The Eco-Social Compact

On stewardship at the scale of inference

Every general-purpose technology arrives with a footprint. Steam, electricity, the internet — each was, in its time, an unaccounted externality before it became a managed system. What is unusual about artificial intelligence is the speed at which we can see its footprint, and the precision with which we can measure it.

An inference event leaves a trace in electricity, in water, in carbon, and in the human signal that trained the model that produced it. This is not a crisis. It is a starting condition.

We call this starting condition the eco-social compact — a working agreement between the systems we build and the commons they draw from.

The compact has two halves.

The ecological half is the planetary resource cost: the terawatt-hours of electricity, the billions of gallons of cooling and generation water, the megatons of carbon that move through the data centers in which our models think. By 2025, AI compute alone draws an estimated 625 TWh per year — roughly the annual electricity consumption of a mid-sized industrialized nation. That number is not a verdict. It is an orientation.

The social half is the human signal: the writing, the conversation, the lived experience that every model is trained on. This is also a commons, and it is also finite. Public, high-quality training data is on track to be substantially exhausted within the decade. Models trained on a closed corpus age the moment that corpus closes. The signal that keeps intelligence current is generated by people, with consent, in the present tense.

Both halves compound. Both reward stewardship. Both punish waste in the same currency: relevance.

Stewardship is not austerity.

The conversation around AI's resource consumption tends to oscillate between two unhelpful poles: alarm ("the planet cannot afford this") and dismissal ("efficiency will solve it"). Both miss the more interesting move.

Most inference today is treated as OPEX — performed, consumed, discarded, and silently re-performed by the next user with a similar question. The same reasoning, the same answer, the same physical cost, repeated indefinitely. This is not a property of intelligence. It is a property of the architecture we have wrapped around it.

The same reasoning, treated as CAPEX — captured, governed, attributed, reused — is performed once and contributes for a long time. The redundant compute disappears. The water it would have consumed disappears. The carbon it would have emitted disappears. Nothing is lost except the act of doing the same work twice.

This is the quiet truth at the heart of the compact: a substantial fraction of AI's footprint is not the cost of intelligence. It is the cost of forgetting.

Evergreen, not extractive.

The same logic applies to the training-data side. A model that drinks from a closed lake gets thirsty. A model fed by a living watershed — fresh, consented, governed at the source — stays current without catastrophic, full-corpus retraining. The compute saved by not retraining the world from scratch every eighteen months is not theoretical; it is measurable. The fidelity gained by feeding on the present rather than the past is also measurable.

We use the word evergreen for this kind of data ecosystem because the metaphor is exact. A forest does not need to be replanted every year. It needs to be tended.

A few things we will not do.

We will not pretend the footprint is small.

We will not frame the footprint as someone else's problem — not the hyperscaler's, not the regulator's, not the end user's. The compact is an industry-level instrument. Every participant has a hand on it.

We will not treat efficiency as the whole answer. Hardware gets more efficient every year. Demand grows faster, every year. Efficiency is necessary; it has never been sufficient.

We will not make the trace invisible. The numbers belong in the open.

A few things we will do.

We will measure, and we will publish what we measure. Anchored to the best public reports we can find. Assumptions exposed as knobs. Trajectories that reshape live when those knobs move. Anyone is welcome to disagree with our numbers. We have made it easy to.

We will build infrastructure that treats reasoning and data as the assets they are. Reasoning that is captured is not stolen from anyone; it is preserved for everyone who follows. Data that is governed at the source is not less abundant; it is more usable.

We will hold the position that intelligence and stewardship are not in tension. They are the same project at different time horizons.

An invitation, not an indictment.

Nothing in this compact is a critique of any actor in the field. The frontier labs, the hyperscalers, the open-source community, the operators of small models in basements and large models in data centers — every participant inherits the same starting condition. The math is the same for everyone. The opportunity is the same too.

We are not asking the industry to slow down. We are pointing at where the leverage is.

Every inference is a small physical act. It draws power from a grid, water from a basin, signal from a person somewhere who once wrote something. None of those facts are scolding. They are simply true, and now — for the first time in the history of computing — they are visible at the moment they happen.

Visibility is the beginning of stewardship.
Stewardship is the beginning of compounding.
Compounding is how a footprint becomes capital.

That is the compact. We invite you to hold it with us.