Overview

Envision has moved the high-voltage DC debate from conference slides into a live data center. According to Sina Finance on Nov. 4, 2026, the company announced commercial operation of the world's first end-to-end 800 V high-voltage DC power system for a large AI data center at its Ulanqab Galaxy campus. The architecture takes 10 kV AC into solid-state transformers and outputs 800 V DC directly to racks. In doing so, it removes line-frequency transformers, AC UPS systems, and the AC-DC stage inside server power supplies. This is not a pilot in a laboratory; it is a deployment in a newly commissioned campus on the grasslands of Inner Mongolia.

Architecture and Efficiency

The technical core is the solid-state transformer. Envision replaces the iron core and copper windings of a conventional transformer with silicon-carbide devices and high-frequency magnetics. The 10 kV AC input is converted to 800 V DC without passing through a line-frequency transformer or an AC UPS. Envision says the SST stage maintains conversion efficiency above 98.5% and cuts power-equipment footprint by more than half. The gain comes from two sources: fewer conversion stages and magnetic components that no longer have to be designed around line frequency. The system still requires protection, control, and coordination, but it eliminates several lossy and space-consuming stages that have long been standard in data-center power chains.

The Three-Way Cost Trade-off

Data-center power design is a contest among three metrics: conversion efficiency, copper consumption, and floor space. Higher efficiency often demands more conversion stages or larger heat sinks; lower copper use demands higher voltage and thinner conductors, which raises device voltage ratings; smaller footprint compresses cooling and maintenance access. Most industry designs optimize two of the three. Envision's answer is to change the transformer itself. The economics become concrete in Ulanqab, where local computing projects pay 0.32 to 0.36 RMB per kWh and comprehensive operating costs are 30% to 50% lower than in eastern tier-one cities. For a million-kilowatt campus, each one-percentage-point efficiency gain saves electricity costs on the order of RMB 10 million a year. Copper savings are equally direct: Nvidia has estimated that a 1,000 kW rack using 54 V DC distribution would need 200 kg of copper in busbars alone; at 800 V, the same conductor cross-section can carry about 85% more power. Existing data centers cannot capture the full benefit because their line-frequency transformers and AC UPS systems cannot simply be removed. The first 800 V deployments therefore have to be new campuses.

Load Volatility and Grid Reliability

The grid can absorb large AI electricity consumption; the harder problem is fluctuation. Thousands of GPUs draw power together, bus voltage sags, and when training ends the load releases suddenly, like every air conditioner in a building switching on and off many times a day. When that load is also tied to wind and solar generation, both supply and demand are variable. The North American Electric Reliability Corporation issued its highest-level reliability alert in May, citing this class of computing load. It recorded a computing center disconnecting more than 1,000 MW unexpectedly, with a frequency impact comparable to a large generating unit tripping and a disconnection speed faster than dispatchers can respond. The Galaxy system was validated on these conditions: 10 kV to 800 V remained stable across operating modes, storage covered second-level gaps and minute-level backup, and the SST and storage coordinated at millisecond level to absorb GPU swings locally. Envision specifies 99.995% power reliability, equivalent to less than half an hour of downtime a year. The control strategy was previously exercised at a 2 GW off-grid system in Chifeng, where all power comes from wind and solar, all equipment is power-electronic, and no synchronous generator provides inertia. There, Envision used algorithms to replace part of the hardware, configured half the storage of comparable designs, and raised green-ammonia capacity by 15%. Galaxy differs because its load is nearly uninterruptible GPU training and it retains grid backup.

Policy and Green-Power Coordination

Policy has moved faster than the equipment market. Computing-power and electricity coordination was written into the Government Work Report for the first time this year and included in the 15th Five-Year Plan outline. In May, four departments including the National Development and Reform Commission and the National Energy Administration issued an action plan encouraging grid-forming storage at computing facilities and improving direct green-power connections. The implementation difficulty is coordination: generators own supply, equipment vendors own storage, and cloud providers own load, and each has a different ledger. Envision owns wind, solar, storage, and control systems and operates the Galaxy campus, so supply-side and control-side data enter one system. Even so, it can shift load only within customer authorization and delivery deadlines, for example by moving flexible training tasks to periods with abundant green power. Galaxy's planned direct green-power connection ratio exceeds 80%, and Ulanqab as a whole reaches two-thirds green power. That resource base is a precondition; the same architecture may not pencil out in locations with weaker wind and solar resources or higher tariffs.

Market and Supply-Chain Constraints

The industry has not settled on a voltage standard. Nvidia advocates moving directly to 800 V, but its companion rack power supplies are planned for mass production only in 2027. Google, Microsoft, and Meta are promoting ±400 V through the Open Compute Project, keeping the first step at the 400 V level. The disagreement sounds like a voltage argument, but on the campus balance sheet it is the same cost question. North America faces additional constraints: large transformers above 100 MW require 2.5 to 3.5 years to order, PJM projects average more than seven years from interconnection application to grid connection, and capacity prices have risen from $28.92 per MW-day in 2024/2025 to $329.17 per MW-day in 2026/2027. When operators cannot wait, ±400 V becomes attractive because it can leverage the electric-vehicle supply chain; 400 V-class devices have been mass-produced for a decade, and insulation and arc protection are easier. Interface standards will likely be set by bodies such as OCP, IEC, and UL. Chinese vendors can compete for equipment share, but that competition is intense and margins will be thin. In China, the solid-state-transformer field already includes TBEA, China XD, Sifang, Sungrow, and dozens of other suppliers. 2026 is being treated as the first year of industrialization, with most expecting volume growth in 2028.

Future Outlook

Envision's next target is 2030, when it plans to build 5 GW of green computing centers in desert and gobi regions worldwide. The demand backdrop is large: data-center electricity use is projected to rise from 170 billion kWh in 2025 to 800 billion kWh in 2030, an increase of more than 600 billion kWh. That volume cannot be met simply by installing more transformers. The Galaxy system is an important proof point because it demonstrates an end-to-end 800 V DC architecture in commercial operation, but it has not run for long. Whether it can be replicated in regions with different renewable profiles and tariff structures will depend on operating records. As Envision chairman Zhang Lei has put it, future AI competition is not only competition over models and chips but competition over power systems, and over how much effective intelligence each megawatt can produce. The first system solves the question of whether the architecture works; the hundredth will determine how much the solution is worth.