Cette publication n'existe pas encore en français. Voici la version originale.
If you read about AI, you read about size. Data centres with the power demand of a large city, chips allocated in supply-chain wars, models with ever more parameters. It is a real race, and it matters. It is just not the race an industrial company has to win.
A plant manager does not ask how many gigawatts stand behind an answer. She asks what a correct answer costs, how fast it comes and whether the data has to leave the building for it. From that angle, the future of industrial AI looks different: smaller, more specialised and closer to the machine.
Power is expensive in Germany, and that stays a factor
For non-household consumers using between 500 and 2,000 megawatt hours a year, electricity cost 18.37 euros per 100 kilowatt hours on EU average in the second half of 2025. In Germany it was 22.64 euros, the third-highest price in the EU; in Finland, 7.48 euros (Eurostat). The Federal Statistical Office explicitly names labour and energy costs that are high by international standards as one reason for the difficult situation of export-oriented industry (Destatis).
For a single company this is first a question of location, not of the price per answer. If you buy AI per call from a cloud, you pay for the power only indirectly, in the provider's price. It counts directly where data centres are built and where models run in-house. So trying to scale AI in Europe mainly through raw compute means playing a game whose rules are against you. That is not a case against AI. It is a case for building it differently.
What factory work really looks like
Look at the questions that come up in an industrial company every day. Does this complaint match a known failure pattern? Which drawing revision applies to this order? Is this measurement within the test limit? Which purchase order matches this delivery note? Is this request complete, or is the material specification missing?
These questions are narrow, frequent and repetitive. They call for precision, not wit. Nobody would hire a polymath for them, and no sensible system should ask the biggest model in the world each of them.
Take the complaint as an example. A customer sends three photos and two sentences. The task is to match the message to one of perhaps forty known failure patterns, name the affected batch and attach the matching earlier cases. For that, the model needs no world knowledge of poetry, tax law or history. It needs the forty failure patterns, the batch data and the old cases, and it has to be reliable within that narrow world.
One research team has put exactly this as a thesis: in agentic systems, language models mostly perform a small number of specialised tasks, repetitively and with little variation. For that, small language models are sufficiently powerful, better suited and necessarily more economical; large models are needed where open conversation is called for (Belcak et al.). It is a position paper, not a law of nature. But it describes the factory better than any headline about data centres.
The curves point the same way
On top of that, small models keep getting more capable. Driven largely by them, the cost of a system performing at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024. The cost of AI hardware fell by about 30 per cent a year, and energy efficiency improved by about 40 per cent a year (Stanford AI Index). And in the cases measured, the price of reaching a fixed level of performance fell between 9 and 900 times a year depending on the task, with a median of 50 times. Whether the fastest drops will last is open, the researchers say (Epoch AI).
Put simply: many models are now good enough for a clearly defined task, and the same task has so far got cheaper every year. If you define a narrow task cleanly today and the trend holds, you get the same quality tomorrow for less money and less power.
What this means for the architecture
The obvious design looks like this: a router decides what kind of question is coming in. The narrow, frequent questions go to small, specialised models. Only the open, unusual questions that really need broad knowledge go to a large model.
This has three advantages beyond power. First, response time: a small model answers faster, and on the line or at the hotline every second counts. Second, the cost per answer, which becomes plannable because the frequent cases stay cheap. Third, location: small models can run close to the data, in your own data centre or in a European cloud. So the architecture also answers one of the most common questions in the Mittelstand: where does our data go?
The bill changes too. A large model called for every small thing creates costs that grow with each new use case and are hard to predict. A fleet of small specialists makes the frequent cases cheap and saves the expensive model for the few cases where it is worth its price. That turns AI from an experiment into an operating expense you can plan.
The price of this design is discipline. Someone has to describe the tasks, collect examples and check whether the answers are right. But that work pays off anyway, whichever model answers in the end.
Europe's edge is not watts
Europe will not win the race for the biggest data centres on electricity prices. But it does not need to win that race to lead in industrial AI. The real advantage lies elsewhere: in decades of production and process knowledge, in precisely described workflows, in standards and test specifications that define exactly what a correct answer is.
These are the best conditions for small, specialised models. They don't need gigawatts. They need good data and clear tasks.
Our assessment
An assessment, not a fact: by the end of 2027, most AI calls in a typical industrial company will go to small or mid-sized specialised models. The share of spending on the biggest general model will fall, even as overall use grows.
So the question for the next investment round is not how much compute you can afford. It is: which ten questions does your plant ask a thousand times a day, and how small can the model be that answers them correctly?
Sources
- Eurostat, 08/05/2026: Non-household electricity prices in 2nd half of 2025: -3.5%
- Statistisches Bundesamt (Destatis), 15/01/2026: Bruttoinlandsprodukt 2025 für Deutschland (Statement zur Pressekonferenz)
- Belcak et al., arXiv, 02/06/2025: Small Language Models are the Future of Agentic AI
- Stanford HAI, avr. 2025: The 2025 AI Index Report
- Epoch AI, 12/03/2025: LLM inference prices have fallen rapidly but unequally across tasks
