From bigger models to faster reasoning

The next battle in AI won’t be about intelligence but about speed and cost. I use agents all day: to develop applications, prepare training materials, research strategy topics, and more. The results are now consistently at the expected level.

Today, I don’t need a “better” model, just a faster one. Imagine: A 10x faster model would develop last week’s hackathon in 30 minutes instead of 4 hours. A 100x model? In 3 minutes.

The key to this scaling is twofold: first, moving from today’s large models to smaller reasoning-oriented models - Small Reasoning Models (SRM). The other innovation is shifting from general-purpose GPUs to specialized hardware. Exactly what Google is doing with Gemini 3 (trained 100% on TPU) or Groq with its LPU (Language Processing Unit).

I think next year will bring us some surprises in this direction, and this is what we need to anticipate now.