AI and responsible digital: combining innovation and sobriety

AI's carbon footprint: a governance issue

Generative AI is one of the most energy-intensive digital technologies. Training a large language model (LLM) like GPT-4 consumes the equivalent of several hundred tons of CO2. Each ChatGPT query consumes around 10 times more energy than a classic Google search.

For Swiss companies, this reality raises a strategic question: how do you tap into generative AI while honouring responsible digital commitments and the federal climate goals? This question is all the more pressing as generative AI is becoming central to B2B strategies in Switzerland.

Key figures of AI energy consumption

Before talking about solutions, it helps to set out concrete orders of magnitude:

• GPT-3 training: about 1,287 MWh of electricity and 552 tons of CO2 emitted, according to a University of Massachusetts Amherst study (2021).
• GPT-4 training: estimates exceed 50 GWh, the annual consumption of more than 10,000 Swiss households.
• A ChatGPT query: between 0.001 and 0.01 kWh per query, against around 0.0003 kWh for a Google search.
• Water consumption: Microsoft reported a 34% rise in water consumption in 2022, largely linked to cooling its AI data centres.
• Global projection: the International Energy Agency (IEA) estimates that data centre electricity consumption could double by 2026, exceeding 1,000 TWh, equivalent to Japan's consumption.

These figures are not abstractions. They translate into energy bills, carbon footprint and reputational risk for companies that deploy AI without discernment.

The three sources of environmental impact

Model training

The training phase is the most intensive. A model like GPT-4 requires thousands of GPUs running for several months. According to the University of Massachusetts, training a single large model can emit as much CO2 as five cars over their entire lifetime.

Good news: this phase is carried out by the providers (OpenAI, Anthropic, Google), not by user companies.

Daily inference

Inference, i.e. each query sent to an LLM, represents a considerable cumulative energy cost. With millions of daily queries, the total inference consumption now exceeds that of training.

This is the lever where companies have a direct impact.

Hardware infrastructure

The data centres that host AI models require massive cooling systems and consume significant amounts of water. The race for GPUs also accelerates rare-earth extraction.

The Swiss landscape of data centres and energy

Switzerland holds a unique position in Europe in terms of data hosting. The country counts more than 90 commercial data centres, mainly concentrated in the cantons of Zurich, Geneva and Vaud. Several factors make the territory particularly suited to responsible AI use:

• Decarbonised electricity mix: about 60% hydro and 30% nuclear. Switzerland emits an average of 26 g of CO2 per kWh produced, against 385 g in Germany and more than 400 g in the United States.
• Climate favourable to cooling: moderate temperatures allow free cooling a large part of the year, reducing energy consumption dedicated to cooling by 30 to 50%.
• Regulatory and political stability: a solid legal framework that fosters trust from international companies.
• Proximity to research centres: EPFL and ETH Zurich are among the world leaders in research on AI model efficiency and sustainable computing.

Hosting your AI workloads on Swiss infrastructure therefore represents a concrete carbon-footprint advantage compared with data centres in the United States or Asia.

5 practices for responsible AI in business

1. Choose the right model for each task

Not all use cases require GPT-4 or Claude Opus. For text classification, email triage or simple response generation, a compact model (GPT-4o mini, Claude Haiku, Mistral Small) consumes 10 to 50 times less energy for results that are often comparable.

The rule: use the smallest model capable of doing the task.

2. Optimise prompts and reduce tokens

Well-structured prompts produce shorter, more accurate answers. Each token saved reduces consumption. Investing in prompt engineering is both a performance gain and an act of sobriety.

3. Cache recurring responses

If your AI chatbot answers the same question 50 times a day, there is no point calling the LLM each time. A semantic cache system serves pre-generated answers for similar queries.

4. Favour committed providers

Some cloud providers power their data centres with renewable energy. In Switzerland, data centres benefit from an electricity mix that is already largely decarbonised (hydro, nuclear). Choosing Swiss or European hosting reduces the carbon footprint compared with a gas-powered American data centre.

5. Measure and report

What is not measured is not improved. Integrate AI consumption tracking into your CSR reporting. Tools such as CodeCarbon or ML CO2 Impact let you estimate the carbon footprint of your AI uses.

Implications for Swiss businesses

FADP compliance and digital responsibility

Since September 2023, the new Federal Act on Data Protection (nFADP) has imposed stricter requirements on transparency and personal-data processing. The use of generative AI raises direct questions: is data sent to LLMs processed in compliance with the FADP? Do models hosted outside Switzerland meet international transfer requirements?

A responsible approach involves:

• Mapping data flows sent to AI services and documenting the legal bases of processing.
• Favouring on-premise or Swiss-hosted solutions for sensitive data (customer data, HR data, financial data). The choice between SaaS and custom development has a direct impact on this question.
• Informing users when AI intervenes in the processing of their personal data.

Green IT certifications and labels

Several frameworks allow Swiss companies to structure and showcase their approach:

• ISO 14001 (environmental management): applicable to the management of IT infrastructure.
• Responsible Digital label (INR): a French framework increasingly recognised in French-speaking Switzerland, covering strategy, procurement, uses and the digital lifecycle.
• Swiss Digital Initiative: led by Doris Leuthard, this initiative promotes trust and responsibility in digital. Its digital trust label is directly relevant for AI deployments.
• Science Based Targets initiative (SBTi): for companies wishing to align their carbon trajectory, including the digital footprint, with the Paris Agreement objectives.

Competitive advantage and customer expectations

In Switzerland, 73% of consumers say environmental practices influence their purchasing decisions (Deloitte 2024 study). For B2B companies, RFPs increasingly include CSR and digital sobriety criteria. Adopting responsible AI is no longer just a matter of ethics; it is a genuine commercial differentiator.

Switzerland, a potential model

Switzerland has unique assets to become a leader in responsible digital applied to AI:

• An electricity mix among the cleanest in the world (60% hydro)
• A culture of precision and optimisation rooted in the economic fabric
• A regulatory framework that encourages transparency (FADP, Swiss Digital Initiative)
• Top-tier research centres (EPFL, ETH Zürich) working on AI model efficiency, with promising applications in sectors such as healthcare

Swiss companies that adopt a responsible approach to AI do not just reduce their impact: they strengthen their credibility with customers increasingly sensitive to these issues.

FAQ

Does responsible AI cost more to set up?

No — quite the reverse, in most cases. Using smaller models and optimising prompts directly reduces API costs. Semantic caching cuts the number of billed queries. Measuring consumption helps identify waste. A responsible AI approach is, first and foremost, an exercise in economic efficiency.

Can you use generative AI while remaining FADP-compliant?

Yes, provided you take the necessary precautions: do not send non-anonymised personal data to LLMs hosted outside Switzerland, document processing, and inform the persons concerned. Solutions hosted in Switzerland or deployed on-premise allow you to combine performance and compliance.

Where do I start concretely?

Three priority actions: (1) carry out an inventory of existing and planned AI uses in the company, (2) define a model-choice policy based on the principle of proportionality, (3) integrate AI consumption indicators into existing environmental reporting. These three steps can be set up in a few weeks.

Operational summary

• Generative AI consumes an order of magnitude more energy per query than a classic Google search.
• Companies can act on inference: model choice, prompt optimisation, semantic cache.
• The rule: use the smallest model capable of doing the task.
• Switzerland has a clean electricity mix and an ecosystem favourable to responsible AI.
• FADP compliance and Green IT certifications structure a credible, valuable approach.
• Contact MCVA Consulting to integrate AI responsibly into your digital strategy and turn digital sobriety into a competitive advantage.