AI training compute has grown more than 300,000x since 2012, a pace that leaves Moore's Law looking leisurely by comparison.
OpenAI released an analysis tracking compute usage in the largest AI training runs from 2012 onward. The headline number is a doubling time of 3.4 months, against the 2-year doubling period associated with Moore's Law. At Moore's Law rates, compute would have grown roughly 7x over the same stretch. The actual figure is more than 300,000x. That gap between 7x and 300,000x captures how different AI scaling has been from ordinary semiconductor progress — not a faster version of the same curve, but a fundamentally different one.
If the trend continues at anything close to this rate, near-future AI systems will belong to a different capability class than today's. OpenAI frames this as a reason to "prepare for the implications" of systems "far outside today's capabilities" — careful language that covers everything from economic disruption to safety concerns to the simple fact that nobody has a reliable model for what happens when training compute keeps compounding.
It is worth noting that OpenAI is one of the largest consumers of AI compute in the world, and a narrative where more compute reliably equals more capability is very much in its interest. Whether the exponential holds or runs into hard physical, economic, or regulatory limits is a question this analysis raises and declines to answer.