Demand appears unlimited, companies are building data centers, and four US corporations are investing around 725 billion dollars in 2026, yet the power supply is becoming the bottleneck. New models impress with speed and intelligence, yet once the subsidy recedes, price becomes the criterion. The applications, finally, range from research through translation to a company memory and beyond, yet the industry is so young that finished solutions exist for only a few areas, and the best ones often emerge tailor-made.

Physics of AI

It starts with demand, measured in tokens, the word fragments of language models. At Google, their volume in May 2026 was around 330 times higher than two years earlier. Cheaper AI does not lead to less computing work but to more, something economists have known since the coal of the 19th century as the Jevons paradox. And the models are getting hungrier: a reasoning text before the answer takes around eight times as many word fragments as a simple chat, an agent across many steps a multiple of that. Even the largest providers ration access; no end to demand is in sight.

This demand is met by a chain of quasi-monopolies. ASML in the Netherlands builds the only lithography machines, the optics come from Zeiss, the laser from Trumpf, fabrication happens at TSMC. Europe thus holds one link of the chain irreplaceably, yet reaches only a single-digit share of effective computing power. What is scarce is not the chips alone but above all the fast memory next to the chip, the joining of the two, and the fabrication slots. Because manufacturers are shifting capacity there, memory prices are rising instead of falling as usual, and the shortage reaches every computer buyer.

Beneath that lie the raw materials, over 60 elements of the periodic table per chip. Three quarters of the ultrapure silicon comes from just two producers, one of them Wacker in Germany, Europe’s second irreplaceable link. China processes the overwhelming share of rare earths, analysts expect a growing copper deficit, and with their water demand the fabs compete for local drinking water. All of that is expensive and politically delicate; a bottleneck it is not, so far.

The real limit is electricity. Running the models, that is answering the requests, will need around 93 gigawatts by 2030 and is growing faster than training at around 62 gigawatts. Together, that equals about 150 large power plants, most of which have yet to come online. The bottleneck is physical, not financial: turbines and transformers have multi-year lead times, grid connections queue for years, several regions are suspending new hookups. The industry’s answer: gas bridges, nuclear takes over, solar flanks.

The structural answer lies where electricity is not scarce, in orbit. In an orbit along the day-night line the sun never sets, a panel delivers eight times as much as on the ground, and the waste heat radiates into space without a liter of water. The first computing satellites are already flying, yet the matter is disputed, costs run at four times ground level, and critics see limits in data transmission.

Models of AI

On this computing power the models are built. New versions appear weekly and differ in speed, depth of reasoning, and price by orders of magnitude, from 14 cents to 50 dollars per million word fragments. The advances no longer come from a single recipe: classical pretraining on ever more text is hitting its limits. Upbringing is handled by human feedback, schooling by reinforcement learning with verifiable rewards, meaning tasks whose results can be checked; on top comes more computing time for reasoning. The quiet engine of the price decline is distillation, in which large teacher models train small students.

The most important figure for a company is still not capability but price. The monthly flat rate of the vendor apps hides it, because it is a subsidy. Anyone billing the same work through the interface, compliantly with a data processing agreement, pays a multiple, and anyone requiring processing in Europe pays more again. Sovereignty thus becomes a line item on the invoice for the first time.

Open weights, meaning freely downloadable models, long came almost only from China, and the strongest still do. Since mid-2026 the West has been countering with open models of its own. Europe is not building a frontier model of its own but runs foreign open models sovereignly within EU jurisdiction.

The models can be run on four levels. A subscription is set up fastest but ties you to one vendor and to a foreign jurisdiction, for closed cloud models can be switched off by state decree, as first happened in 2026. An aggregator bundles hundreds of models behind a single contract, a rented server in EU jurisdiction brings full data sovereignty, and in your own house the individual call costs nothing more, though the hardware is cost-intensive. Model choice is thus not a matter of taste but an architecture decision: per task, what counts is price, capability, speed, and the place where it runs.

Work with the models happens on a ladder of delegation: chat is the first rung, the agent works through tasks with tools on its own, the swarm distributes them across several agents, and the loop orchestrates agents and swarms continuously. With each rung, benefit and cost grow, but so does complexity. Delegation also brings new risks: an agent can misunderstand instructions and lose data, hallucinate with full conviction, and one that reads outside content can be hijacked through instructions hidden inside it. Real incidents have documented this since 2025. The answer is limited permissions, fixed approval points, and traceable intermediate states that shift oversight from the person into the structure.

Application of AI

Inside a company, the path runs in a fixed order from digitization through assistance to automation. Only once processes and documents exist digitally can an assistant work, and it learns by working along. Automation is not a separate project but the maturity level a well-led assistant reaches on its own.

What already carries weight today is less spectacular than the headlines suggest. AI helps where text and knowledge make up the work, and the underrated first step is vetted access for everyone in the house, because that way each person learns the handling. Quick wins lie in reading out long documents, from tenders to contracts, and in opening up grown knowledge and image archives; describing a photo costs a fraction of a cent.

So that this does not dissipate into individual tricks, the know-how belongs in writing. A skill is exactly that, a reusable working instruction the AI draws on for matching tasks. A personal trick becomes project practice, then a department standard, and finally a rule of the house.

The evidence on benefit is thinner than the debate suggests. In every sixth AI-assisted task the work takes longer than before, and nine out of ten German companies report no measurable productivity effect so far. That retrospective, however, covers model generations that could not yet carry longer tasks, for the decisive leaps came only in 2025 and 2026. Before every pilot, therefore, belongs a record of the starting state and ongoing measurement of the relevant metrics.

Since August 2, 2026, none of this is a lawless space anymore. The transparency obligations of the AI Act apply, AI-generated content must be labeled, every company must train staff, and trade secrets do not belong in third-party apps. Where a works council exists, it has a say, because AI access through company accounts makes performance monitorable.

The most valuable entry point has little to do with all this infrastructure. It is a company memory in the company’s own house, one that makes documents searchable and answers questions with the source attached. The occasion is demographic, for with every retirement, knowledge leaves that is written down nowhere. And it gains value every year instead of aging with every model generation.

This entry is not easy, and precisely there lies the opportunity: whoever starts small and measures honestly will be further in a year than anyone waiting for the finished solution. The staff’s concerns belong on the table, too: good rules exclude performance monitoring and leave the final decision with a person. AI is neither a miracle nor a replacement for people, but a tool that amplifies whoever wields it, and wielding can be learned.