ÌÇÐÄÆƽâ°æ

Meta reported an average power usage effectiveness (PUE) of **** for its global data center portfolio in 2023, the latest year for which figures were available. This was slightly lower than Google's average PUE of ****, which covers 2024. PUE is a metric used to quantify the efficiency of a data center facility. It is calculated by dividing the total power supplied to a facility by the power used to run IT equipment within the facility, with a lower figure therefore indicating greater efficiency.

Data center efficiency under the spotlight

The AI revolution has triggered an unprecedented surge in data center demand, with modern AI-optimized facilities drawing massive amounts of power. This stems from the rapidly increasing complexity of AI models, which necessitate cutting-edge GPUs such as those based on ±·³Õ±õ¶Ù±õ´¡â€™s Blackwell architecture. These processors are extremely power-hungry and generate significant heat, driving the need for advanced water cooling systems. This has thrust AI into the heart of the broader sustainability conversation, forcing policymakers, technologists, and the public alike to grapple with the environmental and energy implications of training ever-larger models.

Does PUE show the whole picture?

In the face of growing concern, major cloud and AI operators are seeking to demonstrate progress in data center efficiency. PUE has long been used as a key efficiency benchmark, based on the assumption that directing a greater share of power to computing rather than overheads represents improved efficiency. However, PUE can be misleading, as it only reflects the ratio of total facility power to IT equipment power, not absolute energy use. A lower PUE can simply reflect that a data center is running dense, power-hungry AI workloads, rather than indicating that operations are truly more efficient.