Page 105 - AI Standards for Global Impact: From Governance to Action

AI Standards for Global Impact: From Governance to Action

• Energy-efficient data centres (6 times more compute power than 5 years ago)
• A target for 24/7 carbon-free energy by 2030, through partnerships with geothermal and
small modular reactor (SMR) developers
• Water replenishment goals and AI-driven cooling optimization (e.g. DeepMind’s 40% Part 2: Thematic AI
reduction in cooling energy)

Figure 47: Google approach for sustainable AI

Some of the key takeaways are summarized below:

a) The energy efficiency of AI is a challenging problem to solve.
b) Energy and inference matter: Alongside the energy demands of large-scale model
training, attention should also be paid to sustainable inference – the true computational
bulk of AI.
c) Hardware-software co-design is critical: Emerging architectures (such as neuromorphic,
brain-inspired, or task-adaptive) can help cut energy consumption dramatically, as shown
in the use of small models and analogue systems.
d) Green AI should become the norm, not the exception – through better compute reporting,
evaluation metrics, and research funding incentives.
e) Standards and benchmarks are needed for:
o Model-level energy reporting
o Inference efficiency metrics
o Lifecycle footprint of AI chips and cooling systems
o Frugal AI design methodologies

f) Opportunities for ITU:
o Support the development of standardized AI energy efficiency frameworks, potentially
building on the ongoing work of ITU-T Study Group 5 and consider developing KPIs
to score the enviromental efficiency of AI systems..
o Consider developing new technical specifications on sustainable AI deployment, data
centre optimization, and model auditing.

93