Cryptography approach for measuring AI data center emissions

Context/Problem:

Many medium to large companies are required to report their annual carbon emissions under the EU Corporate Sustainability Directive. As companies purchase enterprise AI licenses, they must report the energy consumed from AI usage under Scope 3 emissions.

However, cloud providers only give proxy data. This is because providing exact energy usage requires disclosure of the efficiency of their training and inference. This directly impacts the valuation of the AI model.

Action/Solution:

I devised a new use case of cryptography to help cloud providers safely transmit sensitive carbon emissions data to clients.

Cloud providers can use Zero-Knowledge Proofs to disclose emissions data. They would insert data like energy usage per token, tokens used, hardware, training, inference, and emissions data. Then, the emissions calculation methodology is encoded as a cryptographic circuit, which takes the provider’s sensitive data and generates a proof that can be independently verified.

Auditors can verify the emissions breakdown for customers and provide them with the end emissions data while the proprietary company data remains confidential.

This would allow companies to meet EU Corporate Sustainability Reporting Directive requirements with precise data. However, right now the directive doesn’t require exact data, so there is no incentive for companies to seek this information or for cloud providers to distribute it.

If regulations became more stringent, this would be an appropriate use case where cloud providers and auditors could participate in a profit-sharing mechanism.

Outcome/Impact:

I created a research poster to present this information. It was selected as the “Best Overall” poster at the London Student Sustainability Conference from 57 academic posters presented at the event.