AI and Compute

When Dario Amodei and Danny Hernandez published this analysis in 2018, it felt like a collective wake-up call for everyone in the software industry. The stat is burned into my mind: the compute used in the largest AI training runs has been doubling every 3.4 months. That is not just slightly faster than Moore’s Law—it is an entirely different order of magnitude, a vertical wall of exponential growth. It explains why the modern AI boom feels so capital-intensive and why startups are suddenly acting more like resource-allocation and logistics companies than pure software outfits.

For an engineering founder, this trajectory is both exhilarating and terrifying. It means that the capabilities of the models we use are expanding at a breakneck pace, but it also means that the cost of training or even fine-tuning these models is ballooning beyond the reach of normal startups. We have to be incredibly strategic about where we sit in this value chain. If we try to compete on raw compute, we lose by default to the tech giants. Our opportunity lies in efficiency—finding ways to prune, distill, and optimize these models so they can run cost-effectively within our constraints, turning their brute-force breakthroughs into practical, lean products.

What stuck with me

• Hyper-exponential growth: The 3.4-month doubling rate of AI training compute far outstrips the traditional 18-to-24-month cycle of Moore's Law.
• Capital as a moat: The sheer cost of acquiring the compute necessary for state-of-the-art training has made hardware access a major barrier to entry.
• The efficiency frontier: Since we cannot out-spend the hyperscalers on raw silicon, our startup's engineering focus must shift heavily toward model optimization and distillation.