Generative artificial intelligence and responsible digital: what Switzerland can hold

Note revised on 25 May 2026. Article originally published in December 2025 — full rewrite.

Generative artificial intelligence is one of the most energy-intensive technologies in contemporary digital. This characteristic is now documented by several converging sources, and it enters into tension with the sobriety commitments that Swiss companies have progressively bound themselves to over the past decade. For the executive preparing the adoption of generative models in their organisation, the question is no longer whether generative AI consumes energy — it consumes substantially — but how to integrate it in an approach compatible with the climate objectives displayed and with the growing expectations of stakeholders.

This note sets out the mechanics of the environmental footprint, identifies the levers on which a user company can effectively act, and qualifies the specific assets of the Swiss context for responsible use.

The environmental footprint: where it is concentrated

The environmental footprint of a generative-AI system breaks down into three distinct sources, whose relative importance is not the same depending on the point of view.

The training phase of foundation models is the most intensive in absolute value. Training a large language model requires the mobilisation of thousands of specialised accelerators over several weeks or months, in data centres designed for that compute density. This phase is carried out by the publishers of the models — OpenAI, Anthropic, Google, Meta, Mistral and their equivalents — not by user companies. It is therefore largely out of direct reach for a Swiss SME, except to favour models whose publishers document and limit their training footprint.

Daily inference — every query sent to a model feeds this category — represents the cumulative energy cost that weighs on the model's service life, because it is exercised on massive volumes of queries operated continuously. A query to a generative model can consume more than a query to a classical search engine depending on the model, the length of the response and the infrastructure mobilised. This dimension is documented by the International Energy Agency and by several sector studies^[1]. It is on this lever that user companies have a direct effect, because it depends largely on their usage choices.

The underlying hardware infrastructure — data centres, cooling systems, water consumption, rare-earth extraction for components — carries the third source of impact. This dimension depends largely on the choice of host, and Switzerland presents particular assets that deserve to be identified.

Five levers where a user company can effectively act

On the daily-inference dimension, five concrete levers are available to the user company, in decreasing order of efficiency.

The choice of model for each task constitutes the first lever, and probably the most impactful. Not all uses require a very large model. For tasks of text classification, information sorting, generation of simple responses or translation, compact models consume several orders of magnitude less energy for results often comparable to those of the leading models. The practical rule consists of favouring the lightest model capable of accomplishing the task, rather than systematically resorting to the most powerful available.

Prompt optimisation constitutes the second lever. A structured prompt, which clearly formulates the request and specifies the expected response format, generates a shorter and more precise response. Every saved response token directly reduces consumption. This discipline belongs to a competence — prompt engineering — that does not require high technical prerequisites but does require iterative practice.

Caching of recurring responses constitutes the third lever. When a conversational agent handles the same question several times, or semantically very close questions, it is rarely useful to call the model each time. A semantic cache system serves pre-generated responses for similar queries, reducing both operational cost and cumulative consumption.

The choice of infrastructure provider constitutes the fourth lever. Hosts are not equivalent on the carbon footprint of their data centres — the energy mix varies considerably from one geography to another. Favouring hosting in Switzerland or in Europe on infrastructure powered by renewable or decarbonised energies can reduce the footprint of an equivalent compute workload, depending on the effective energy mix and the efficiency of the data centre mobilised.

Measurement and reporting constitute the fifth lever. Without measurement, no improvement. Integrating the tracking of energy consumption linked to the use of generative models into the environmental and social-responsibility reporting produces two beneficial effects: it identifies the waste zones that settle in by inadvertence, and it transforms the sobriety commitment into a demonstrable approach rather than a declarative intention.

The Swiss context: three structural assets

Switzerland presents three structural assets that make responsible use of generative AI more accessible than in the average of European geographies.

The national electricity mix is dominated by decarbonised energies — hydroelectricity mainly, complemented by nuclear. This characteristic translates into a carbon intensity per kilowatt-hour produced significantly lower than that of neighbouring countries heavily dependent on fossil fuels. For an equivalent compute workload, the carbon footprint of a processing operated on Swiss infrastructure is significantly lower than that of the same processing operated elsewhere.

The temperate climate favours passive cooling — free cooling — for a substantial part of the year. This characteristic reduces the energy consumption dedicated to the cooling of data centres, which can represent a substantial fraction of total consumption in warmer climates.

The regulatory framework structured by the Federal Act on Data Protection^[2] encourages transparency on data processing, which indirectly includes the traceability of the flows that feed generative models. This regulatory discipline, combined with the growing interest in data sovereignty, pushes Swiss companies towards local hosting solutions that by construction have a more favourable carbon footprint than solutions hosted on more carbonised infrastructure.

Data compliance: a strand of overall responsibility

Compliance with the Federal Act on Data Protection articulates with the responsible-digital approach without merging with it. Compliance obligations bear on transparency, purpose, proportionality and security of personal-data processing. Responsible-digital commitments bear on the overall environmental footprint and on the reasoned use of compute resources.

The two dimensions intersect in practice. Mapping the data flows that feed model-driven tools makes it possible both to objectify the compliance obligations and to identify the compute volumes that make the footprint. Favouring local hosting responds to both demands: it satisfies sovereignty expectations for sensitive data, and it benefits from the Swiss energy mix for the environmental footprint. Informing users about the use of models in the processing of their personal data belongs to compliance, and at the same time participates in the general transparency that characterises a mature responsible approach.

The choice between a subscription to an external publisher and a custom development hosted locally has, in this respect, a direct impact on both dimensions. The note SaaS versus custom development in Switzerland treats this strand on a larger scale.

Responsible digital as a competitive advantage

The environmental argument is no longer solely ethical. It is progressively becoming commercial. Tenders from major buyers, particularly in the public sector and in regulated industries, now integrate corporate social and environmental responsibility criteria in the evaluation of providers. A structured responsible-digital approach ceases to be an accessory differentiating argument; it becomes a condition of eligibility in certain markets.

This dynamic is observed beyond institutional B2B alone. A growing share of consumers declares that the environmental practices of a brand influence their purchasing decision, according to several converging studies. This declaration is not always followed by effect in real behaviour — the gap between declared intention and actual behaviour remains documented — but it weighs at the margin on brand perception, and it weighs more heavily in the younger customer segments.

For a Swiss company adopting tools driven by generative models, the trade-off is therefore not between operational performance and environmental responsibility. It is a trade-off between opportunistic adoption, which maximises immediate gains without qualifying the footprint, and disciplined adoption, which combines operational gains with coherence with the commitments displayed. The second approach is more demanding at initial framing. It produces an approach the company can document, present to its stakeholders, and defend against future regulatory evolution.

A discipline that does not oppose adoption

Generative artificial intelligence will continue to spread in Swiss organisations in the coming years. This diffusion is not negotiable by an individual or sector decision. What is negotiable, by contrast, is the quality of the diffusion: the intentionality of the initial framing, the choice of models and providers, the measurement of the footprint produced, the integration of the approach into the organisation's overall reporting.

This discipline does not slow down adoption. It consolidates it. A company that frames its use of generative models on the five levers identified above integrates the technology more durably than a company that piles up subscriptions and uses without qualification, exposing itself to the bad surprises — pricing, regulatory and reputational — that the rapid evolution of the market will continue to produce.

Switzerland has an energy mix, a regulatory framework and an economic culture that make this discipline more accessible than elsewhere. This accessibility is an advantage. It deserves to be seized.

Sources

[1] International Energy Agency (IEA), Energy and AI — Energy demand from AI. www.iea.org/reports/energy-and-ai/energy-demand-from-ai [↩]

[2] Federal Act on Data Protection (FADP), revision of 25 September 2020, in force since 1 September 2023. www.fedlex.admin.ch/eli/cc/2022/491/en [↩]

Jérôme Deshaie is CEO of MCVA Consulting SA, a Swiss firm specialising in strategic consulting on artificial intelligence, based in Valais.