Software, not concrete, can fix the AI energy crisis

Jan de Vries · 2026-07-02

The AI energy crisis is not what it seems like. You may not know, Stanford research found that advanced economy power grids run at an average utilisation of just 30%. While research from Duke University shows electricity providers can already meet data centre energy needs on 350 out of 365 days a year. The hardware exists. The electrons exist. What is missing is the intelligence to coordinate them -- and that is a software problem, not a construction problem. The dominant response has still been to build more. This means more generation, substations, and transmission capacity. Multinational giants such as Amazon, Google, Meta, and Microsoft are signing power purchase agreements at a scale the energy industry has never seen. Data centre electricity consumption is on track to more than double by 2030. But, grid interconnection queues in Europe's main data centre markets already run to seven to ten years. In Dublin, where many major multinationals operate, the strain on the power grid has led the operator to pause new connections until 2028. For companies unable to secure billion-dollar power deals, waiting for new infrastructure is not a viable strategy. The capacity gap will not be closed by construction alone. It will be closed by software that makes existing infrastructure perform closer to its potential -- and right now, almost nobody is funding that. After mapping more than 100 companies across the data centre efficiency landscape, the conclusion is clear: software is the most capital-efficient solution to the AI power crisis, and it is the most underfunded layer in the stack. The inference cost of running AI models declined 280 times between 2022 and 2024 -- almost entirely through software improvements. Google doubled its data centre energy efficiency through software optimisation alone. The capital flowing into physical AI infrastructure dwarfs what is going into the software that makes that infrastructure perform. ### Four layers, one opportunity There are four distinct layers where software can recover capacity. Grid efficiency software manages interconnection queues, data centre flexibility, and distributed energy orchestration. This market is estimated to reach $4.12 billion by 2030. Much grid congestion is a coordination failure, not purely a physical one: outdated systems cannot route available power efficiently, and software can close much of that gap without new infrastructure. Facility efficiency covers the software and controls that reduce data centre cooling and infrastructure energy use, lowering operating costs and increasing usable capacity. Public market estimates for adjacent facility-management software suggest a market of roughly $4 billion to $5 billion by 2030-2031. Crucially, every improvement in rack density creates additional sellable AI capacity without a new site, making facility efficiency a capacity expansion story, as well as a cost reduction one. Compute efficiency focuses on recovering stranded GPU capacity. Software that identifies unused capacity and orchestrates workloads more intelligently can lift GPU utilisation materially: research has shown one production GPU-sharing system deployed across 20,000+ GPUs improved utilisation from 26% to 76%, while, according to Fujitsu, over 75% of organisations still report GPU utilisation below 70% even at peak load. An analysis found the broader AI infrastructure market is projected to reach $223.5 billion to $394.5 billion by 2030. Software efficiency -- model compression, inference optimisation, intelligent model routing -- is one of the largest and least funded layers, with the AI workload-management market implying a projected $197.5 billion opportunity by 2030, while the broader AI inference market is forecast to reach around $253.75 billion by 2030. Yet, the majority of AI infrastructure capital still goes into concrete and copper. We're building power plants and substations when the real bottleneck is a lack of intelligent software. Think about it: your phone's processor is millions of times more powerful than the computers that sent humans to the moon, yet our power grids run on systems designed decades ago. The AI energy crisis isn't a shortage of electrons -- it's a shortage of intelligence to use them wisely. The companies that figure this out first won't just save money on electricity bills. They'll unlock capacity that everyone else assumes doesn't exist. And in a world where data centre wait times stretch to a decade, that's a competitive advantage worth billions.