The physical reality of training and deploying complex artificial intelligence (AI) systems presents profound environmental challenges, notably regarding carbon emissions, energy consumption, and water depletion. To wit, training a single AI deep learning model can generate approximately 31 tonnes of carbon dioxide (CO2), a volume nearly five times the lifetime emissions of an average human. To mitigate these externalities, we might think of the integration of digital sustainability frameworks as a necessity.
The Swiss AI Action Plan
In this context, the Swiss AI Action Plan, co-created by digitalswitzerland and the Swiss Federal Administration, establishes a collaborative national roadmap designed to secure Switzerland’s global competitiveness while preserving core societal values of trust, privacy, and innovation. The plan spans five core pillars:
- AI Literacy: scaled national AI literacy and education
- Research: promotion of cutting-edge research and targeted knowledge transfer to the Swiss economy
- Infrastructure: building stable, domestic data center capacities
- AI-Ready Data: providing high-quality, linkable, and legally secure datasets
- Governance: an innovation-friendly, lean regulatory approach that protects fundamental rights and is supplemented by voluntary industry self-commitments
Our Anlaysis
At D4P, we had a closer look at all five above pillars. Regarding the ‘AI Literacy’ pillar of the above Plan: the national focus on upskilling the workforce and society to adapt to AI-driven changes is indispensable; however, it must transcend simple functional training to incorporate green & holistically sustainable computing competencies.
In practice, this means that, beyond the proposed ‘Playbook for AI Adoption’ for SMEs and targeted AI certificates for administrative professions, as well as ‘[…] a large-scale, national awareness campaign […] training one million people in the responsible use of AI’, the plan should be supplemented by upskilling researchers and building a R&I workforce capable of developing sustainable AI systems from the outset. Educating developers on computational efficiency, alongside accuracy, transforms the engineering culture. Rather than treating computational power as an infinite resource, a green- and sustainability-literate workforce integrates model optimisation as a fundamental engineering requirement.
As to the above ‘Research’ pillar, Switzerland is in no shortage already: e.g. the Swiss AI Initiative, launched in December 2023 with an initial seed investment of over 10 million GPU hours on the Alps supercomputer and a 20 million CHF grant from the ETH Domain, still represents the largest open-science foundation model effort worldwide. This initiative, representing the first joint endeavor of the Swiss National AI Institute (a partnership between the ETH AI Center and the EPFL AI Center), collaborates with over 800 researchers. Additionally, the International Computation and AI Network (ICAIN) connects this initiative with international organisations, including underserved regions of the world, promoting inclusive access.
On the ‘Infrastructure’ pillar of the Plan: resilient digital infrastructure requires smart investments in high-performance computing (HPC) to retain cutting-edge R&D within Switzerland. At the same time, Switzerland is among the top ten fastest-warming countries in the world, which is widely documented by the United Nations and the World Meteorological Organisation, as well as recently reported by the Swiss Academy of Natural Sciences. Consequently, the physical optimisation of Swiss data centers must balance peak computational demands with localised ecological cooling and a renewable energy integration.
The ‘AI-Ready Data’ pillar seems to be the most realistic and aligned with Federal data strategy: i.e. to expand AI use in policy-making, requiring high-quality, secure, and standardised data ecosystems. Indeed, unlocking existing data potential through standardised metadata reduces duplicate data collection and redundant model training, directly conserving energy. This phase integrates with the post-2030 agenda for long-term sustainability.
The ‘Governance’ pillar of the Plan is the most controversial one. The proposed in the above Plan a ‘lean regulatory approach’ ‘supplemented by voluntary industry self-commitments’ sounds a little like the recent EU Digital Omnibus package, allegedly increasing Europe’s competitiveness, while actually challenging some of the EU GDPR’s and AI Act’s founding principles regarding relevant sensitive areas, as well as reinforcing existing power imbalances, including the growing dominance of hyperscalers, rather than benefiting SMEs and SMCs.
We only hope that the ‘lean’ in regulatory approach of the Swiss AI Action Plan does not mean similar or comparable ‘lean’ in EU Digital Omnibus simplifying crucial digital regulations. Indeed, while reducing regulatory burdens could improve competitiveness, we must also acknowledge that most regulations serve important purposes, and even the strictest ones have valid justifications.
As regards the ‘voluntary industry self-commitments’ part of the Governance pillar of the Plan—if we want a truly sustainable AI in Switzerland, on the contrary, Swiss public should minimise passive reliance on supplier self-declarations and maximise active, legally defensible verification of safeguard of and respect of fundamental rights, national data traceability norms, international labour standards, and other applicable law and regulation. A Swiss AI Action Plan should not come at the expense of weakening protections, but rather should aim to enable innovation that is both technically robust and socio-ecologically aligned.
Conclusion
AI systems depend on data, yet their development must be guided by principles that ensure sobriety, circularity, energy efficiency, fairness, trust, accountability, and societal benefit.