What AI runs on, and where

Every answer from an assistant is computed somewhere. A chip in a data centre does the arithmetic, the grid delivers the power, and a cooling system carries the heat away. For most companies all of that has been invisible.

It is turning into a business decision, because it sets the price of AI, its speed and the place where your data goes.

Compute grows faster than anything around it

Stanford’s AI Index 2026 estimates that global AI computing capacity has grown 3.3 times a year since 2022, to the equivalent of about 17.1 million of Nvidia’s H100 chips. Nvidia accounts for more than 60 percent of it. Google and Amazon supply much of the rest with chips of their own, and Huawei holds a small but growing share. Almost every leading AI chip is made by one foundry, TSMC in Taiwan.

The money follows. The International Energy Agency reports that the capital spending of the largest technology companies exceeded 400 billion dollars in 2025 and is expected to rise by another 75 percent in 2026. It also notes a shortage of high bandwidth memory, a component integral to AI chip production, which it expects to last until at least the end of 2027.

Power is the new limit

Chips need electricity. The IEA counts a 17 percent rise in data centre electricity demand in 2025, and 50 percent in data centres built for AI. Its central projection has data centre consumption roughly doubling from 485 terawatt hours in 2025 to 950 in 2030, around 3 percent of global electricity demand. Consumption in data centres focused on AI triples over the same period.

The AI Index puts AI data centre power capacity at about 29.6 gigawatts at the end of 2025, comparable to the peak demand of New York state. By its estimate, the power demand of all AI systems together is comparable to the national electricity consumption of Switzerland or Austria.

Europe wants more of that capacity at home. The European Commission has set a goal of tripling data centre capacity in the EU within five to seven years, the IEA notes. On 30 July 2026 the EuroHPC Joint Undertaking opened a call for up to seven AI gigafactories in at least seven member states, with public money as an anchor customer and more than 20 billion euros of private investment expected. Proposals are due on 12 November, and the selection is expected in early 2027.

Each answer costs less

While the total grows, the single answer gets cheaper. The IEA writes that the energy used per AI task has dropped by at least an order of magnitude a year recently, and that a simple text query now typically uses less electricity than running a television for the same time.

Prices tell the same story. Epoch AI measured how fast the price of reaching a given level of performance falls: between 9 and 900 times a year, depending on the task. The fastest drops were recent, and Epoch is careful to say they may not last.

The hardware underneath has followed. The AI Index, using IEA data, finds that the cost of GPU computation has fallen by more than 99 percent since 2006, and that leading chips deliver about ten times more computation per watt than those of a decade ago.

The catch is the work itself. Video, reasoning and agents can use hundreds or thousands of times more energy per query than a simple text answer, the IEA notes. The unit gets cheaper. The bill depends on what you ask the machine to do.

Small enough for the office

The second shift is in size. A capable model no longer needs a data centre. Nvidia’s DGX Spark is a desktop machine with up to 128 GB of memory. Nvidia states that it runs models of up to 200 billion parameters, can fine tune models of up to 70 billion, and that up to four can be linked for larger ones.

Smaller models are getting better at their size. The AI Index points to an open model with 32 billion parameters, nearly 90 times fewer than Grok 4, that reaches comparable results on several benchmarks. Its makers trained it on far fewer tokens and selected the data with great care.

Switzerland builds its own. Apertus, released in September 2025 by EPFL, ETH Zurich and the Swiss National Supercomputing Centre, was trained with more than 10 million GPU hours on the Alps supercomputer, in sizes of 8 and 70 billion parameters, on 15 trillion tokens in more than 1'000 languages. Its weights, its recipes and the documentation of its training data are public. Version 1.5 followed in July 2026, with image and audio understanding and a family of compact models.

The country has the people for it. The AI Index counts 110.5 AI authors and inventors per 100'000 inhabitants in Switzerland, ahead of every other country, with Singapore close behind.

Swiss and European hosting

The large providers have moved closer. In June 2025 Microsoft announced 400 million dollars to expand its four data centres near Zurich and Geneva, with AI for regulated sectors such as healthcare, finance and administration in mind. Heise, citing Reuters, reported that the expansion should keep data within Swiss borders.

Swisscom builds on a platform of its own. In February 2026 it launched an assistant for company knowledge that is operated entirely in Switzerland, on several open source models used unchanged. All data is stored and processed in its Swiss data centres, and company data does not train the base models. The platform underneath offers GPU capacity and model interfaces to others, and Swisscom named Apertus among the models it plans to add.

Choosing where your AI runs

For a small or mid sized company, this comes down to three options. A frontier model rented through an interface, with enterprise terms and a European region. An open model hosted in a Swiss or European cloud. Or a model on hardware of its own, in its own building or its own cloud account.

We decide with four things in mind: how sensitive the data is; how much work runs through the model, because a fixed cost only beats a metered one above a certain volume; how fast an answer needs to be; and who will look after the system once it runs.

We use more than one of these ourselves. Twin works for us in Claude and ChatGPT, and on our own hardware with a model of our own. Studio holds its data in Zurich. For clients, Twin runs in Switzerland or the EU, or on the client’s own hardware.

Eigna, built by our partner Eigna Labs, sits at the private end. It runs in a company’s own cloud or on its own servers, from one machine to a cluster, and is designed to run fully offline in closed networks. Corrections retrain it on the company’s own hardware, and a new version goes live only when it tests better.

The machine can be chosen on a spreadsheet. Which data may leave the building is a judgement, and it belongs to a person who answers for it. Start with the data. The chip, the building and the country follow from it.