The whole world is betting on artificial intelligence. Companies build data centers, states fund chips, and the four big tech companies invest about 725 billion dollars in AI infrastructure this year between them. Anyone deciding about AI today should know what holds up and what is hype. The AI Economy report by Maxwel Consulting pursues that question in three parts: the physical side, the models, and the application in business. Part 1 asks what the boom physically depends on and follows the chain from demand through chips and raw materials to energy.

The chain starts with demand, and demand is measured in tokens, the word pieces in which language models read and write. Google’s token volume today is 330 times what it was two years ago, ChatGPT has more than a billion users per week. Cheaper AI does not lead to less compute but to more. When DeepSeek cut its prices by ninety percent, volume rose almost three hundred percent on the first day. On top of that, the models themselves are getting hungrier. Reasoning models, which work through intermediate steps before answering, need about eight times as many tokens as a simple chat; agents, which complete a task on their own across many steps, up to a thousand times. There is no end to demand in sight.

The chips come from a chain of quasi-monopolies. ASML in the Netherlands builds the only lithography machines for the finest chip structures, the optics come from Zeiss, the laser from Trumpf, and the chips are made at TSMC in Taiwan. The United States has many chips and little electricity, China much electricity and few chips. Europe holds three top positions in this chain and still only reaches 4.8 percent in AI compute itself. What is scarce is not the compute chips themselves but the fast memory right next to the chip, the packaging that joins the two, and the fabrication slots.

More than sixty raw materials go into a chip. High-purity silicon comes three quarters from two producers, the world market leader being Wacker in Germany. China processes ninety percent of rare earths, and copper is heading for a deficit of six million tons by 2035. That is expensive and politically sensitive, but not a bottleneck.

The bottleneck is energy, and it is shifting right now from training to operation. Training the models was the big build-out; electricity is now consumed by operation, the answering of queries. By 2030, according to the report’s estimates, operation needs 93 gigawatts of capacity and training another 62, together about 150 large power plants. A single query today needs 33 times less electricity than a year ago; total consumption rises anyway. Gas turbines are sold out until about 2031, large transformers have lead times of more than five years, and in Ireland 22 percent of electricity already goes to data centers. The industry is routing around the grid, with its own turbines, with nuclear contracts, and with solar fields plus batteries, for which there is no queue at the grid connection.

The industry is now looking for the solution off the planet. In the right orbit a satellite is almost always in sunlight, a solar panel delivers eight times the energy it does on the ground, and cooling works by radiating heat into space. The first high-performance GPU has been orbiting since November 2025 and has trained a language model there, SpaceX presented a data-center satellite in June 2026, Google and Nvidia plan prototypes for 2027. China is in it with hardware in orbit: the first twelve compute satellites of the Three-Body constellation have been flying since May 2025, and a constellation of 2,800 compute satellites is planned. Operating in orbit still costs four times as much, the chips are not radiation-hardened, and at scale it will only carry in the 2030s. But with these projects and the money behind them, the direction is set, the AI economy is serious, all the way into orbit, and its limit is electricity.

Part 2 on the models follows shortly.