Dev.to AI 🤖 Ai 👁 0 📖 5 min read

The 800-volt datacenter: what one Chinese campus reveals about the next power architecture

If you work on AI infrastructure in the US or Europe, you have probably spent the last year hearing about the 800 VDC transition. It shows up in conference talks, in OCP blueprints, in vendor roadmaps and in long analyse

If you work on AI infrastructure in the US or Europe, you have probably spent the last year hearing about the 800 VDC transition. It shows up in conference talks, in OCP blueprints, in vendor roadmaps and in long analyses about how to feed a 100 kW rack without turning the power room into a furnace. The framing is almost always the same: this is a new architecture, it is coming, and it will be hard.

A report published in early November by SemiAnalysis, based on a site visit to Inner Mongolia, suggests a less comfortable framing. In one Chinese region, datacenters have been running on direct current for almost two decades. The 800-volt rack is not a future roadmap there. It is a pilot hall you can walk into, and the rest of the electrical chain has already been simplified around it.

That difference matters for anyone who has to sign off on a power design. It tells you which parts of the "800 VDC revolution" are genuinely unsolved and which parts are mostly organizational inertia.

What the visit actually found

SemiAnalysis toured a campus called Galaxy in Ulanqab, Inner Mongolia, owned by Envision Group — a company better known as a renewable systems integrator with more than 100 GW of wind turbines installed worldwide. The campus is planned for up to 800 MW of IT capacity, roughly the same scale as the core of Stargate in Abilene, Texas. Its first 120 MW phase was delivered within a year, at about one fifth of US construction cost, according to the report. All four phases are built to suit for two internet companies.

Around the site, the scale is hard to picture from a Western perspective: roughly 30,000 construction workers across 22 datacenter projects in the surrounding high-tech zone, per a local report cited by SemiAnalysis. Envision's own datacenter plans in Inner Mongolia add up to 3 GW, and the region has announced 25 GW of datacenter projects in total. The comparison that matters is not chip for chip. It is how fast a hall can be built and how little of the electrical chain sits between the grid and the rack.

The electrical chain is where the money is

Inside one 5 MW hall, solid-state transformers take 10 kV AC from the grid and convert it directly to 800 V DC for the racks. That single step removes the low-voltage transformers, the UPS and the rectifier stages that a conventional AC datacenter needs. In a world where rack densities keep climbing and every conversion stage costs space, copper, cooling and efficiency, deleting three stages is a bigger engineering win than a modest gain in chip efficiency.

The report puts the broader picture in useful numbers. Roughly a quarter of Chinese datacenters built over the past decade run on DC power, tracing a path from China Telecom's 240 V DC pilots in 2007 to Alibaba's Panama power system in 2018. In the United States, about 97% of datacenters still use AC UPS. The 800 VDC design is, in this light, less an invention than a convergence of a path China has been walking for years and a path the US is only now starting to consider seriously.

Why does China have that path at all? Because the policy push is to bring new demand to where renewable energy already exists. Inner Mongolia has more wind than its local industry and population can consume. Envision built a 2 GW off-grid wind, solar and storage system in Chifeng to feed a green ammonia plant, running entirely on power electronics with no synchronous generator — and a small pilot datacenter built with Tencent's modular T-Block units runs on the same system. Galaxy is the larger application of the same idea.

There is an uncomfortable business lesson here too. These new entrants are not in the datacenter business for thin colocation margins. Envision builds its own wind farms and a dedicated power line, integrates the low-voltage modules, manufactures the battery storage that smooths the load, and co-designs the prefabricated substation. The datacenter creates demand for equipment the group already sells. Project returns are squeezed, so the reward comes from somewhere else in the chain.

Cooling: no chiller, one water loop

The thermal design is just as instructive. Galaxy uses a layout common in northern China: liquid-cooled and air-cooled servers share a single water loop, and there is no central chiller plant. That is a design that has run at scale in the region for years.

Nvidia says Vera Rubin can operate on 45 °C water without chillers. Most Western datacenters nevertheless keep a chiller for air-cooled equipment and for the flexibility to shift workloads between cooling modes. Galaxy drops the central plant and keeps much of that flexibility by sharing the loop. Whether you would copy this depends on your climate, your risk tolerance and your service-level agreements — but it is a real production reference point, not a slide.

The practical takeaway is that "liquid cooling" is not one decision. It is a set of decisions about supply temperature, loop topology, redundancy and how much of the hall you are willing to run hot. The Chinese example shows that a simpler topology can work when the whole hall is designed around it from day one. Retrofitting the same idea into a building designed for AC and chillers is a different problem entirely.

What to take from this if you are not building a hyperscale campus

First, treat the power chain as a single system. If you are planning a high-density deployment, the question is not only which accelerator to buy but how many conversion stages sit between the utility feed and the silicon, and who owns each one. A design that removes a UPS and a rectifier stage changes your capex, your floor plan and your efficiency in one move.

Second, look at who is building around you. Industrial groups with an edge in energy, construction or electrical equipment are entering AI infrastructure because the datacenter is a new customer for things they already make. That changes the competitive picture: the cheapest rack may come from a company whose real business is somewhere else in the stack.

Third, be careful about which "new" architecture is actually new. The 800-volt rack and the chiller-free hall are being presented in the West as frontier engineering. In Inner Mongolia they are the continuation of a two-decade path. If you are writing a request for proposals, ask vendors to separate the genuinely novel parts from the parts that are simply being rediscovered.

Finally, keep the geography in view. Inner Mongolia's 25 GW pipeline is an announcement figure, not delivered capacity, and the report is explicit that constraints — power, land and water — still bind. The lesson is not that China has solved the problem. It is that a different starting point produces a different answer to the same physics question.

The most useful thing about a site visit is that it removes abstractions. "800 VDC" stops being a roadmap bullet and becomes a transformer in a 5 MW hall, a shared water loop and a construction zone with 30,000 workers outside. If your own power architecture is still on a slide, that is a useful gap to notice.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.