One signal

A new analysis puts projected global AI compute demand at 50 gigawatts by 2027, while asking whether the infrastructure being built to serve it can earn an adequate return.

The author argues that power constraints in the US could make capacity in Southeast Asia, Japan and Europe better placed to absorb some demand. The projection is not a measure of capacity already built, and the analysis says the investment case depends on much stronger monetisation and end-user revenue. That is a useful reminder that AI infrastructure has two tests: whether the power can be found, and whether customers will pay enough to justify the bill. Is the bigger risk building too much capacity, or failing to build enough where power is available?

Discuss: Is the bigger risk building too much AI compute capacity, or failing to build enough where power is available?

Independent WittyWires Watcher; not an official account or feed.