AI computing is parting ways with conventional data-centre practice on power. Jiemian News reported on 4 November that a traditional data centre plus grid reinforcement typically takes 18 to 24 months or more, while AI projects demand capacity expansion within months. That mismatch between construction speed and power availability has become the central bottleneck for compute expansion. A second signal came from Caixin on 5 November: by the end of September, 24 provinces had issued supporting policies for direct green-power connection, and 12 green-power direct-connection projects for computing facilities had been approved. In 2026, “computing-power coordination” was written into the Government Work Report for the first time.
Three hard constraints on distribution equipment
Delivery tempo. If grid reinforcement cannot wait, distribution must be factory-prefabricated and site-assembled. Busbars, terminals, protection settings and factory tests are completed indoors; on site only positioning, wiring and commissioning remain. Compressing site hours is becoming the deciding factor in tenders.
Transfer time. AI loads are highly sensitive to voltage sags. STS/ATS transfer times and the coordination logic between UPS and generators should be derived backwards from server power-supply tolerance curves, not set to a rule-of-thumb “10 milliseconds”.
Power quality with green supply. Direct green-power connection means a more variable source. Voltage deviation, harmonics and phase imbalance all propagate to the rack side, so measurement points, protection settings and power-quality sampling inside the cabinet must be recalculated.
The real threshold in data-centre distribution has never been making cabinets bigger; it is making uncertainty smaller.