Tim DaviesCEO Projekt Rising - The expert AI partner for SMEs

The exponential growth of AI processing power and its implications for the future

2 May 2025

Watch on YouTube

This is a short transcript excerpt from a May 2025 YouTube talk by Tim Davies discussing increases in AI model inference speed and falling compute costs. Davies mentions a 70-billion-parameter model running on Cerebras hardware at ~1,600 tokens/second, notes Cerebras is a dedicated inference chipset, and references Ray Kurzweil’s logarithmic computations-per-second-per-dollar trend while raising power/cost tradeoffs.

But now, DeepSeekar1, the 70 billion model is running on cerebras and outputting 1600 tokens a second. Right. I don't know what Cerebras is. It's a specific inference chipset. So it's compute, designed just for processing LLM output. Okay? So that rate of thinking is increasing. And while that's probably not quick enough to play a game against an experienced gamer, the impact it's making in terms of playing these basic games is it's getting faster and faster all the time. And the cost of that COMPUTE is coming down. So Ray Kurzweil has got this graph where he shows from 37 to now, this just logarithmic improvement in computations per second, per dollar. And is that cost as in the cost of the chips, or is it cost of the. The power? Because there's a power consumption. There is a power consumption thing, but it's a combination.