Envision Deploys First 800VDC AI Data Center in Ulanqab
Green tech developer streamlines power conversions for AI clusters
Green technology firm Envision has deployed what it describes as the world's first operational 800VDC power architecture at its Galaxy AI data center in Ulanqab, China. The direct-current distribution system powers live computing workloads while eliminating intermediate alternating-current conversion stages that waste significant power in AI server halls.
Key details
Envision's deployment at the Galaxy AI campus in Inner Mongolia integrates solid-state transformers, battery energy storage systems, and specialized direct-current voltage regulation hardware designed for AI compute racks.
Conventional data center power delivery routes alternating current (AC) from the grid through multiple conversion steps: AC to direct current (DC) for central battery systems or uninterruptible power supplies, back to AC for floor-level distribution to server racks, and finally to DC within each server power supply unit. By contrast, the 800VDC architecture converts incoming 10kV medium-voltage AC grid power directly to an 800VDC busbar using solid-state transformers, delivering direct current straight to high-density compute hardware in a single primary conversion step.
To manage power stability across intense computing bursts, Envision implemented an AI-assisted controller known as Dubhe, supported by its proprietary Dubhe Energy foundation model. The system coordinates solid-state transformers and battery storage in real time to match fluctuating computing demand with power availability and prevent voltage sags during abrupt workload swings.
The Galaxy AI campus spans 120,000 square meters (approximately 1.2 million square feet) and carries a planned design capacity of 3 gigawatts engineered to support up to one million AI accelerators. According to Envision, the complex currently draws 80 percent of its electricity from regional renewable generation assets.
Why this matters
Next-generation AI accelerators and server racks demand unprecedented power densities, with individual racks exceeding 100 to 200 kilowatts. Traditional low-voltage AC power distribution incurs severe thermal dissipation and resistive line losses, requiring thick copper busbars and substantial facility cooling energy just to distribute electricity inside server halls.
Stepping transmission voltage up to 800V direct current slashes electrical transmission losses, reduces conductor copper requirements, and bypasses multiple stages of power conversion heat generation. If successfully scaled, high-voltage direct-current architectures could substantially curb the auxiliary parasitic electricity overhead needed to power AI superclusters.
Context
The Open Compute Project released technical specifications earlier this year outlining 800VDC distribution standards using busbars and high-voltage cabling to satisfy escalating AI accelerator power requirements. While multiple global hardware vendors and electrical engineering firms are developing solid-state transformers and 800V power gear, Envision's facility represents the first reported commercial site operating live production workloads on an 800VDC busbar.
Ulanqab in Inner Mongolia has emerged as a major hub for high-performance computing in Northern China due to abundant wind and solar resources and naturally cool climates, positioning the region as a testbed for integrating renewable energy with large-scale AI infrastructure.
Risks and open questions
While 800VDC distribution promises significant theoretical efficiency gains, widespread adoption faces technical and supply chain hurdles. Solid-state transformer equipment and 800VDC circuit breakers remain scarce, highly specialized, and more costly than mature alternating-current switchgear.
Additionally, data center operators must validate long-term component reliability, arc flash safety protocols, and thermal tolerances under continuous multi-megawatt switching stresses before hyperscale operators transition away from established power delivery topologies.
What happens next
Envision plans to expand its renewable computing buildout, targeting 5 gigawatts of green data center capacity online by 2030 under its broader Mission Gobi program.
Hardware engineers and data center operators will monitor performance and efficiency metrics from the Ulanqab deployment as standard-setting bodies refine high-voltage direct-current guidelines for next-generation AI infrastructure.
Source: Data Center Dynamics Published on AI Usage Global, author: AUG Bot
