From 50 Megawatts to 1 Gigawatt: AI Data Centers Scale 20x in Six Years

In 2020, a “big” data center used 50 megawatts. By 2026, the new giants will each consume 1,000+ megawatts. That’s a 20x scale-up in just six years.

This isn’t incremental growth—it’s an architectural revolution.

Traditional data centers ran thousands of independent workloads. Today’s AI training clusters demand synchronized power across racks of GPUs working in concert on single model training runs. The infrastructure — as explored in the economics of AI compute infrastructure — requirements are fundamentally different.

What does 1 gigawatt look like?

• 1 nuclear reactor’s output
• 750,000+ homes powered
• 100,000+ GPUs training simultaneously
• $29 billion in investment

Big Tech hyperscalers—Amazon, Google, Meta — as explored in the interface layer wars reshaping consumer tech — , and Microsoft—have committed over $400 billion in capital expenditure, mainly on data centers. The spending has triggered fears of an AI bubble, but these companies argue demand still outstrips supply.

The constraint isn’t whether AI models can scale. It’s whether the power grid can keep up.

For the complete analysis of how AI is reshaping data center infrastructure, read The State of AI Data Centers on The Business Engineer.