Training AI models now requires massive amounts of internet bandwidth — here's why your data bills are about to matter more

A detailed analysis shows that the infrastructure cost to build large AI models (LLMs — systems like ChatGPT that understand and generate text) is dominated by the cost of moving data across the internet, not by computing power or electricity as previously assumed.

As AI models grow larger and training takes longer, the volume of data that must flow through internet pipes increases exponentially — meaning bandwidth costs now rival or exceed the cost of the GPUs and chips doing the actual computation work.

For companies building AI systems — from startups to tech giants — this shifts where money gets spent: cheaper compute in one region becomes pointless if you can't affordably move terabytes of training data there. This makes data center location, internet connectivity, and ISP contracts as strategically important as chip procurement.