senior vice president of Envision Energy

Three different timescales for keeping power stable

The control problem is not simply supplying enough electricity overall. Envision says computing loads can swing sharply from moment to moment, while operators must deliver more power to each rack within limited space. Its Dubhe Controller, supported by the Dubhe Energy Foundation Model, coordinates the transformers and batteries in real time to manage those swings.

The company describes three distinct timing functions. Coordination between the transformers and storage operates at millisecond scale to absorb local fluctuations in demand from graphics processing units, the chips used for AI computing. Battery storage supplies power support on a seconds timescale, while its backup function operates on a minutes timescale.

An operating system inside an unfinished campus

Galaxy began operations in August 2026, according to Envision. Bloomberg’s account, published by Energy Connects, says Envision expects construction to finish by the end of 2028.

At full build-out, the campus has a planned capacity of 3 gigawatts and is designed to support up to one million AI accelerators operating in parallel, the company says.

The renewable-energy figure is also a target: Envision plans for more than 80% of Galaxy’s power to come through direct renewable-electricity connections.

Envision intends to scale this approach through Mission Gobi, its initiative targeting 5 gigawatts of green AI data-center capacity in desert regions worldwide by 2030.