The rapid expansion of AI infrastructure is forcing data centre operators to revisit assumptions that have shaped power design for decades.
Much of the industry discussion has centred on GPUs, compute capacity and rising rack densities. Yet the implications for electrical infrastructure are just as significant. As individual racks move towards 100 kW and large AI campuses scale into hundreds of megawatts, relatively small inefficiencies in the power chain can translate into significant operating costs, additional cooling demand and greater pressure on supporting infrastructure.
In markets like China, the rapid expansion of AI and high-density computing infrastructure is pushing operators to examine not only how much power a data centre can secure, but how efficiently that power can be converted, distributed and protected. Similar questions are becoming more pressing across the Middle East as Saudi Arabia, the UAE and other regional markets expand sovereign cloud, hyperscale and AI infrastructure.
For operators planning facilities that may consume hundreds of megawatts, the efficiency of the power architecture can no longer be treated as a secondary engineering consideration.
Rethinking the efficiency trade-off
Traditional uninterruptible power supply systems have typically relied on online double-conversion architectures. Incoming AC power is converted into DC and then converted back into AC before reaching the IT load. The approach provides high power quality and isolates critical systems from fluctuations in the utility supply, which is why it has remained a standard choice for mission-critical environments.
The trade-off is that every conversion introduces electrical losses and generates heat.
These losses may appear relatively small at system level, but their financial impact changes considerably as facilities become larger. A one percentage-point difference in efficiency across a 100 MW environment represents a very different cost equation from the same difference in a conventional enterprise data centre. Additional heat also has to be removed, creating a secondary effect on cooling requirements and power usage effectiveness.
ECO operating modes emerged as one way of reducing these losses. When utility power meets the required quality thresholds, the UPS bypasses the rectifier and inverter stages and supplies electricity directly to the load.
The efficiency gain comes with engineering compromises that have limited the use of conventional ECO modes for highly sensitive environments. If utility power becomes unstable, switching back to inverter power can introduce an interruption measured in milliseconds. That may sound negligible, but high-value computing infrastructure can be sensitive to even extremely short disturbances.
Power quality presents another concern. AI data centres are not made up solely of servers and accelerators. Cooling systems, fans and other mechanical infrastructure introduce nonlinear loads that can generate reactive power and harmonic distortion. In a traditional bypass configuration, these harmonics may not be actively compensated.
Large facilities also tend to operate multiple UPS systems in parallel. Small differences in components, cable lengths and electrical impedance can produce uneven current distribution, leaving one UPS carrying more load than another. At sufficient scale, that imbalance becomes a reliability issue rather than simply an engineering inconvenience.
From conventional ECO to adaptive power
These constraints are leading power infrastructure vendors to rethink the traditional distinction between high-efficiency bypass operation and double-conversion reliability.
Huawei Digital Power’s Adaptive S-ECO architecture is one example emerging from this shift. Proven against the requirements of increasingly dense computing environments, the system focuses on three areas that have historically constrained ECO operation: transfer interruption, harmonic distortion and current imbalance between parallel UPS systems.
Its hardware clamping architecture keeps an inverter in hot-backup mode at a voltage slightly below the utility supply. If utility voltage falls below the required level, the inverter can take over without the startup delays associated with a conventional ECO design. Huawei reports a transfer time of 0 ms and system efficiency of up to 99.1 percent, with the system designed to meet Class 1 dynamic response requirements under IEC 62040-3.
The inverter also remains active in managing power quality. Rather than allowing harmonics generated by nonlinear loads to pass back through the system unchecked, it analyses load current and generates compensating current in the opposite phase. Huawei’s testing found that input current distortion exceeding 70 percent THDi under a specified nonlinear-load condition could be reduced to below 10 percent, and to below five percent at full load.
For parallel UPS installations, the system monitors current across individual bypass paths and adjusts the firing angle of the silicon-controlled rectifier to influence their effective impedance. Huawei reports that, under its test conditions, current imbalance could be held within five percent for resistive loads and within 10 percent for nonlinear loads even where cable lengths differed substantially.
What this means for Middle East operators
The engineering detail matters because efficiency at this scale has a direct relationship with operating economics.
Huawei estimates that a 100 MW data centre using Adaptive S-ECO could deliver significant annual electricity cost savings compared with traditional double-conversion operation. Actual savings will vary depending on electricity tariffs, utilisation, facility design and operating conditions, but the underlying economics become increasingly important as Middle East operators build larger, more power-intensive campuses.
The region’s data centre market is moving into a phase where capacity is being planned not in individual megawatts but, increasingly, in hundreds of megawatts across large sites and clusters. At those levels, operators have to account for the cumulative cost of every conversion loss across the power chain. Efficiency also influences how much of the available electrical capacity ultimately reaches revenue-generating IT equipment rather than being consumed by supporting systems.
Component life adds another dimension to the calculation. Traditional double-conversion architectures keep power components under continuous electrical and thermal load. Huawei says operating components in a lower-stress hot-standby state for much of their service life could potentially extend equipment life from approximately 10 years to 15 years.
That claim will depend on operating conditions and individual deployments, but it highlights a broader issue for data centre owners. Power architecture affects not only monthly electricity consumption, but maintenance cycles, asset replacement and the long-term economics of the facility.
As AI pushes compute density higher, data centre design will increasingly be judged by how effectively the entire infrastructure supports that density. The industry has already invested heavily in improving server utilisation, cooling technology and facility design. Power conversion represents another area where incremental efficiency gains can become substantial once they are multiplied across hundreds of megawatts.
For Middle East operators building the next generation of AI infrastructure, experience from high-density markets, including China, provides a useful reference. The objective is no longer simply to maximise available power capacity. It is to ensure that more of that capacity reaches the computing infrastructure reliably, with as little lost along the way as possible.
Discussion about this post