A datacenter dedicated to AI training and runtime processing (inference). The top companies developing AI (OpenAI, xAI, Amazon, Google, Microsoft and Meta) have deployed thousands of servers and tens of thousands of GPUs in their datacenters to train large language models (LLMs). Spending on new construction in 2025 and 2026 is in the hundreds of billions of dollars. See
GPU.
A Lot of Computing
There are many passes required to train AI models, and once trained, the inference engine generates answers for millions of users at the same time. A lot of computing. See
inference engine and
large language model.
Which Is Why We Need 50 Times More Energy
A server rack in a datacenter uses five to 10 kilowatts (kW), but AI racks use up to 100 kW. Huge datacenters require from megawatts to gigawatts of energy (see
Project Colossus,
Hyperion datacenter and
Project Prometheus). See
rack mounted.
Estimates say by 2030 the energy required will be fifty times higher than 2025 because of the hundreds of new datacenters worldwide. In the U.S., Wyoming, Indiana, Iowa, Texas, Oregon and Washington State are locations. Because of its reliable power supply, Northern Virginia is popular.
Cooling Takes a Big Chunk
Depending on the age of equipment, cooling can use up to 40% of the electricity. More modern facilities use only 10%, but the cooler the climate the better. See
datacenter cooling.
The Network Is a Huge Factor
In AI processing, all servers are generally connected by optical fibers for data and metal wires for control.
Potential Electricity Source
If the grid is not sufficient, power has to be generated on site. Small nuclear reactors are a fast-growing industry (see
small modular reactor), and some hope nuclear fusion will save the day by 2030 (unlimited power source). See
nuclear fusion.
Use Less Energy to Start With
Several companies attempt to lower energy use with efficient hardware. Startups such as Normal Computing claim its chip will reduce power for image generation by 1,000% (see
Normal Computing).
Forget Earth!
To solve the energy requirement, Google plans to build AI datacenters in orbit. See
Project Suncatcher.