China AI campus tests 800V DC power for future Nvidia racks

Testing the 800V DC architecture

In Ulanqab, Inner Mongolia, Envision has opened a 5MW hall at its Galaxy AI campus that operates using 800V DC power sent directly to the server racks. This facility acts as a functional trial for a bigger version intended for the campus's upcoming phase. The project is part of a 120MW initial build that was completed in less than a year.

The architecture is significant because it aligns with Nvidia's stated roadmap for future data center infrastructure. Nvidia has announced that it will lead the transition to 800V DC to support 1 MW IT racks and beyond starting in 2027. This move is intended to address the physical limits of current 54V DC in-rack distribution as racks exceed 200 kilowatts.

By using 800V DC, the system can transmit 85% more power through the same conductor size compared to older standards. This efficiency allows for a 45% reduction in copper requirements. The Galaxy pilot represents a major leap forward, as its 800V capacity is 3.3 times higher than the 240V DC pilots China Telecom tested in 2007.

How solid-state transformers work

The technical core of this setup involves solid-state transformers, or SSTs. These devices take 10kV AC power from the grid and transform it immediately into 800V DC for the server racks. This direct conversion is a major departure from traditional power distribution methods used in most current facilities.

By using SSTs, the campus removes several intermediate stages that are standard in older designs. Specifically, it eliminates the need for low-voltage transformers, UPS (uninterruptible power supply) systems, and rectifier stages between the grid and the rack. This streamlined path reduces the complexity and energy loss associated with multiple conversions.

Nvidia is currently studying both traditional transformer-based and solid-state transformer approaches for its own 800V DC plans. However, Envision has already implemented the SST version in a live hyperscaler site. This gives the Galaxy campus a head start on deploying the specific technology that many believe is the natural next step for high-density AI power.

Cooling and construction efficiency

The Galaxy campus also utilizes a specific cooling design that avoids a central chiller plant. Instead, liquid- and air-cooled servers share a single water loop. This design has been running at scale in northern China for several years and allows for high efficiency without the need for a massive central cooling unit.

This approach contrasts with many US data centers that maintain central chillers to provide flexibility between cooling methods. Nvidia has noted that its Vera Rubin chips can run on 45°C water without needing chillers, but most current US sites still keep them for air-cooled equipment. Galaxy’s design maintains much of that flexibility while dropping the central chiller requirement.

Construction speed is another factor in the campus's design. Envision co-designed a prefabricated substation that can be delivered in less than 3 months. This is significantly faster than the approximately 6 months required for a conventional indoor substation. Such speed is critical for developers trying to keep pace with the rapid demand for AI compute.

The shift toward integrated power

The competition in the AI space is moving beyond just the chips to the entire power chain. Because the structural bottleneck of AI is often the availability of electrons, the architecture from the grid connection to the rack is being redesigned. This shift provides integrated players a new way to compete against traditional data center developers.

Envision, a renewable systems integrator, designed and integrated the low-voltage power modules and the 800V SST system for the Galaxy campus. Its affiliate, AESC, supplied the battery storage. This type of vertical integration allows for a more cohesive system than when different vendors cover only isolated parts of the power chain.

The move toward 800V DC is a path that China has been exploring for nearly two decades. It began with 240V DC pilots in 2007 and moved to Alibaba’s Panama power system in 2018. The current 800V SST pilot at Galaxy represents the latest evolution in this long-term strategy to optimize how power reaches the AI hardware.

What comes next for the grid

While the 5MW pilot is currently a small portion of the Galaxy campus, it serves as a reference point for the US roadmap. Many questions remain open regarding the efficiency and uptime of these high-voltage systems, as those specific figures are currently held behind paywalls. However, the move toward 800V DC is clearly the direction for high-density AI factories.

Nvidia has identified several key partners in its 800V DC electrical ecosystem, including Delta, Eaton, Schneider Electric, and Vertiv. These companies will likely play a major role in the rollout of the Kyber rack-scale systems scheduled for 2027. The transition will require a massive overhaul of how data centers are wired and cooled.

For data center builders, the primary takeaway is the move toward prefabricated components and high-voltage DC. The ability to deliver power more efficiently with fewer stages will be a requirement for supporting 1 MW racks. As the industry moves toward 2027, the focus will shift from just buying more chips to building the infrastructure capable of powering them.

The bottom line

The Galaxy AI campus is already running a 5MW hall using 800V DC power and solid-state transformers to deliver electricity directly to the racks. This setup removes traditional intermediate stages like UPS and low-voltage transformers, creating a more direct path from the grid to the AI hardware.

Nvidia plans to adopt this 800V DC architecture for its Kyber rack-scale systems in 2027 to support 1 MW racks and beyond. By switching to this higher voltage, data centers can transmit 85% more power through the same conductor size while reducing copper requirements by 45%.

This development highlights a shift where the power chain is becoming as important as the chips themselves. As AI demands more electricity, the ability to deliver power efficiently and quickly will define the next generation of data center construction.