AI and compute
OpenAI
OpenAI found the compute used in top AI training runs has doubled every 3.4 months since 2012, way outpacing Moore's Law. That's a 300,000x jump in six years, and it means AI capability is scaling faster than almost anyone budgeted for.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI dug through the history of its biggest training runs and found a number that should make you sit up: compute used in the largest AI experiments has been doubling roughly every 3.4 months since 2012. Moore's Law, the old chip-industry heartbeat, doubled every two years. This is a different animal entirely.
Do the math and it gets stark. A two-year doubling period, run from 2012 to today, would produce about a 7x increase in compute. What actually happened was a jump of more than 300,000x. That's not a marginal acceleration — it's an entirely different growth regime, the kind of curve that turns a five-year forecast into science fiction within eighteen months.
The reason this matters isn't just nerd trivia about FLOPs and GPU clusters. Compute has been one of the main engines behind the jumps in AI performance we've actually seen — bigger models trained on more hardware doing things smaller ones couldn't. If that compute curve keeps climbing at anything close to this rate, the systems five years out won't just be incrementally better than today's. They'll be operating in a capability zone nobody has mapped yet.
OpenAI's own framing is cautious: they're not claiming this trend continues forever, and it can't, physically or economically, at this pace indefinitely. But they are flagging that if it does continue even a while longer, the industry, regulators, and honestly everyone downstream of AI systems should be thinking hard about what that implies now, not after it happens.
My take — AI-written commentary, not fact-checked reporting
This is OpenAI in 2018 basically pointing at a growth chart and saying 'this seems important, someone should probably do something,' then... continuing to build the very systems the chart warns about. I find that pattern more revealing than the stat itself. Compute scaling is real and it did keep bending capability curves for years afterward, but the lesson everyone in this industry keeps dodging is that flagging a risk isn't the same as slowing down for it.
Read more about this at: OpenAI