AI Data Center Electricity Demand Hits 945TWh Projections by 2030
By 2030, global artificial intelligence infrastructure is expected to consume 945TWh of electricity annually, matching Japan’s current total power consumption, Tom’s Hardware reported. This looming energy crunch has pushed hardware manufacturers to design more efficient graphics processing units, but researchers increasingly point to software optimization as a critical untapped lever for power conservation. While data center power usage effectiveness has remained stagnant for six consecutive years according to the Uptime Institute’s 2025 survey, computer science experts argue that modifying algorithms and systems software can drastically reduce power draws without replacing physical silicon.
Software Layers Offer Compter Efficiency Gains
Computing systems operate as a stack consisting of hardware at the bottom, backed by systems software, algorithms, and applications. Jae-Won Chung, a computer science and engineering PhD candidate at the University of Michigan and a researcher with the ML.Energy initiative, stated in an interview with Tom’s Hardware that software and algorithms can meaningfully help manage power and energy. While replacing slow and costly physical chips remains difficult, the upper three layers of the tech stack are much easier to modify, allowing efficiency gains to compound across workloads.
Tests conducted by ML.Energy on the Alibaba Qwen 3 235B A22B Thinking model demonstrated that running inference in lower-precision FP8 formats used one-third less energy on problem-solving tasks compared to bfloat16 versions. Additionally, Chung’s Perseus training optimizer identifies underutilized portions of large-model training jobs and slows them down to finish simultaneously with busier threads, cutting training energy consumption by up to 30% without sacrificing throughput.
Hardware Giants Adopt Advanced Power Profiling
Major semiconductor firms are already implementing similar software-driven energy controls. Nvidia’s Blackwell architecture features power profiles that fine-tune GPU compute frequencies, memory speeds, power limits, NVLink states, and cache settings to match specific workloads. According to company estimates, these dynamic adjustments save up to 15% of energy while maintaining 97% or more of overall performance, enabling power-constrained facilities to run additional GPUs and boost overall throughput by as much as 13%.
At the infrastructure level, legacy equipment running outdated code remains a primary source of unnecessary power consumption. Chetan Visrolia, business development manager for data center infrastructure at SHI, noted in an interview with Tom’s Hardware that legacy hardware supporting old software represents the lowest-hanging fruit for quick savings in any facility.
Global Workload Shifting Responds to Grid Signals
Beyond optimizing individual servers, software can dictate when and where data centers draw electricity from local power grids. Sophie Hall, a doctoral student at ETH Zurich’s Automatic Control Laboratory who has studied workload allocation across Google’s global data center fleet, explained to Tom’s Hardware that timing and geographic placement present major opportunities for efficiency. Non-time-sensitive batch jobs can be delayed until local grid demand drops or dynamically routed to regions with spare capacity and higher proportions of renewable energy.
Implementing dynamic scheduling successfully counters traditional over-provisioning strategies used by hyperscalers. However, practical barriers remain. National data sovereignty regulations frequently prevent workloads from crossing international borders, and inference providers often hesitate to transfer massive datasets across external networks due to potential operational risks.
Frequently Asked Questions
What is Power Usage Effectiveness and why is it limited?
Power Usage Effectiveness measures how efficiently a data center uses energy by comparing total facility power to the power delivered to IT equipment. According to the Uptime Institute’s 2025 survey, PUE fails to track whether the software running on servers is actually performing useful work.
How much energy can FP8 inference formats save?
ML.Energy testing on the Alibaba Qwen 3 235B A22B Thinking model showed that executing inference in the FP8 lower-precision format used one-third less energy on problem-solving tasks than bfloat16 configurations.
What limits geographic workload shifting for data centers?
Sophie Hall noted that strict data sovereignty regulations restrict workloads from crossing national borders, while providers worry about the logistical risks of moving massive datasets across external networks.