On May 16, 2018, OpenAI published “AI and Compute,” a short analysis by Dario Amodei and Danny Hernandez. Its central finding was striking: since 2012, the amount of compute used in the largest AI training runs had been growing exponentially with a doubling time of 3.4 months, an increase of more than 300,000 times over that period. (Because OpenAI’s own site blocks automated fetching, this is corroborated through search and an independent archival summary; the canonical URL is openai.com/index/ai-and-compute/, verified June 6, 2026.) For comparison, Moore’s Law had a doubling period of about two years. The piece measured training runs in petaflop/s-days, a unit equal to performing roughly ten to the twentieth operations.
The point of the analysis was not the number itself but what it implied: the recent leaps in AI capability had been driven substantially by throwing more computation at the problem, not only by better algorithms. This empirical framing dovetailed with the “scaling laws” research that would follow and with Richard Sutton’s “Bitter Lesson,” which argued that general methods leveraging computation tend to win out over hand-crafted approaches.
A later, broader study by Epoch AI, “Compute Trends Across Three Eras of Machine Learning” (Jaime Sevilla and colleagues, 2022), put the trend in longer perspective using openly licensed data. Epoch identified three regimes: a pre-deep-learning era where training compute roughly tracked Moore’s Law with a 20-month doubling time, a deep learning era from around 2010 with a doubling time of about 6 months, and a large-scale era from roughly 2015 onward. Their data showed training compute had grown by a factor of ten billion since 2010, far outpacing hardware improvements alone.
Why business readers should care: this work reframed compute as the central, measurable input to AI progress. It explains why companies now spend billions on chips and data centers, why access to compute has become a strategic moat, and why the cost of frontier AI has climbed so steeply year over year.