One Grad Student's Software Is Making AI's Energy Habit Measurable

Energy used to be an afterthought in AI development. When the generative AI race kicked off in the early 2020s, developers cared about one thing: getting bigger models out the door faster than the competition. Nobody was tracking watts.

That's changed fast. In parts of the US, AI-driven electricity demand is already outstripping grid capacity, pushing up prices and straining reliability. Data centers have gone from infrastructure footnotes to hot-button political issues. Efficiency isn't optional anymore. It's a constraint developers have to design around, whether they're training a new model or just answering a user's prompt.

The problem is that measuring AI's energy usage has always been messy. Modern training and inference runs happen across clusters of servers built from multiple types of chips, plus all the cooling infrastructure needed to keep them from melting. Each chip type comes with its own measurement tools, its own quirks, and its own margin of error. Stitching that together into a single, trustworthy number has been a genuine engineering headache.

Jae-Won Chung, a 30-year-old doctoral student at the University of Michigan, decided to solve that problem directly. He built Zeus, software that pulls energy data from across a heterogeneous cluster into one interface. That alone would be useful. But Zeus goes further: it can actually suggest optimizations.

One of Chung's more counterintuitive findings is that a model tuned to respond as fast as possible often burns way more energy than necessary. Slow it down slightly, even by a hair, and you can see big energy savings with barely a dent in perceived responsiveness. It's the kind of tradeoff that's invisible unless you're actually instrumenting the system closely enough to see it.

Turning Measurement Into a Public Benchmark

Zeus isn't just a research prototype sitting in a lab. Its measurements form the backbone of the ML.Energy Leaderboard, a public initiative led by Chung and colleagues that publishes energy data on dozens of AI models.

There's a notable gap in the Leaderboard's coverage. Models like OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini are closed systems. Their internals aren't open for inspection, so they can't be benchmarked the way open-source models can. Instead, the Leaderboard tracks open models like Alibaba's Qwen.

That might sound like a limitation, but Chung argues it's still practically useful:

• Open models share enough architectural DNA with their closed counterparts that their energy profiles serve as reasonable approximations.
• Developers building on top of open models get real, actionable data instead of guesswork.
• The methodology itself, unifying measurement across chip types, is reusable even if the specific numbers don't transfer perfectly to closed systems.

The work has gotten attention well beyond academia. Nvidia, Meta, and Google have all taken notice, and Chung is now exploring collaborations with all three to bring energy optimization into their systems directly. That's a meaningful shift. A few years ago, pitching a chipmaker or a hyperscaler on energy tooling built by a grad student would've been a hard sell. Now they're coming to him.

Part of what's changed is industry mindset. Chung says the field has broadly come around to treating energy use as an essential metric, on par with latency or accuracy, rather than a secondary concern to worry about later. A big chunk of his job now isn't just building better measurement tools. It's convincing people that the problem is tractable at all. As he puts it, the message he's spreading is that "you can actually tame it."

That framing matters because energy efficiency in AI has often been treated as a black box: expensive to measure, hard to act on, and easy to deprioritize when there's a product deadline looming. Zeus and the Leaderboard are an attempt to change that calculus by making the data legible and the optimizations concrete.

There's also a broader industry pattern here worth noting. Efficiency work in AI infrastructure tends to lag behind capability work by a few years, then catch up in a rush once the constraints become undeniable. Compute costs did this. Memory bandwidth did this. Energy looks like it's following the same arc, except now it's colliding with actual grid capacity limits and public backlash over data center siting, which raises the stakes considerably.

Chung's timing, in that sense, is good. He built tooling for a problem before it became existential, and now that it has, the infrastructure and credibility are already in place. The chipmakers and model developers now sniffing around his work aren't doing so out of academic curiosity. They're doing it because power consumption has become a bottleneck that directly threatens their ability to scale.

Key Takeaways

Zeus solves a real, unglamorous engineering problem: unifying energy measurement across heterogeneous chip fleets that previously required a patchwork of chip-specific tools with inconsistent margins of error. That unification is what makes the ML.Energy Leaderboard possible, giving developers a public, standardized reference for energy costs across dozens of open models, even as the biggest closed models remain opaque by design.

The more actionable insight might be Chung's finding around response latency: deliberately slowing inference slightly can yield outsized energy savings, a tradeoff that's easy to miss without the kind of granular instrumentation Zeus provides. As Nvidia, Meta, and Google explore collaborations with Chung, the underlying signal is clear. Energy efficiency has moved from a nice-to-have metric to a core design constraint, and the tooling to actually measure and act on it is only now catching up to the urgency of the problem.