OpenAI reports that the compute behind the biggest AI training runs has been increasing at an exponential rate since 2012.
The analysis reveals a 3.4‑month doubling time for compute, compared with the roughly two‑year cycle of Moore’s Law. Over the past decade this translates to a more than 300,000‑fold jump in raw processing power for flagship models—something a two‑year doubling schedule would have delivered as a modest seven‑fold rise.
Why it matters: Compute growth has been a primary driver of recent breakthroughs in language models, vision systems, and reinforcement learning. If the trend holds, future models could operate at scales that dwarf today’s capabilities, raising questions about cost, energy use, and the feasibility of training ever larger systems.
In short, the data suggest that AI’s computational hunger is accelerating far beyond traditional semiconductor trends, a fact that will shape both research agendas and the economics of AI development.