The Energy Bill AI Never Shows You
Data centres used around 420 terawatt-hours of electricity in 2024, and by 2030 that figure is expected to more than double as demand for AI blooms in every sector. [1] Those figures describe what the machines draw from the grid. They say nothing about how much of that power produces useful work and how much vanishes into inefficiency that no dashboard ever reports. This is the quiet deception at the heart of the AI boom: the technology presents itself as sleek and immaterial, a few words typed into a box, while behind the interface sit warehouses of spinning hardware burning through electricity. The University of Michigan, the University of Illinois Urbana-Champaign and NAE Systems have launched a project called E-CALC to close that gap, backed by $1.83 million in DARPA funding and aimed at cutting AI energy consumption by up to 50 %, as reported by Open Source For U. [1] The ambition is not to build bigger computers but to make the existing ones work smarter — and the first step is admitting that nobody currently knows where the power actually goes.
The Ledger Nobody Keeps
E-CALC sets out to give developers something they have never had: a running account of how much energy an AI system uses at each stage of its life, from preparing data and training models to running them in everyday applications. The team plans to combine physical sensors with software tools to measure consumption more accurately than current methods allow, and to release those tools freely to other developers. The honesty problem here is structural rather than malicious. An AI model does not misreport its appetite; it simply has no mechanism for reporting it at all. The likely reason, in the project’s framing, is that developers optimise what they can see, and energy has remained largely invisible, so it has gone unoptimised. A tool that makes the invisible visible changes the incentives without changing anyone’s intentions.
The Gap Between the Demo and the Bill
Consider what a user experiences when an AI system answers a question or generates an image. The response arrives in seconds, feels weightless, and carries no signal about the computational effort behind it. That effort is real, and it is repeated across millions of interactions. Energy waste does not stay contained within the industry that generates it. It inflates cloud bills, strains local electric grids, and drains batteries faster than users expect. Each of these costs lands somewhere — on a company’s operating budget, on a utility’s capacity planning, on a device owner’s charging routine — yet none of them appears in the product itself. The deception is not a claim the system makes. It is a claim the system’s smooth surface implies: that nothing much is happening.
Building on What Was Already Measured
The two-year project, led by University of Michigan professor Mosharaf Chowdhury, extends earlier work on optimising AI training and on comparing the energy use of different models. [1] That foundation matters because it establishes that measurement is possible and that models differ meaningfully in what they consume. What E-CALC adds is scope: a lifecycle-wide view rather than a snapshot of one phase. Training a model is energy-intensive but finite; running it in production is continuous and, over time, may be the larger draw. A developer who measures only training sees half the picture and may optimise the wrong half. Extending the ledger across the whole lifecycle corrects a bias that has quietly shaped how the industry thinks about efficiency.
What Success Would Actually Look Like
If the project delivers, AI becomes cheaper to run and more energy efficient without losing the capabilities that make it useful. That framing is important, because efficiency gains in computing have a way of being spent rather than saved. A model that uses half the electricity could either reduce the total load or enable twice as much usage at the same cost. The E-CALC goal of up to 50 % is a target for cutting AI energy consumption, and whether it translates into lower aggregate demand depends on what happens next. The project’s contribution is to make that choice visible. You cannot decide whether to spend a saving until you can see it on the books.
The Professional Consequence
For the people who build and deploy AI systems, accurate energy accounting changes the job. Developers currently make trade-offs between speed, quality and cost, with cost measured mostly in money and latency. Once energy per stage is a known quantity, it could become a further axis in every design decision, and the tools being built are meant to be freely available so that the axis is available to everyone rather than only to firms with large engineering teams. The shift is not dramatic in any single decision. It accumulates across thousands of small choices about data pipelines, model sizes and deployment configurations, each of which may have previously carried an invisible environmental and financial surcharge.
The Deception That Ends With a Sensor
The most consequential form of AI deception may turn out to be the least dramatic: a system that appears to run on nothing. That appearance was never a lie told by anyone in particular. It emerged from the absence of instrumentation, the way a house with no electricity meter feels like it costs nothing to heat. Sensors and open software will not make AI virtuous, and they will not by themselves guarantee lower global electricity demand. What they do is remove the excuse. Once the energy cost of every stage is measured and published, the gap between what AI seems to be and what it actually does narrows to the width of a readout — and every subsequent decision to waste power becomes a decision someone made on purpose.
