EFFICIENCY / ↓
LESS ENERGY
PER TASK.
Hardware, software and model design continue to improve the energy efficiency of individual AI tasks.
HACK4EARTH 2025 · GREEN AI
RETROSPECTIVE / 01FROM IDEA
TO PROOF.
Could AI help the planet and run cleaner?
In 2025, HACK4EARTH treated Green AI as a question to test in public. The brief asked builders to either reduce the footprint of AI itself or apply AI to a measurable sustainability problem — then show the work.
MISSIONS2
Build Green AI / Use AI for Green Impact
KAGGLE ENTRANTS267
Public competition participation count
TEAMS44
Teams recorded on Kaggle
KAGGLE SUBMISSIONS169
Leaderboard submissions recorded
01 / THE STATE OF GREEN AI
THE SIGNAL IS NOT ONE-DIRECTIONAL
MORE EFFICIENT.
MORE DEMAND.
The Green AI story is now a systems story: efficiency gains are real, but they sit beside rapidly growing use.
The IEA reports that energy use per AI task has been falling sharply as software and hardware improve. At the same time, AI-focused data-centre electricity use rose quickly in 2025 as adoption and more compute-intensive use cases expanded. Green AI is therefore less about claiming that AI has become “green” and more about measuring trade-offs, choosing proportionate models and infrastructure, and making resource use visible.
LESS ENERGY
DEMAND / ↑
MORE AI
EVERYWHERE.
Scale, agents, multimodal systems and wider deployment can increase total electricity demand even while individual tasks improve.
02 / WHAT GREEN AI MEANS IN PRACTICE
THREE LAYERS.
ONE SYSTEM.
A useful way to read the field without reducing it to one carbon number.
COMPUTE
Do the same useful work with less.
Right-size models, reduce runtime, compress, prune, route tasks intelligently and avoid unnecessary computation.
INFRASTRUCTURE
Run workloads where and when impact is lower.
Consider hardware utilisation, electricity mix, cooling, scheduling and the embodied cost of infrastructure — not only model accuracy.
OUTCOME
Ask whether the application improves a real system.
Energy, water, agriculture, mobility and other domains need evidence of benefit, not sustainability language alone.
03 / 2025 · FROM IDEA TO PROOF
NOT “PROOF THAT AI IS GREEN”
PROOF THAT THE
QUESTION COULD BE BUILT.
The 2025 experiment turned an abstract concern into a build brief with observable outputs.
01QUESTION
Can AI deliver useful outcomes while making its own resource use part of the design?
02BRIEF
Two missions: make AI itself lighter, or use AI for measurable environmental benefit.
03BUILD
Open-source prototypes, public experimentation and a month-long competition format.
04EVIDENCE
Teams were asked to show baselines, deltas, technical decisions and visible project outputs.
05CONTINUE
The strongest lesson was not a single winning model. It was the value of making sustainability inspectable.
THE SHIFT
2025 asked: “Can AI be greener?”
2026 asks: “What does responsible engineering look like when whole systems are under pressure?”
04 / THE 2025 DESIGN
TWO WAYS
IN.
MISSION 01
BUILD
GREEN AI.
Reduce the environmental footprint of an AI model or system. Establish a baseline, change something material, and compare the result.
MISSION 02
USE AI FOR
GREEN IMPACT.
Apply AI to a real sustainability problem and make the claimed benefit concrete enough to interrogate.
Competition framing and participation figures: Kaggle Community Olympiad · HACK4EARTH Green AI ↗