China Switches On World's First 800-Volt DC Power System for AI Data Center

Envision Energy says it has completed the world's first commercial deployment of an integrated 800-volt DC power system inside a data center, and the site it powers is built to scale to 3 gigawatts and, eventually, one million AI accelerators. The system went live at the company's Galaxy AI data center in Ulanqab, Inner Mongolia, where Envision has been testing it since August with its Dubhe Controller and the Dubhe Energy Foundation Model doing the balancing.

The commercial stakes are straightforward. AI racks are starving power grids, and every conversion stage between the grid and the GPU burns energy and eats floor space. Envision's answer is to skip the intermediate steps.

What does the 800V architecture actually change?

Conventional data centers push grid power through several conversion stages before it becomes a usable supply. Envision's Dubhe technology converts electricity directly from 10 kV medium-voltage AC to an 800V DC bus, which the company says raises efficiency and slashes the size of the power conversion hardware.

The system rests on three capabilities, according to the company's announcement:

• Stable solid-state transformer (SST) conversion directly from 10 kV medium voltage to an 800V DC bus across operating conditions.
• Integrated battery storage delivering both second-scale power support and minute-scale backup.
• Millisecond-level coordination between the SST and the storage, absorbing GPU load fluctuations locally.

That last point matters most for AI workloads. Training clusters don't draw steady power; they spike. The batteries in the loop respond to workload bursts and cover short-term backup duties, while the AI-driven Dubhe controller matches power supply against computing demand in real time.

Why does density matter here?

Higher rack density means more computing power in the same floor area, and Envision says the 800V architecture supports exactly that. The company also claims a more compact footprint overall and a reduced requirement for copper and other cabling — a real cost lever at gigawatt scale, where cading and conversion hardware can dominate the electrical budget.

The design is also meant to pair well with renewables and storage, which fits its geography. Ulanqab sits in Inner Mongolia, a region with abundant wind and solar resources, and Envision frames the whole architecture as green-compute infrastructure rather than a pure efficiency play.

By the firm's own account, the Dubhe system is performing well since testing began in August. Those are vendor claims, not independent measurements, but the deployment itself is operational — a first among data centers worldwide, per the announcement.

What comes next?

The Ulanqab installation is a beachhead. Envision says the tech will be instrumental in expanding compute resources at the Galaxy AI data center, which current plans grow to 3GW of capacity and, ultimately, support for up to one million AI accelerators.

The bigger target sits farther out in the desert. Envision also plans to roll the 800V DC technology into its Mission Gobi initiative, which aims for 5GW of green AI data center capacity in desert regions by 2030.

If the vendor's efficiency and density claims hold at scale, the Ulanqab deployment becomes the reference case for 800V DC power as the default architecture for hyperscale AI compute — and a data point competitors building similar high-voltage DC schemes will have to answer.