QQuestion: Why does AI need so many chips and so much electricity?

Training an AI means doing an enormous number of simple multiply-and-add calculations. The chip that is good at doing lots of them at once is the GPU. The AlexNet paper, presented in December 2012, established the combination of GPUs, large amounts of data and deep neural networks, and it is usually taken as the starting point of today's deep learning era¹.

Then came the finding that bigger means better. The scaling laws study from January 2020 showed that performance improves predictably as model size, data and compute grow, and it became the rationale for spending big to build bigger². In March 2022, the Chinchilla study, drawing on over 400 models, concluded that model size and training data should grow in equal proportion³. The bigger the model and the more it trains, the more chips it needs.

More chips need more electricity. Colossus, which xAI brought online in September 2024, had 100,000 H100 GPUs installed in 122 days, but its first grid connection was only 8 MW, so it had to rely on temporary gas turbines⁴. The International Energy Agency (IEA) reported that data center electricity demand grew 17% in 2025 and projected that AI-specific demand will triple by 2030⁵.

That's why power and cooling now matter as much as chips. In August 2024, Epoch AI laid out the order of the limits on scaling AI through 2030: power binds first, then chip production capacity⁶. In September 2026, NVIDIA started a program that certifies even data center batteries and cooling units — a sign that the bottleneck is moving from chips to power and cooling⁷.