Day 1 · Keynote · 10:40–11:10
Looking beyond the model: the physical footprint of AI
Vrushali Gaud · Google Sustainability
Tap a dashed word or a blue reference to open it below.
In one line
AI's footprint is physical (electricity, grids, water and hardware), so the burden of using it responsibly shouldn't fall on users: build the system right in the first place, measure it, and use AI where the benefit is clear.
Why SMASH exists
SMASH began about two years ago as a conversation between Mila and Google, to make room for topics most curated conferences don't. Her background is automotive, not ML research. Her analogy: cars changed lives, then the industry spent decades fixing the unintended consequences. How do we avoid doing the same with AI?
Look beyond the model
People see the app, and maybe the model, but not the physical side underneath. Teams working on energy, chips and models are often siloed, and few people look across which chip, which model and which region a workload uses.
Information alone doesn't change behaviour
Most of the room didn't know the energy or water cost of their own AI use, and most wanted to. But in sustainability, information rarely changes habits on its own. Her position: don't put the burden on users. Give them real options, and design the system well from the start.
The physical stack
She described four layers: hardware and infrastructure, software (serving and orchestration), models, and applications. At the infrastructure level, three things matter most:
- Clean energy: electricity is the biggest factor in emissions, and where a data centre sits matters. The aim is clean energy around the clock, rather than as a yearly total.
- Grids: flexible work can move in time and place, but reaching clean power through old, fragmented grids is the hard part.
- Water and cooling: water cools chips best but competes with communities in water-stressed regions. Air cooling saves water but uses more energy.
The biggest challenge is speed: data centres are being built faster than clean energy and grids can follow.
Measuring it, and AI used for good
She shared Google's published per-prompt figures: over 12 months, energy per median text prompt fell 33× and its carbon footprint 44×, and a median prompt now uses less energy than watching about nine seconds of TV. She was candid that this is a starting point, since nobody sends just one prompt. On the benefit side she pointed to , which forecasts river floods for over 700 million people in 150 countries, FireSat for early wildfire detection, and work on reducing .
Unofficial notes by Rai Vaheed, written up after the talk. The ideas are the speaker's; any mistakes are mine. Corrections: vaheed.raihan@gmail.com · View the source of these notes