China Claims First Commercial 800V DC Power System for AI Data Center
Envision says an 800-volt direct-current power architecture is running at an AI campus in Inner Mongolia, though independent performance data is still pending.
According to TechRadar, China's Envision has announced that it has put an integrated 800-volt direct-current power architecture into commercial operation at an AI facility in Inner Mongolia. The installation is at the Galaxy AI campus in Ulanqab, where trials of the company's Dubhe controller and supporting energy model began in August 2026.
The reporting describes the system as converting 10-kilovolt grid electricity directly into an 800-volt DC supply. It also pairs solid-state transformers with battery storage, a combination the outlet presents as a new approach for AI facilities.
Envision describes the deployment as the first commercial use of this configuration. TechRadar's summary notes that independent performance figures have not yet been published, so the efficiency and conversion claims rest on the company's own account for now.
The story's headline frames the work as a "world first," a label that comes from the company's framing rather than from an independent verification in the coverage available. Readers should treat that phrasing as a claim.
Background
Data centers have become one of the fastest-growing sources of electricity demand, driven largely by the chips used to train and run large AI models. Those chips draw far more power per rack than the servers of a decade ago, and the electrical systems that feed them have struggled to keep pace.
Most conventional data centers receive alternating current from the grid, convert it through transformers, and then convert it again in uninterruptible power supplies and server power units before it reaches the processors. Each conversion step loses some energy, and the losses turn into heat that cooling systems must then remove, adding another layer of consumption.
Direct-current distribution has been discussed for years as a way to cut some of these conversion steps. Supplying DC closer to the load can reduce the number of stages between the grid and the chip. Higher voltages, such as 800 volts, can also carry the same power with lower current, which allows thinner cabling and less resistive loss over a given distance. Engineers have pointed to these benefits for high-density AI racks, where power demands per cabinet can be very large.
Solid-state transformers are an alternative to traditional bulky iron-and-copper transformers. They use power electronics to change voltage and can offer more control over power flow, though they are still less common at scale and must meet strict reliability requirements. Battery storage on site can smooth demand spikes and provide backup. None of these ideas is unique to one company; the open question in any deployment is how well the full system performs under real and sustained workloads.
Why it matters
If the architecture performs as described, it points to a practical route for cutting energy waste inside AI facilities. Power losses in data centers translate directly into higher operating costs and larger cooling loads, so even modest efficiency gains can matter at the scale of a campus filled with accelerator hardware.
The broader stakes are about grid strain. Operators are under pressure to connect new capacity quickly, and anything that reduces the power needed per unit of computing could ease some of that demand. It could also shape how future facilities are designed, since power architecture is difficult to retrofit once a building is constructed.
For the industry, a working commercial example from a large Chinese operator adds to a growing field of competing approaches to AI power distribution. Companies building chips, racks and power equipment are all watching which designs prove reliable and cost-effective. Standards for high-voltage DC in data centers are still developing, so early deployments can influence which approaches become common.
There is also a supply-chain dimension. Power electronics, solid-state components and battery systems involve specialized manufacturing, and a design that depends on particular equipment may affect procurement decisions for other operators.
What to watch
The most important signal will be independent performance data. Efficiency, uptime and conversion figures measured by third parties would show whether the claimed gains hold across real workloads and seasons, not only in controlled conditions.
Watch whether Envision publishes technical details on the Dubhe controller and the energy model, and whether other operators adopt similar 800-volt designs. Adoption beyond a single campus would suggest the approach is more than a pilot.
Also watch how the system handles reliability and maintenance over time. Power failures in AI facilities are costly, and a new architecture must prove it can match the uptime standards operators already expect.
Finally, note whether regulators or standards bodies begin addressing high-voltage DC in data centers, since clear safety and interoperability rules would shape how widely the design can spread.
Viewpoints
Commentary from our columns: three angles on this story.
The Skeptic open questions and weak points
Every press release about a "world first" deserves a few questions before anyone gets excited. The reporting says the system is in commercial operation, but the efficiency and conversion claims come from the company, and independent performance figures have not been published. Those are the numbers that matter. A power architecture is judged over months of heat, load swings and failures, not during a launch window. How long has the system been running under full AI load? What happens during a grid disturbance or a battery fault? Does the efficiency gain hold once cooling, maintenance and redundancy are counted? A bold label is not a measurement. Until outside engineers can check the results, the most honest position is interest paired with patience. The industry has seen many promising power ideas that looked clean in a demo and became expensive in daily use. This one may prove out. It has simply not yet earned the word "proven."
The Builder the engineer's view
From an engineering standpoint, the appeal of an 800-volt DC bus is easy to understand. Fewer conversion stages mean fewer points where energy turns into waste heat, and higher voltage lets you move the same power with less current and thinner conductors. Pairing solid-state transformers with on-site batteries is also a sensible combination for dense AI racks with spiky loads. The hard part is everything around the headline. Protection and fault isolation on DC are harder than on AC, because there is no natural zero crossing to help interrupt arcs. Maintenance crews need trained procedures, and component standardization matters for keeping spares available. If I were evaluating this design, I would want the controller logic, failure modes and commissioning data before anything else. Clever architecture is only as good as its behavior at 3 a.m. when a module trips.
The Rulebook privacy, safety and regulation
Changes to core power infrastructure raise questions that go beyond efficiency. High-voltage DC systems in populated or shared facilities bring electrical safety obligations, including arc-fault protection, clear labeling and trained personnel. Battery storage adds its own fire-prevention and thermal-management rules, and local fire codes may not yet address this combination at scale. Grid connection is another area for scrutiny. Large AI campuses draw substantial power, and how they interact with regional grid planning and any emergency demand-response requirements affects the public. The reporting does not describe the regulatory approvals behind this deployment, and that gap is worth noting. Clear standards for DC distribution in data centers would help operators, insurers and communities judge risk consistently. Innovation in this space can be welcome while still asking that safety documentation and oversight keep pace with the engineering.