The environmental cost of AI depends on your position in the value chain

Sabrina Ramonov recently shared data on TikTok regarding the environmental footprint of Mistral AI's large language models. According to her report, a single LLM query produces incremental CO2 emissions equivalent to watching 10 seconds of online streaming in the US or 55 seconds in France. She suggests that using AI for learning instead of watching Netflix could offset these emissions. We believe these figures provide necessary context, but comparing a single prompt to a video stream ignores the massive energy requirements of training these models.

Why it matters

Business owners often hesitate to integrate AI because of vague warnings about energy consumption. These specific metrics help lower that barrier to entry. If a single prompt costs as much as a few seconds of video, the operational footprint of an AI-driven tool is manageable. You can scale your automation without fundamentally altering your company's carbon profile.

However, you must distinguish between inference and training. The data provided by Mistral AI focuses on the energy used when a user asks a question. This is the cost of running the model, not the cost of building it. Training a massive model requires immense power and water for cooling data centres. While the daily use of AI is efficient, the initial investment in the technology remains resource-heavy. We advise looking at the total lifecycle of your AI implementation rather than just the cost per query.

How it actually works

When you send a prompt, a sequence of computational events occurs to generate a response. This process relies on pre-existing models that have already undergone their heavy training phase. The environmental cost discussed by Sabrina Ramonov on TikTok refers to this active stage, where the model processes your specific input. The energy used during this stage is relatively low because the heavy lifting of learning patterns happened months earlier.

1. Input processingYour text is converted into numerical data that a machine can understand.
2. InferenceThe model runs these numbers through its neural network to predict the next most likely words.
3. Output deliveryThe system converts those predictions back into text and sends them to your screen via the internet.
4. Hardware operationServers in a data centre consume electricity to perform these calculations and keep the hardware cool.

What it changes in practice

For most small and mid-sized Swiss companies, these figures change very little about your Monday morning operations. If you are simply using existing tools like ChatGPT or Mistral to draft emails or summarise documents, your carbon footprint remains stable. You are consuming a service, not building the infrastructure. However, the perspective shifts if you move from being a user to a developer. The implications depend entirely on where you sit in the AI value chain.

• If you only use third-party APIs, your environmental impact is tied to their efficiency and your usage volume.
• If you decide to train a custom model on your own hardware, your energy and water consumption will spike significantly.
• Companies building proprietary models must account for the massive energy required during the initial training phase.
• For those integrating existing models into software, the focus should remain on optimising prompt efficiency to reduce unnecessary compute cycles.

Where it goes wrong

Misunderstanding the environmental cost leads to poor technical decisions. Many leaders focus on the wrong metrics when planning their AI integration. They either over-correct by avoiding useful tools or ignore the real resource drain caused by inefficient implementation. The true cost is rarely the single prompt you send, but the way you architect your entire system.

• Building custom models from scratch when an API would suffice.
• Running large-scale inference on hardware that is not optimised for the specific task.
• Creating infinite loops of automated agents that query models without a clear objective.
• Storing massive, redundant datasets that require constant, energy-intensive maintenance.

Who should act, and who should wait

You should act now if your business model relies on proprietary data and custom model fine-tuning. Building your own AI infrastructure requires a deliberate strategy to manage energy consumption and hardware lifecycle. If you plan to host large models on your own servers, you must account for the cooling and electricity requirements immediately. This is not a task for a casual experiment.

Leave it alone if you are simply integrating existing software into your workflow. Using a tool to automate a repetitive task is a matter of efficiency, not an environmental crisis. You do not need to optimise your carbon footprint if you are not the one running the data centre.

What we would do

We do not aim for zero emissions by default, as that often compromises model performance. Instead, we focus on the ratio between computational cost and the actual value generated. If a query requires a massive model to perform a task a smaller, specialised one could handle, we consider that a failure in design. Our approach is to select the smallest architecture that meets your specific accuracy requirements.

1. Model selectionWe choose the smallest parameter count possible for your specific use case to reduce unnecessary compute.
2. Inference optimisationWe apply quantization techniques to reduce the memory and energy required for every single prompt.
3. Hardware mappingWe match your workload to data centres that utilise renewable energy sources to lower the carbon intensity of your operations.

The industry lacks a standard metric to calculate the precise carbon cost of a specific API call versus local hosting. This remains unanswered because providers keep their compute data private and hardware efficiency changes every month.