• MIT Technology Review's 2026 climate list credits Jae-Won Chung with software that measures the energy demands of open-source AI models, but it states no savings figure, hardware or workload.
• The ML.ENERGY Benchmark paper's abstract reports 'significant (sometimes more than 40%)' inference energy savings without changing what the model computes, which is a best-case figure, not a typical one.
• The Zeus paper's 15.3%-75.8% energy-efficiency improvement applies to DNN training, not to serving models in production.
• The two figures answer different questions, so quoting either as a cost-per-call saving for a serving fleet would be a category error.
• The benchmark is open-source, so the defensible step is to run it on your own model, batch size and hardware and weigh energy against your latency budget.

MIT Technology Review's 2026 Innovators Under 35 climate list, published on September 17, 2026, names Jae-Won Chung for software that measures the energy demands of open-source AI models. The profile gives no figures. The papers behind his work do, and their scope is narrower than a production team might assume: the largest savings range comes from training, while the inference result is a best case.

The profile is a pointer: it says what Chung built, not what it saved

MIT Technology Review's entry says Chung built software to make AI more energy-efficient and, by measuring the energy demands of open-source models, hopes the industry can better understand and address AI's impact. It states no percentage, hardware or workload. Two other honorees also use AI: Jing Wei for pollution tracking and Zhonghua Zheng for climate models built for cities. The remaining six work on lithium, steel, bioplastic packaging, fish feed and refrigerants.

The inference benchmark reports savings of 'sometimes more than 40%'

The ML.ENERGY Benchmark paper, by Chung and seven co-authors and a spotlight at NeurIPS Datasets and Benchmarks 2025, reports energy measurements across 40 model architectures and six task categories. Its abstract says automated optimization recommendations yield 'significant (sometimes more than 40%) energy savings without changing what is being computed by the model.' 'Sometimes' matters: this is a best-case figure from the abstract, not a typical or guaranteed saving. This desk has not rerun the benchmark.

Zeus's 15.3%-75.8% range is for training, not serving

Zeus, Chung's earlier NSDI 2023 paper with Jie You and Mosharaf Chowdhury, reports energy-efficiency improvements of 15.3% to 75.8% for diverse DNN training workloads. Its abstract covers recurring training jobs, not inference. A team quoting that range for a serving fleet would be applying a training result to a different workload. The width of the range also suggests the outcome depends heavily on the workload, so no single figure from it should be expected to transfer.

What a production team can take from these numbers

Neither paper is a cost-per-call forecast for a given deployment. The benchmark measures a fixed set of models and tasks, and a team's own model, batch size, hardware and latency target can differ. The ML.ENERGY leaderboard repository describes the benchmark as covering large language models and diffusion models, and the paper describes it as open-source and extensible to customized models and scenarios. The defensible step is to run it on your own configuration and compare energy against your latency budget.

Verdict: credible measurement work, but the headline numbers do not transfer

Chung's work is peer-reviewed, open-source and tied to specific numbers, which puts it well ahead of a profile blurb. The two figures on offer answer different questions: savings of up to 40% or more on inference in some configurations, and 15.3% to 75.8% on training. Treat both as hypotheses to reproduce on your own hardware. Treat the MIT list as an index of people worth following, not as evidence of savings. This analysis rests on the papers' abstracts, not their full method sections.