Sasha Luccioni, who spent four years building emissions-transparency tools at Hugging Face, is leaving to launch Sustainable AI Group, a consultancy co-founded with former Salesforce sustainability chief Boris Gamazaychikov. The firm will advise companies on reducing the environmental cost of AI deployments, with an initial focus on agentic workloads and model right-sizing. The bet is that enterprise demand for emissions data has outpaced what model providers are willing to disclose.
Luccioni built her reputation at Hugging Face by publishing an energy-efficiency leaderboard for open-source models and pressing closed-model labs to release per-query energy and water figures. She has been one of the more visible critics arguing that companies including OpenAI, Anthropic, and Google are deliberately withholding sustainability data. Sustainable AI Group will sell that critique as a service.
The pitch lands at a moment when corporate buyers are getting squeezed from two directions. Boards and ESG officers want quantified emissions from AI usage; employees are pushing back on mandated Copilot rollouts whose footprint nobody can measure. "You need to be quantifying this," Luccioni said of the pressure her clients are fielding. The other complaint she hears: "You're forcing us to use Copilot—how does it affect our ESG goals?"
Key facts
- 01Sasha Luccioni is leaving Hugging Face after four years to co-found Sustainable AI Group.
- 02Her co-founder is Boris Gamazaychikov, the former Salesforce sustainability chief.
- 03The firm will advise companies on reducing the emissions footprint of AI agents and matching workloads to right-sized models.
- 04Luccioni argues the EU AI Act's sustainability clauses are now generating the first reporting requirements on data-center emissions.
- 05She wants ChatGPT and Claude to display per-query energy and emissions data in their UIs.
Luccioni argues the conversation has moved past abstinence. "It's not about not using AI. I think we're past that. It's choosing the right models, for example, or sending the signal that energy source matters, so customers are willing to pay a little bit more for data centers that are powered by renewable energy," she said. The consultancy's working thesis is that procurement leverage, not consumer guilt, is what changes data-center sourcing.
“I wish there was a little meter or info box on the ChatGPT or Claude UI that tells you at the end of each query or conversation how much energy was used.”— Jaeden Schafer
Regulatory pressure is uneven by geography. The EU AI Act included sustainability clauses from its first drafts, and the initial reporting requirements are now coming online. The International Energy Agency has been compiling AI energy-use reports but depends on member countries for the underlying numbers, and many governments do not yet collect data-center-specific figures. Some countries have started rejecting new data-center builds as a result.
Asked what single disclosure she would extract from Sam Altman or Dario Amodei, Luccioni named the user interface. "I wish there was a little meter or info box on the ChatGPT or Claude UI that tells you at the end of each query or conversation how much energy was used," she said, along with the emissions and the generating source. She thinks the first major lab to commit to renewable-only data centers would gain a marketing edge: "It's like when Anthropic said no to the US government for military use. It did give them a boost. A cultural boost."
The second strand of her argument is technical: most enterprise AI work does not require a frontier LLM. Classifier models still do the bulk of what businesses call "AI productivity," and token-level telemetry — which Google publishes — lets buyers route simple queries to cheaper, smaller models. "If you guys want to just search company documents, this is the model to use. It's simple and cheap. And if you actually want to do deep research, here's a more complex model," she said, describing the internal policy enterprises should be writing.
Luccioni is blunt about why that message struggles to land: vertical integration. "Many of the big companies making the models are also the ones that are selling you the compute. It makes sense for them to sell you the largest model, because then you need the most compute," she said. The same point applies to first-party products built on top of those models, which is the full stack at every hyperscaler.
The counterweight is data quality. Luccioni concedes the numbers she wants regulators and buyers to act on are still patchy, and per-query figures are small enough that public framing tends to oscillate between dismissal and panic. "It's true that maybe each individual query is not a big deal. But then you multiply it by the number of people that use these things—it is a really hard conversation," she said. Her response is that transportation and nutrition data are also imperfect and still useful: "We still need the numbers on energy and water use in order to make informed decisions."
Sustainable AI Group is entering a market that did not exist two years ago and is now driven less by climate advocacy than by procurement reality. Enterprises buying agentic AI at scale need to log the same emissions and supply-chain data they already track for cloud and travel, and the model providers have so far declined to ship it by default. A consultancy that can normalize that data across vendors — and tell a CFO whether to route a workload to a frontier model or a classifier — has a clearer revenue path than most AI-policy outfits. The harder question is whether the labs respond by opening up, or by continuing to bet that compute lock-in keeps the disclosure question at arm's length.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
