Anthropic's 39% of flesh

I'm midway through an AI transformation engagement with a climate client and one thing I agreed to chase down for them is a reliable way to measure AI's climate impact.

After some research and 6K of ElevenLabs tokens, I had hacked together a short 5-minute podcast between three fake people to explain what I found. Wanna listen?

Let's acknowledge the absurdity for a second: using AI voices to explain an AI-generated report about AI's climate impact. While waiting for the audio to generate, I wondered if I had burned a tree in order to learn that AI burns trees. The fact it was for a climate organization felt like I was ordering "Burn Trees For AI" hats from China via a coal-powered steamship.

If you don't want to listen to the podcast, here's the best-guess numbers you should know from the research:

• A standard text prompt to an AI is about .24 watt-hours, or turning on am emergy-efficient LED lightbulb for two minutes
• A 30-person team chatting with AI daily for a full year consumes about two gallens of gas worth of energy. Total.

I consider myself a snowflake, though, and I don't think I'm just sending standard text prompts. I send fat, context rich text prompts. I ask AI to analyze reams of PDFs. I like using the most modern models because why would I use the Nissan Altima model when the Ferrari model is a click away.

I wanted MY number so I could properly assign my tree-burning guilt to myself.

And here's where it got interesting: I couldn't get it.

I pay Anthropic $100 a month for their power-user plan. I opened my usage dashboard expecting an odometer. What I found was a progress bar telling me I had consumed 39% of my weekly allowance.

39% of what? They don't say. There is no published number. No tokens, no watt-hours, no kilowatt-anything. I am 39% of the way to a ceiling whose height is a trade secret.

Imagine a water bill that just said "you've used: some."

I went back to the research to try to hack my way to something a little more concrete. The best independent estimates put a full week's usage cap somewhere around 2,500–3,500 standard messages. Soft number, and I'm telling you it's soft. When I multiplied this soft number across the year, my entire year of AI— drafting trainings, editing this newsletter, the tarot deck — is somewhere between one and four gallons of gas.

One Costco fill-up covers me until 2027, with change.

Connected, Not Compared

This new perspective didn't give me a guilt-free conscience, but a weirdly intimate connection to the headlines.

We all know those headlines- DATA CENTERS- but now my late night conversations with Tiddlywinks have a real (albeit small) cost. Not a proprietary 39% of mysterious hand waving, but something I can visualize at a gas station.

I've always known that my chats pass through fiber to a physical building where racks of hot metal think my thoughts back at me, cooled by real water moving through real pipes near a real town. Some of those buildings are marvels of self-contained engineering. Some of them are wrecking the communities next door. Before my research, this was an intellectual exercise.

Now, my one to four gallons flow through this intellectual exercise. I stopped being a spectator of the AI-climate story and became a character in it: a very minor character, a guy holding a shot glass of gasoline, but standing inside the story instead of cogitating at it.

I used to treat carbon accounting as either guilt-tech (feel bad with precision!) or absolution-tech (feel fine with precision!). But now, I'm seeing it more as connection-tech. The twenty minutes of napkin math did what a hundred headlines couldn't: it made the whole system feel real, because now I know which sliver of it is mine.

Don't Let Recycling Happen To You

There's a very real possibility that climate activism doesn't sufficiently pressure frontier AI labs into real disclosure. All the soft numbers above are thanks to little transparency and a lot of complexity with measuring AI's climate impact. It's not in Anthropic's financial interest to make it clear how many trees were mowed down to train the newest model.

This story has unfolded badly before: recycling.

Did you know that consumer recycling is largely a societal opiate distributed by Big Plastic? Look closely: you will never receive your own individual consumer recycling metrics. Some recycling systems just drop your carefully sorted cans into landfills with the rest of the trash. The potential impact from fervent individual recycling is miniscule compared to industrial recyclable waste. Learn more about this great deception from Yale's Climate Connections and from one of my favorite Ologies podcast episodes.

We've been snookered once, society-wide, by something we can physically hold in our hands.

Now, the AI industry writ large has the potential to pull the same con, except instead of cans, prompts are ephemeral and intangible.

I wish I had a great "So now what?" for you all to follow or take home, but measuring your own AI usage and connecting your part to the larger narrative is all I can offer today. Sure, you can email Anthropic Support and ask for your energy consumption metrics, but good luck.

Actually, here's your homework: be suspicious.

Big AI is disincentivized to be transparent. Your local Main Street doesn't want to be transformed into a data center shantytown. Construction firms want to build big things everywhere. Advertisers want lots of eyeballs on breathless articles about AI's climate impact. Energy infrastructure is complicated.

Don't layer on guilt to yourself or your organization simply because of a headline. Instead, ask yourself, "How do I center myself in this system and accept a right-sized climate cost?"

I think I'll walk to brunch today, instead of driving.

A TL;DR from the CRO

My carbon pawprint is difficult to measure, so instead I use a proprietary Snoot Usage Graph and you're currently at 12%.

-Roman Noodles, Chief Ruff Officer

Next Week: SMSCWAI

On Friday, August 28th at 1pm ET, we're going to tour a few Strategy Rooms.

Every SMSCWAI, I issue an optional challenge and this month's is to build a Strategy Room, a place where the AI already knows your flaws and helps you find fixes for your foibles. Read more about this in the last TL;DR, but here's the short version.

1. Pick your fighter: a Claude or ChatGPT Project, a Gemini Gem, whatever lets you save persistent instructions.
2. Give it a job: to be your advisor on one real thing you’re actually deciding — a business call, a hire, a pivot, a purchase.
3. Tell it explicitly to push back before it agrees with you. Name the failure mode you don’t want: flattery, hedging, telling you what you want to hear.
4. Argue with it. Ask it to provide evidence. Be humbled. Then decide what you will actually do.

Bring whatever you get — fully built, half-built, or “I made it argue with me about lunch and it won” all count. Messy is great.

I'm in the process of relaunching my website to transition off Squarespace and save on hosting fees. Instead of just saving a couple hundred bucks, it's been a crash course in Github and glee.

Getting "paid" to learn is always a good price.

Here I am, a non-coder, using Claude Code, to help me self-host a website. I asked Claude to teach me the fundamentals of hosting, to draw a diagram of what the new backend will look like, to tell me all the ways it can go wrong, to ideate all the new opportunities this freedom can give me, and to mock up some styles and copy that got my ole thinker chugging on overdrive.

TL;DR- This is so cool.

However, it was when I hopped from Claude Cowork to Claude Code that I really felt glee. I was able to describe something vague, and then watch as Opus turned it over, triaged it, made a demo, reworked the demo, and then presented the final product to me. And it was GREAT.

It really is magic, and what's even more stunning is that my neophyte AI-generated code has meaningfully created the front window of my business.

That feeling- glee- is hard to manufacture, so when it hits, it hits. I know that eventually I'll be embarrassed at all the no-nos I did in making this thing, but drinking in this unexpected glee is a reminder that feelings can be deliverables and gosh, it's a hella good deliverable.

This is my dog food and I will continue to eat it until the website goes live. Even if it tastes a little like gasoline.

Burping with joy,

Dan from Learn to Scale​

​Opt-out from the newsletter | Unsubscribe from all emails | Update your Preferences | www.learntoscale.us, Boston, MA 02119 |

​PS. This gives me zero climate guilt for watching twice.​