AI’s next bottleneck is physical

The AI race now runs through the grid

The most important AI infrastructure story may not be a faster accelerator. It may be whether a new facility can get enough electricity, cooling, and equipment to operate at all.

The International Energy Agency estimates that data centers used about 415 TWh globally in 2024 and projects roughly 945 TWh by 2030. At the same time, grid connections, transformers, and transmission lines are becoming long-lead infrastructure.

Heat is part of the compute bill

Higher-density AI racks push operators toward liquid cooling and new facility designs. That can unlock more compute per square meter, but it also adds plumbing, maintenance, heat-rejection, and retrofit complexity.

The durable trend is systems thinking: the future of AI will be shaped as much by power planning and local permitting as by model benchmarks.

Questions readers ask

Why does AI need so much more data-center power?

AI workloads use dense accelerator systems that consume substantial electricity and produce significant heat, increasing both compute and cooling demand.

What is liquid cooling in an AI data center?

Liquid cooling moves heat from high-density components through cold plates or other fluid systems, allowing more compute per rack than conventional air cooling in some designs.

What is the main AI infrastructure bottleneck?

The bottleneck is a connected system of grid capacity, transformers, transmission, facility construction, cooling, permitting, and access to high-density hardware.

AI’s next bottleneck is physical

What is the key point of AI’s next bottleneck is physical?

The AI race now runs through the grid

Heat is part of the compute bill

Questions readers ask