The Problem
How much water does a single AI query consume? OpenAI's CEO has claimed about 0.3 ml — a few drops. Independent researchers at UC Riverside, using broader lifecycle accounting that includes electricity generation, put the figure closer to 10 ml. That's a 30x discrepancy for the same activity, and that's before you get to the harder questions.
For context, a single cup of coffee carries a lifecycle water footprint of roughly 140 liters, and a cheeseburger somewhere around 2,500 liters — but those numbers count agricultural rainfall and irrigation, while the AI figures count operational cooling water. The comparison itself breaks down. And that breakdown is the point.
There's a narrow set of AI use cases where the resource tradeoff is actually calculable. An AI system optimizes a process that consumes 10,000 gallons of water per month, reduces consumption by 20% — you can compute a payback period for the compute resources required. The units match. The math works. But most AI adoption decisions aren't like that. They trade resource costs denominated in energy and water for returns denominated in productivity, speed, or "better decisions." The numerator and denominator aren't in the same currency — and in many cases, the numerator (value delivered) is assumed rather than measured.
Meanwhile, AI adoption is accelerating across industries while energy and water infrastructure are under measurable stress from AI-driven demand growth. Regulatory interest in AI environmental disclosure is building. Organizations that aren't asking hard questions about resource tradeoffs today will be forced to soon, and the consultancies advising them should have a framework ready before that happens.
So the interesting question isn't how to build a universal payback calculator. It's understanding where a given use case falls on the spectrum from calculable to fundamentally incommensurable — and what organizations should be asking themselves at each point along it.
Research Design
I'm combining a broad survey for pattern identification with qualitative interviews for depth, grounded in published resource consumption data. The logic is pretty straightforward: survey data alone tells you what people report — not why. Interviews alone give rich insight but can't establish patterns across roles, industries, or organization types. The quantitative reference layer keeps everything anchored to real resource numbers instead of abstractions.
That said, I'll admit I went back and forth on whether interviews would add enough to justify the timeline. What convinced me was testing the survey instrument — even in pilot responses, I could see patterns forming that I couldn't explain without talking to people. The survey surfaces that a gap exists between what people know and what organizations do. Only conversations can get at why.
Survey Instrument
The survey is 12 questions across four sections, designed for about 3 minutes of completion time. It's structured to do more than collect opinions — the question sequence tests specific hypotheses about how organizations think about AI costs.
Section 1 (About You) captures role, organization type, and experience level — the segmentation variables that make cross-tabulation meaningful. A strategy director at a consultancy and a junior engineer at a startup may both be "somewhat aware" of AI resource costs, but the implications for organizational decision-making are very different.
Section 2 (AI Usage) establishes frequency, use cases, and personal stance. This isn't just demographic data — it creates the baseline for testing whether heavy AI users are more or less aware of resource impacts, and whether personal stance correlates with organizational behavior.
Section 3 (Resource Awareness) is where the analytical design gets specific. Five questions in deliberate sequence:
Q7 asks about prior awareness of AI resource consumption. Q8 asks whether resource costs have come up professionally. Q9 — a question I added specifically to test the "what gets measured gets managed" hypothesis — asks whether the organization tracks financial costs of AI tool usage (API spend, token costs, per-seat licensing). Q10 asks what factors were considered in AI adoption decisions. Q11 presents a concrete scenario: if a specific AI task consumed 10x the energy or water of the non-AI alternative, would that change usage?
The Q9→Q10→Q11 sequence creates an analytical arc: are you tracking cost → what did you weigh → would information change behavior? But the cross-tabulation I'm most interested in is Q9 against Q7 — whether organizations that measure financial AI costs are also more aware of resource costs, or whether financial tracking and resource awareness operate as completely separate concerns. Then Q9 against Q11: does tracking financial costs predict willingness to act on resource information?
Here's what I'm watching for. If organizations that actively track AI spend are also more receptive to resource impact data, that's a lever — financial discipline as a gateway to environmental awareness. If there's no correlation — if financial tracking and resource awareness are orthogonal — that's a different finding, and honestly a more important one, about how organizations compartmentalize different types of cost. I don't know which way this will go yet. I have a hunch that financial tracking correlates with awareness but not with willingness to change behavior, which would be its own kind of interesting.
Section 4 is an optional interview opt-in, which also serves as a self-selection signal: who cares enough about this topic to volunteer 20 minutes?
Qualitative Interviews
8–12 conversations with practitioners at consultancies and organizations who sit at the intersection of personal AI adoption and organizational AI strategy. The interview protocol is designed around a core tension the survey can identify but can't explain: the gap between what people know about resource costs and what their organizations do about them. These conversations explore how adoption decisions are actually made, what gets weighed, and where resource costs sit (or don't) in that calculus.
Quantitative Reference Layer
Published resource consumption data — energy per query, water per training run, ranges by model and task type — compiled into a structured reference dataset. Combined with original survey data to ground the framework in real numbers rather than assumptions. This part is more straightforward than the primary research, but getting the comparability right is fiddly — different studies use different system boundaries, and being honest about that matters more than having a clean table.
What This Produces
A decision typology that maps AI adoption scenarios by how calculable the resource tradeoff actually is. On one end: concrete payback scenarios where AI optimizes water usage and you can measure the savings. In the middle: mixed-unit tradeoffs where AI speeds up design review — saves time, costs energy, and you're stuck comparing apples to kilowatt-hours. On the far end: fully incommensurable value claims where AI improves "decision quality" and the benefit is unmeasured against a measurable resource cost. Each category gets a different set of recommended questions, because the mistake I keep seeing is people trying to apply the same cost-benefit logic across all three.
Survey findings visualizing organizational awareness patterns, the relationship between financial tracking and resource consciousness, and the gap between stated concern and likely behavior change — segmented by role, organization type, and experience level.
A resource consumption reference structuring published data on AI resource impacts into a usable, comparable format.
A shareable report packaging findings and framework for practitioners, with the full research methodology documented transparently.
Process and Tools
This project is built with AI tools throughout — and that's intentional, not incidental. The research builds on foundational data and framing from MIT's IAP Sustainable AI program, extended through original primary research. AI is used for research synthesis, thematic analysis, secondary source review, survey instrument development, and technical implementation (the survey page itself was built in a single working session with Claude).
I should be direct about the tension here: I'm using AI to study AI resource consumption. Every query I run to synthesize research or test survey logic adds to the resource footprint I'm studying. I'm tracking that — this project's resource usage is logged in an Airtable usage log alongside the research it produces. I don't think that tension invalidates the work, but I'd be uneasy if I didn't name it. The tool that enables the research is also the subject of the research.
Data collection runs through Airtable for speed and simplicity. Analysis moves to Supabase — SQL queries, cross-tabulations, and structured analysis in a proper database environment. I picked tools based on what the stage of research needed, not on having a unified stack, and I'm being transparent about those choices because that's part of what this project is about.
Take the Survey
This research depends on perspectives from people who work with AI tools and make or influence adoption decisions. If that's you, the survey takes about 3 minutes and is completely anonymous.
If you'd prefer a conversation, reach out at jnangle@alumni.cmu.edu. I'm particularly interested in talking with people at consultancies, enterprise teams, and climate/energy organizations about how AI adoption decisions actually get made.
This is a live project. Research is underway — this page will evolve as findings emerge.