Профиль

Thanks, Brendan!

Thanks, Jon!

Not every grid is fading. Texas went the other way: clean won 8% of days in 2019, 56% so far in 2026. Clean Energy Stripes are now live, an interactive site covering different U.S. grids, plus open-source code so anyone can make their own: johnbistline.com/clean-energy...

Take a guess which year the Indian Point nuclear reactors closed.

This comes from a new paper in Nature: www.nature.com/articles/s41...

A global shift to healthy diets would shrink agricultural land by 274 million hectares by 2050. This would be the single largest contraction in over 2,000 years, bigger than the drop after Rome’s collapse or the Black Death combined.

For more on IRA’s impacts, check out our review paper here: www.nature.com/articles/s44...

The Inflation Reduction Act turned 4 today. One of its side effects nobody predicted: the U.S. now makes more EV batteries than it makes EVs.

That's why we built an open framework for measuring AI emissions: a standard system boundary, common unit (kgCO2e per million tokens), and three-tier approach so companies can measure with whatever data they have today. More here: watershed.com/en-GB/blog/a...

For the same company with the same AI usage, one measurement method says 54 tons of CO2e, while another says 3. Both are “correct, which is the problem with how AI emissions get counted right now.

Breaking: U.S. Department of Energy discovers that solar panels do not work when the sun is not out.

Built with Stanford, Tsinghua, and practitioners at Block, Okta, and BC3. Check out the white paper: cdn.sanity.io/files/3ogo9b... A tl;dr blog: watershed.com/en-us/blog/a... Live tomorrow w/ Cooper Elsworth (Google) and Neil Jorgensen (Block): watershed.com/events/measu...

Several important levers we highlight: - Efficiently route tasks to the right model version (which can vary by orders of magnitude in its electricity/emissions) - Pick the region: US grids vary more than 5x in carbon intensity Decisions teams can make today.

Importantly, it reports kgCO2e per million tokens alongside electricity. So you can tell whether your footprint moved because of efficiency or the grid, and which levers to pull to address them.

It's tiered, so you don't need perfect data to start: 
Spend → Activity → Provider A spend-based backstop today; provider-reported per-token figures as disclosure matures. It meets you where your data are.

Ask three sources what one AI query emits and you'll get answers orders of magnitude apart. There's no agreed way to count training, inference, idle capacity, or which grid you assume. So we built one.

Two AI systems can do the same task with a ~30x difference in electricity use. Many companies have no idea which one they're running. Our new open framework for measuring corporate AI emissions... and actually acting on them. 🧵

Read the full QJE paper here: academic.oup.com/qje/advance-...

Insane paper: a Spanish city’s prayers for rain predicted actual rainfall for 200 years. Turns out it’s just drought hazard curves, the same math that tells you a heat wave is more likely to break the longer it runs.

Ask three sources what a single AI query emits, and you'll get answers that span orders of magnitude. Sustainability teams are being asked to account for AI emissions with no accepted way to do it. Tomorrow, we publish a framework to fix that.