Nidhal Jegham joins Sustainable AI Group as Research Engineer to expand empirical measurement of AI's environmental impact

Nidhal is the lead author of "How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference," one of the most widely cited empirical methodologies for estimating the environmental footprint of large language model inference. The paper was recently accepted for publication in Communications of the ACM following peer review. His research has been cited in The Guardian, Fast Company, and L'Humanité and referenced across academic, industry, and policy work, with more than 680 citations to date.

At SAIG, Nidhal Jegham joins Chief Scientific Officer Dr. Sasha Luccioni in running the firm's empirical AI model testing program and developing statistically defensible methodologies that extend this work to proprietary frontier models. Today, most of the AI systems enterprises actually deploy operate as environmental black boxes: vendors disclose little, and existing measurement frameworks were designed for open models. SAIG's research stream is built to close that gap and put defensible numbers in the hands of enterprise buyers.

The firm is committed to open-sourcing all methodologies, measurements, and tooling developed through this work. Upcoming research includes extending empirical measurement to video generation models — where compute demand is growing fastest, and disclosure is weakest — and aligning AI energy measurement with the GHG Protocol so that emissions from AI operations can be reported in the same framework enterprises already use for the rest of their footprint.

A conversation with Nidhal Jegham

Your closed LLM inference energy estimation methodology has been cited hundreds of times and is now headed to Communications of the ACM. What gap were you trying to close?

Nidhal: For a long time, the conversation around the large environmental impact of model training seemed like the obvious culprit since it can take hundreds of thousands of GPU hours. But this misses the elephant in the room. Inference, although pretty negligible at the prompt level, happens billions of times a day, and cumulatively, its impact eventually adds up. The goal of “How Hungry is AI?” was to build a reproducible, empirically defensible, standardized, and transparent methodology to evaluate model inference side by side and become a robust methodology that researchers, enterprises, and policymakers can adopt for their specific use cases. 

The biggest measurement gap today is in proprietary frontier models — the ones enterprises actually deploy. Why has this been so hard, and what are you doing about it at SAIG?

Open-source models’ measurement frameworks exist and there has been real success in evaluating these models, notably the AI Energy Score. With open-source models, researchers can probe the system at every level: varying input and output length, batch size, quantization, and hardware configuration. The model is fully observable, and that observability is what makes rigorous measurements and evaluations possible. With proprietary models, everything is hidden in a black box:  you send a request, you get a response, and everything in between – the hardware it runs on, the batch size it is grouped into, and the infrastructure that serves it – is hidden.

Yet, these are precisely the models that enterprises are actually deploying at scale, meaning the systems with the largest footprint and the broadest adoption are also the least measured, and closing this gap is central to what SAIG is building. By deriving scaling laws from architectural principles and validating them through exhaustive empirical testing, we develop statistically robust methodologies to estimate this hidden footprint without relying on vendor self-reporting. The result is a non-partisan framework that is standardized, consistent, empirically validated, and most importantly, defensible against scrutiny from regulators, procurement teams, and the vendors themselves. And like any scientific instrument, it keeps on maturing with the field, as new open-source models are launched, more disclosures are published, and more empirical testing is conducted. 

Why does open-sourcing this work matter?

I’m convinced that the antidote to opacity cannot itself be opaque. The field is already struggling with black box systems and companies that disclose very little, leaving enterprises and regulators in the dark about the true environmental cost of this technology. Building closed methodologies to study closed systems would simply compound the problem. Open-sourcing our work transforms it from a SAIG product into an industry standard and infrastructure that researchers, regulators, and enterprises can all rely on, verify, and potentially improve upon. It also raises the bar for vendors, since they can engage with the numbers directly instead of dismissing them.

What's the theory of change? How does better measurement actually bend the curve?

Right now, sustainability is whatever vendors say it is, which means it functions as a marketing claim, not a decision criterion. Empirically defensible methodologies change that entirely—when you put defensible numbers in the hands of buyers, the market takes care of the rest. The real lever is market pressure, and the engine of market pressure is enterprise buyers that deploy and utilize these models at scale. When enterprises can compare models on environmental footprint with the same confidence they compare them on cost and latency, sustainability stops being a talking point and becomes a procurement criterion. 

About Sustainable AI Group (SAIG)

Sustainable AI Group (SAIG) is a research and advisory firm helping enterprises measure, compare, and act on the environmental impacts of AI. The firm was founded by two pioneers in the field: Dr. Sasha Luccioni, a renowned AI researcher with a decade of experience building field-defining benchmarks, and Boris Gamazaychikov, a strategist specializing in AI enterprise implementation. Together, they help organizations make AI sustainable in practice by grounding deployment decisions in empirical evidence. SAIG supports organizations adopting AI by conducting rigorous studies on fundamental environmental questions, developing evidence-based procurement guidelines, and engineering practical tools — such as the AI Energy Score — that bridge the gap between climate science and real-world corporate decision-making. Learn more at: sustainableaigroup.com