Transparency as a baseline, not a concession.
Companies operating large-scale compute should disclose water withdrawal, energy source mix, and local grid impact for their facilities as a matter of course, not in response to investigative reporting.
Company / Public-interest statement
AI Development
and the Public Interest
The current trajectory of AI infrastructure development reveals a widening gap between private incentive and public cost.
The current trajectory of AI infrastructure development reveals a widening gap between private incentive and public cost. Data centers are being sited and scaled at a pace that outstrips the environmental review processes meant to govern them — drawing on municipal water systems for cooling, pulling from local power grids often still carbon-intensive, and doing so in communities that see little of the economic upside generated by the models trained and served on that infrastructure.
None of this is inherently unsolvable. It's a governance and incentive problem before it's a technical one. A healthier balance would mean:
Siting decisions made with, not for, affected communities.
Water rights and grid capacity are shared public resources; decisions about them shouldn't be made purely on the basis of tax incentives and speed-to-deployment.
Efficiency treated as a first-order design constraint.
Model architecture, training approach, and inference cost all have real environmental footprints, and "bigger is better" as a default strategy externalizes those costs onto the public.
Independent verification.
Claims about efficiency, sustainability, or reduced environmental impact — from any company, including our own — should be falsifiable and checked by someone other than the company making them.
Architectural constraint
Sustainability should be an architectural constraint, not a mitigation strategy
Reducing the environmental impact of current AI infrastructure is necessary, but it is not a sufficient goal. We should also question the premise that large-scale intelligence must require continuous extraction of water, energy, and material resources from the communities surrounding the infrastructure that supports it.
Water-intensive cooling, in particular, should not be treated as an unavoidable cost of computation simply because it is common in current data-center design. Where engineering alternatives can eliminate routine freshwater consumption for cooling, the appropriate question is not how much water a facility can responsibly withdraw, but why that withdrawal needs to occur at all.
The same principle should extend deeper into the computational stack. GPUs are one implementation of machine computation, not a physical requirement for artificial intelligence. Future AI systems should be evaluated not only by capability or speed, but by how efficiently they transform physical resources into useful computation. That means actively researching architectures that reduce heat generation, memory movement and power demand; specialized accelerators; closed-loop and non-evaporative thermal systems; photonic, neuromorphic and in-memory computing; longer-lived and reusable hardware; and model architectures that accomplish more with less computation.
Objective
Our objective should therefore go beyond efficient AI toward ecologically bounded computation: systems designed from the beginning to minimize their dependence on continuous resource extraction.
That standard must include the full lifecycle. Eliminating water consumption at the data center while increasing water use in electricity generation or semiconductor fabrication would merely relocate the cost. Likewise, improvements in efficiency are not sufficient if lower costs simply produce enough additional computation to increase total resource consumption.
The long-term engineering question is more fundamental:
How much intelligence can we produce while disturbing the surrounding physical environment as little as possible?
That should be treated as a technical research problem alongside model capability itself—not as an environmental problem to address after the computing architecture has already been chosen.
Accountability
The standard we intend to build toward.
This is the standard we intend to build toward, not just point at. As a company, we're committing to address these issues directly as we grow rather than treating them as someone else's problem to solve later: building efficiency and resource accountability into our design decisions from the start, publishing what we can verify rather than what sounds best, and inviting scrutiny of those claims instead of deflecting it. We don't have this fully solved today, and we won't pretend otherwise — but it's the direction we're accountable to.
Technology built without regard for these tradeoffs isn't advancing human prosperity so much as borrowing against it. The measure of responsible AI development isn't just what a model can do, but what it costs the people who never asked to live near where it runs.