As part of my work, we are grappling with how the apparent inexorable growth of AI might be contained within environmental limits. We think this requires system level or structural  intervention, and that contracts could play a critical role in this. Below is some early thinking, which I’d be delighted to receive feedback on.

The problem

Running AI models (the day-to-day use, which engineers call inference) is now the dominant and growing environmental cost of AI, larger than the one-off cost of training the models in the first place. Global data centre electricity use was around 415 TWh in 2024, roughly 1.5% of world electricity, and the International Energy Agency expects it to more than double to about 945 TWh by 2030. In the UK the trajectory is sharper: around 2.5% of national electricity now, projected to roughly quadruple by 2030. The contracts that govern this expansion contain almost no environmental obligations. The European Union’s template contracts for buying AI, published in March 2025, contain none of substance. No major law firm has filled the gap.

Why efficiency alone does not work

Individual fixes can cut the energy use of a single AI task significantly. The organisation I run – The Chancery Lane Project – has developed a Markdown for Agents tool, for example, which cuts the data an AI agent has to process by around 80% on a standard benchmark (16,180 units of text reduced to 3,150). On their own, savings like this make AI cheaper to run, demand grows, and the saving is absorbed elsewhere in the system. This is the rebound effect, sometimes called the Jevons paradox: when something becomes more efficient, people use more of it, and total energy use stays the same or rises. The energy does not disappear. It reappears further up the chain, most visibly as more data centres being built.

What we are exploring

A contract-based mechanism that connects savings at the small end of the chain (one AI query, one tool, one application) to the large end (cloud services and data centre construction), so that efficiency gains translate into less energy use overall, not more. This has two parts:

• A floor. A clear, shared standard for what good AI procurement looks like, built from existing rules and risks: EU law on company reporting (CSRD), the EU AI Act, software carbon measurement standards (ISO/IEC 42001, SCI), and growing litigation risk. Companies operating below the floor accumulate legal and regulatory exposure. The floor rises over time as adoption spreads.
• A cascade. Contract terms that pass an absolute energy limit (not a per-query target) up the chain, from demand to supply: from the application using an AI model, through the AI provider’s service contract, through cloud purchasing, into data centre construction. Efficiency creates headroom that cannot be filled by extra demand, because the ceiling does not move.

Why this is a pipe dream worth testing

Voluntary climate commitments by individual companies break down under competition. A company that absorbs the cost is outcompeted by one that does not. The mechanism we want to test does not depend on goodwill. It uses contracts and existing legal exposure to make the cooperative move the commercially rational one. The route in exists, via the data centre trade associations. Whether the mechanism holds across a real value chain is what we, at The Chancery Lane Project, want to find out.