For about four months I was very proud of a number that turned out to be roughly a third of one percent of the actual number.
Here is the number I was proud of. I built an environmental baseline for our AI use, starting where any reasonable person would start, which is the tool everybody can see. Microsoft Copilot. Roughly 4,000 employees. Something like a million queries a year, give or take, depending on how you count a query, and I promise you people will argue about how you count a query. Run the energy math, apply the grid factors, be conservative in the places where you’re guessing.
Answer: about 200 kilograms of CO2e per year.
Two hundred kilograms. That’s a large dog. That is, for an enterprise of our size, so small it’s almost embarrassing to put on a slide. I remember thinking: well, that’s the good news, and also that’s a strangely anticlimactic four months.
Then Randy Kuite and Brennon opened a door for me, which is the polite way of saying they got me access to the AWS Carbon Footprint Tool, and I pulled the org-wide number for July 2025 through June 2026.
68.5 metric tons CO2e, market-based. 221.6 location-based. Plus about 69 cubic meters of water, which is its own conversation and one we are not nearly ready to have.
Sixty-eight point five metric tons against my carefully-constructed two hundred kilograms. Roughly three hundred and forty times larger. I had been measuring the flashlight while standing next to the furnace.
It is the most common failure in corporate sustainability, and I would like credit for at least committing it thoroughly. I did not measure the wrong thing. The Copilot number is real and correctly derived. I measured the visible thing. The thing with a logo and a license and a line item and an obvious owner, the thing that produces a dashboard when you ask it to. Meanwhile the infrastructure underneath, which has no logo and belongs to eleven different teams and never once introduced itself as an AI cost, ran three orders of magnitude hotter and produced no dashboard at all, because nobody had asked it to.
I got seven minutes to present this. Seven. My outline, which I wrote for myself and am reproducing without cleanup because the cleanup would ruin it:
Own it: The numbers. With the asterisk. Conservation of energy principle. You don’t have to use it. But if you do, know what it costs.
That’s the whole talk. Own it means the correction is mine, out loud, first, before anyone else has to be gracious about it. The asterisk is because market-based and location-based are two honest answers to the same question and if you only give one you’re picking your favorite. Conservation of energy because the compute did not appear from nowhere and does not vanish afterward, and if it feels free that only means the cost landed somewhere you aren’t looking. And the last two lines are the ethic: no mandate, no shame, just don’t pretend it’s weightless.
The equivalence I landed on was twelve Eames Lounges.
That’s 68.5 metric tons expressed in the embodied carbon of the chair. Twelve of them. A dozen chairs, and a room full of our people can see twelve chairs, can stand them up in their heads in two rows of six along a wall.
I originally had a second equivalence in the deck, something about road trips, because I have never met a carbon number I didn’t want to convert into driving. Somebody told me to cut it, and they were right for two reasons. First, the moment you give an audience two equivalences they stop listening and start doing arithmetic, comparing your metaphors against each other instead of absorbing the point. Second, and this is the part I actually care about: the road trip belongs to everybody. Every sustainability deck on earth has a road trip in it. The Eames Lounge is mine. Nobody else in that field can use that line, because nobody else makes the chair. If your comparator could appear in a stranger’s presentation unchanged, it isn’t yours and it isn’t working very hard.
I got a note that the talk needed a “why this matters to us” opening, and I pushed back, possibly with more confidence than the situation warranted: I don’t think anyone coming to this talk needs a reminder on “why” this is important.
I was wrong in an interesting way. Not because the audience needed convincing, but because I hadn’t articulated it even to myself, and when somebody pressed me later the real answer came out fully formed, which is usually a sign it had been sitting there a while:
This company has spent a century arguing that how you make something is part of what you make. We measure the carbon in a chair down to the foam. Down to the foam. So walking past the carbon in our AI would contradict the entire premise of the brand. It wouldn’t be an oversight, it would be a contradiction. And practically: if we’re going to put these tools in 4,000 people’s hands, the credible move is to know the cost and manage it, not to discover it later when somebody outside the building does the math for us and asks us why we didn’t.
Underneath the headline number there’s a governance finding I find more alarming than the tonnage. About 60% of the Copilot footprint runs through Copilot Chat, which was never procured, arrived as part of something else, and cannot be switched off at the tenant level. We do not have a knob. And Copilot Studio agent sessions are a genuine data gap with something like a 30x energy multiplier attached, which means the fastest-growing category is the one I can see least clearly. Right now most of us get usage dashboards, not impact dashboards. That gap is the ask, and it’s an ask for vendors, not for users.
One more thing I’ve started saying, mostly to myself: the most sustainable prompt is the one you didn’t have to write twice. Bad prompting is an emissions problem and not only a quality problem. Every rerun is real electricity and real water for an output you already should have had. Craft turns out to be a climate practice, which I did not expect and find slightly delightful.
And the last line of the outline holds. Don’t dodge those with concerns. Somebody told me not to soften that, and I haven’t, because the people raising their hands about the environmental cost of this stuff are not obstacles to the program. They’re doing the work I was doing badly for four months, and doing it earlier.