Software could be the easiest fix for hyperscalers’ AI power squeeze, researchers say — data center demand is…

The urgency of this question becomes clear when examining the stubborn stagnation of traditional efficiency metrics within the data center ecosystem. Average power usage effectiveness, commonly known as PUE, has barely changed for six successive years, according to the Uptime Institute’s 2025 survey. While PUE remains a vital metric for measuring how efficiently a facility delivers energy to its computing equipment—effectively showing whether cooling and power distribution systems are wasteful—it fundamentally fails to account for whether the software running on the servers is actually doing useful work. PUE tells facility managers how much overhead energy is burned to keep the lights on and the chips cool, but it remains blind to the computational inefficiency happening inside the machines themselves.

This oversight is especially glaring when looking at where electricity is actually consumed inside a modern facility. Servers account for around 60% of electricity demand in a modern data center, making them by far the single largest consumer of power in the building. Meanwhile, cooling systems—long the primary target of engineering optimization and green initiatives—range from consuming about 7% of total energy in an efficient hyperscale site to more than 30% in a less-efficient, older enterprise facility. Given that the vast majority of power is drawn directly by the hardware executing code, focusing solely on ambient room temperatures and advanced liquid cooling loops leaves the software side of the equation largely ignored. Eking out more efficiency from the software and workloads running on those servers suddenly seems not just sensible, but entirely necessary to avert a severe environmental and economic bottleneck.

The relentless rise in demand driven by large language models, generative AI applications, and enterprise data processing has pushed infrastructure to its absolute limits. Traditionally, the tech sector approached scaling through brute force: if a model was too large or processing took too long, the solution was simply to deploy more servers, buy faster accelerators, and pipe in more electricity. Tech giants have recently signed multi-billion-dollar deals to restart decommissioned nuclear power plants, contract directly with renewable energy developers, and construct sprawling campuses near major metropolitan sub-stations. Yet, the lead times for these energy infrastructure projects often span years, while the deployment of new AI clusters happens in a matter of months. This widening temporal gap between energy supply and computational demand is precisely why the physical-only strategy is facing renewed skepticism from energy analysts and software engineers alike.

When evaluating the operational realities of modern data centers, the disconnect between hardware improvements and software bloat becomes starkly evident. While semiconductor manufacturers achieve remarkable generational leaps in performance-per-watt efficiency, the software ecosystems built on top of those chips frequently expand to consume every available cycle. Unoptimized code, redundant data processing, inefficient model architectures, and continuous background operations mean that servers are often burning massive amounts of electricity to execute tasks that could be streamlined or eliminated entirely. Yet, corporate sustainability reports and industry conference keynotes continue to heavily favor hardware innovations—such as immersion cooling, advanced silicon design, and microgrid integration—over deep algorithmic efficiency and workload rationalization.

The findings from the Uptime Institute’s 2025 survey highlight the limitations of relying exclusively on facility-level metrics to drive genuine sustainability. For over a decade, the industry rallied around PUE as the holy grail of data center efficiency, driving down ratios from historical highs of 2.0 or worse down toward the ideal theoretical limit of 1.0. However, as the survey demonstrates, progress on PUE has plateaued across the board. Most modern hyperscale operators have already squeezed out nearly all the easy gains in facility cooling and power distribution architecture. With physical facility efficiency reaching a plateau, any further significant reductions in carbon footprint and energy consumption must come from the IT layer—specifically from how servers are utilized and how software code is written and executed.

This shift in perspective forces a hard look at the nature of workloads entering modern data centers. AI training and inference tasks demand unprecedented levels of parallel processing and computational throughput, but they also introduce vast amounts of redundant calculation and over-provisioned computing. As companies race to deploy artificial intelligence features across every consumer and enterprise product, the sheer volume of queries, data ingestion pipelines, and continuous model re-training creates an immense baseline load. If the industry continues to prioritize hardware expansion without simultaneously addressing software efficiency, the projected 945TWh consumption figure for 2030 could easily be surpassed, placing an intolerable strain on global energy grids and climate goals alike.

Addressing the software side of the data center energy equation, however, presents its own unique set of cultural and economic challenges. Software development historically prioritizes speed to market, developer productivity, and feature richness over energy minimalism. Writing lean code that minimizes CPU cycles and memory overhead often requires specialized expertise and extra development time—resources that are frequently deprioritized in fast-moving technology markets. Furthermore, cloud computing business models, which charge customers based on resource consumption or instance hours, have historically offered little direct financial incentive for software developers to optimize their applications for power reduction. Unless energy costs are baked directly into software engineering metrics or regulatory pressures force a reevaluation of computational waste, the path of least resistance will remain anchored in hardware upgrades.

As the International Energy Agency’s projections draw closer, the conversation around data center sustainability is clearly at a critical crossroads. The heavy reliance on securing more physical power and building hyper-efficient silicon will remain essential components of the strategy, but they can no longer be treated as the sole solution to an escalating crisis. With servers dominating the vast majority of facility power demand and traditional cooling metrics hitting a wall of diminishing returns, the industry must reckon with the efficiency of the code itself. Whether through smarter workload scheduling, algorithmic optimization, or a fundamental cultural shift in how software is engineered for the AI era, mitigating the impending energy crunch will ultimately require looking past the hardware rack and into the logic running within.