Shout out to Dan Mall. I asked him last week, to shoot holes through some of my thoughts on the argument about the environmental impact of AI vs our personal usage of things that are bad for the environment.
I whined “Isn’t it interesting how we want ChatGPT to “change” a company we don’t control. Yet we’ll swear by almond milk and eat hamburgers without thinking twice.”

He said:
“You need to be more upfront about what you actually think about AI and the environment, exact quote “you are the dang leader, tell the team what is in bounds and out of bounds”.
I got tired of the argument, so I went and looked at myself first
We’re a B-Corp. So my company should be asking me tough questions as the environment is a stakeholder. But something about it kept bugging me, and it took me a while to name it.
We hold AI to a standard of scrutiny we do not hold ourselves to.
Not one person in the company has ever said, hey wil, I care so much about the environment that I want you to show me your flights and your impact on the environment and discuss what we’re going to do about that.
Most real change starts with looking at yourself, and if I’m going to ask people to bring me data, I’d better show up with my fucking own.
Have you ever compared your AI usage vs your travel impact? It’s brutal.
01 Start with me
I flew 70k miles last year.
Mostly in business class.
Forty-one segments between August 2025 and August 2026 — roughly 2.8 times around the Earth. Brazil twice, Munich, London, Mexico City, Oaxaca, and a dozen domestic hops. (Claude analyzed my calendar this could be off).
43.6 t
CO₂e — one year of my flying
22.6 t
If I’d flown economy the whole way. This is the floor.
43.6 t
Adjusted for the seats I actually sat in.
87.1 t
Including contrails and NOx at altitude,
14–16 t
The average American’s footprint, all sources. My flying alone is about triple it.
Now the AI number, done honestly
There’s a number floating around that says a median AI prompt uses about 0.24 watt-hours. It’s a real measured figure from Google.
I modeled the expensive version: the IEA’s 2026 estimate for an agentic request with reasoning, which is about 50 Wh a pop — roughly 200× the chatbot median.
Two hundred heavy reasoning runs a day. Every single day. For a year. That’s about 1,350 kg of CO₂e.
Which is 3.1% of my flying.
To offset one year of my flying with AI, I’d have to kick off a heavy reasoning run every 4.5 seconds — every working second of every working day, for a full year.
AI usage is growing fast, which is exactly why it’s fair to examine todays multiples by this aggressive growth rate. But my personal AI use, run harder than almost anyone at Seer runs it, is a rounding error next to how much I fly.
Right-size your model usage.
Sonny recently posted internally that using a more efficient model could have an impact on our usage and I was like we should do that.
What if I used a smaller model?
Shifting 20% of my agentic work from Opus to Haiku saves 264 kg a year. That’s slightly more than my July round trip to Boston.
Almonds, hamburgers, golf courses, and pools
This is where my head really went, data, always and it started with water, ok if we care so much about water usage….
One almond takes about 3.2 liters of water to grow.
One quarter-pound beef burger: about 1,700 liters.
One simple AI prompt: 0.26 milliliters. Five drops.
Then I went to Arizona, because Arizona is a desert and they got tons of golf courses and pools (used deep research).
| Use | Annual water | vs. data centers |
|---|---|---|
| All Arizona data centers (2025) | 0.9 billion gal | |
| ~505,000 residential pools | ~8.1 billion gal | ~9× |
| Golf courses, Phoenix AMA alone (174 courses) | ~32.4 billion gal | ~36× |
| Pools + golf combined | ~40.5 billion gal | ~45× |
| Agriculture, statewide | >70% of all AZ water | ~1,700× |
Statewide, about 74% of Arizona golf water is potable. In the Phoenix AMA only 21% is reclaimed. Phoenix golf on drinking water alone is roughly 28× every data center in the state.
Don’t you take away my pool…
Arizona has around 505,000 residential pools — one for every 13 residents, more per capita than any state in the country.
We built half a million private pools in a desert without a single public meeting about it. Nobody called that an environmental crisis. Something else is at play here, AI is striking a chord that I think all of us need to dig in on. Why THIS and never THAT?
Wil, what is “in bounds” at Seer?
Everything above is research. This next part is the thing Dan wouldn’t let me leave out, and it’s the reason this post exists.
My number one job is keeping people employed. That’s it. That is the top of my list and nothing is close to it.
There are roughly 180 people whose mortgages and tuition payments and health insurance run through this company.
We’re operating in a market where organic traffic is falling off a cliff and the entire mechanism by which people discover brands is being rewritten in real time.
If Seer isn’t leaned all the way into AI, we don’t get to have a values conversation at all, because we won’t be here to have it.
So the environment sits lower on my priority list than winning as a company. Not because it isn’t something I want to work on and fix. Because I can’t make it my driving force and still do the first job.
I’d rather say that plainly than let people assume I secretly hold it as priority one and just never act on it. Dan’s point was that my team deserves to know the real order, even if they don’t like it.
So here’s exactly what I will and won’t do as of today.
What I believe
- We hold AI to a scrutiny standard we don’t apply to our own choices. Business-class flights, vacations, pools, dogs, burgers, almond milk — all bigger, all mostly unquestioned.
- AI’s footprint is growing fast, so it’s fair to examine. What bothers me is examining it while ignoring our own.
- Data center water use in Arizona is genuinely a problem. Golf courses are a much bigger one and get almost zero protest.
- Most change starts with looking at yourself before you point at anybody else.
- Right-sizing models from Opus to Haiku where it fits is worth doing — real savings, no loss of capability. Just don’t sell it as a climate program.
- Individual, voluntary evals are the actual lever. Taylor moved a chunk of work from reasoning calls to static SQL queries — had the idea, tested it, proved it, shipped it, without being asked. That’s the model.
What I’m willing to do
- Fund an intern to work on this, on Seer’s time and with Seer’s money.
- Share Seer’s own data publicly and with government orgs so other companies can learn from what we find.
- Use my connections to push this work forward.
- Build the internal infrastructure so people can run real evals – Haiku vs. Opus, local models, etc
- Back anyone who wants to spend their spare pockets of time on efficient model environmental impact evals.
- Support anybody who tests an idea on their own initiative and brings the data back.
What I’m not willing to do
- Personally implement other people’s ideas. I care enough to look at myself, do the research, fund interns and test my own usage. If you got an idea, you gotta care enough to work with me but not assume I’m going to do it.
- Carve out extra paid time specifically for this work (outside of the intern)
- Change my travel
- Let this become the priority. The priority is survive and thrive at this time, so we can live into our values – win for our clients, so we can win for our co-workers, so we can win for our community. In that order.
FYI this post is about 40% written by AI, as it was research heavy. I spent 3 hours talking to Claude, sharing my ideas getting pushback, pushing it back, reading it outputs. Then I recorded a video for the team of my thoughts, and had claude read that transcript + all of our chats and help write a draft, and with my new wordpress MCP connector it pushed the draft right in.
The job to be done with any post is to think, and share not to “avoid em dashes”. So If you seen an em dash, in the immortal words of DMX…get at me dog, I’m ready.
Method. Flights: 41 calendar segments, Aug 2025 – Aug 2026, great-circle distances, 0.255 kg CO₂/passenger-km under 1,500 km and 0.195 above, business multiplier 2.5× long-haul and 1.5× short-haul based on booked seats. AI: IEA 2026 estimates of 50 Wh for an agentic request with reasoning and 1.1 Wh for a standard agent request; 0.37 kg CO₂e/kWh US average grid. Excludes training amortization, device energy, and non-agentic chat — all of which push the AI number up, not down. One acre-foot = 325,851 gallons.
Sources. IEA, Key Questions on Energy and AI (2026) · Our World in Data / Hannah Ritchie (2026) · Google Cloud, AI inference measurement paper (2025) · Bluefield Research via Circle of Blue · Ceres via Grist · ADWR and University of Arizona golf water studies · WRRC Phoenix AMA data · Arizona State Law Journal on allotment overages · KOLD Fact Finders on per-person pool water · US Bureau of Reclamation 2026 Level 1 Shortage declaration.

Leave a Reply