One Station, One Step: How AI Data Centers Are Rethinking Power Delivery

One Device Is Holding Back the Entire AI Industry

Training large AI models requires compute power, and compute power requires electricity. That much is not a secret.

What few people realize, however, is that the real bottleneck for AI is often not GPUs—it is the equipment that delivers electricity from the grid to the servers.

A traditional data center substation takes 12 to 18 months from design to energization. AI data centers are built in 6 to 8 months. By the time you finish building the substation, the GPUs have already gone through two generations. Worse still, traditional substations follow a “build first, energize later” model—every expansion means starting from scratch. AI data center capacity expansion happens in phases—100 racks today, another 200 next month—and traditional substations simply cannot keep up with that rhythm.

Putting an Entire Substation into a Prefabricated Enclosure

The logic behind the “Power Island” is simple: stop building substations on site; instead, make them as prefabricated enclosures, manufactured in the factory and assembled on site like building blocks.

All 167 functional modules are 100% factoryprefabricated and fully systemtested before shipment. Onsite civil work is reduced by 70%, installation work by 80%, and the entire substation delivery cycle is compressed from the traditional 12–18 months to 150 days.

But “fast” is only part of the story. What truly changes the game is the power architecture itself.

The traditional data center power delivery chain is: highvoltage grid → multiple stepdown stages → AC distribution → rectification to DC → servers. Every stage introduces losses, requires equipment, and consumes space.

The “Power Island” chain is: 110 kV or 220 kV direct HV input → SST solidstate transformer singlestage conversion → 800 V DC direct supply to the server hall.

From AC to DC, from high voltage to low voltage—in one step. All the intermediate stepdown transformers, distribution cabinets, and rectifiers are eliminated.

Two Key Numbers: 98.5% and 30%

The core enabler of this chain is the SST (SolidState Transformer) . It uses silicon carbide highfrequency topology to replace traditional copperandiron cores, reducing footprint by 60% and achieving power conversion efficiency of 98.5%. Conventional transformers run at 96–97% efficiency. Do not underestimate that 1.5percentagepoint difference—an AI data center consumes hundreds of millions of kWh per year; 1.5% represents millions of kWh.

More importantly, the “Power Island” does not just “deliver power”—it enables “computepower synergy.” It is equipped with an AI platform that monitors wind and solar generation, grid load, and battery storage in real time, then allocates power based on the priority of compute tasks. When green energy is abundant, it schedules “flexible tasks” like data cleaning and model pretraining; when the grid is under stress, it ensures that core training and inference workloads are never interrupted.

The results: Token energy cost reduced by 30%, total substation cost reduced by 20%, and onsite O&M labour reduced by 40%.

Why This Matters

The 2026 Government Work Report explicitly included “computepower synergy” as a national strategy for the first time, calling for the implementation of “ultralargescale intelligent computing clusters, computepower synergy, and other new infrastructure projects.” The NDRC has further clarified that direct investment in the computing power network will reach the trillionyuan level, with computepower synergy becoming a new growth pole for investment.

The “Power Island” essentially answers one question with a fully domestic power equipment solution: what kind of power does an AI data center actually need? Not just “any power will do.” It needs highvoltage direct input, DC direct supply, green energy integration, and intelligent dispatch. Traditional substations cannot meet any of these four requirements.

SST solidstate transformers entered intensive commercialization in 2026. They are no longer just lab concepts—they are physical assets that are transforming the power delivery logic of AI data centers.

A “Power Island” does not just save one substation—it saves the time and electricity wasted in AI infrastructure construction. In the second half of the compute race, the competition is not just about chips. It is also about power.