The Real Cost of AI

The idea that “AI is cheap” may be one of the most expensive illusions in tech today. When we look only at the token prices of frontier models, it’s easy to get the impression that AI will always be dirt cheap to run. But before you lay off half your company and go all-in on “cheap intelligence,” come hear this talk. AI inference has to run on something and someone has to pay for it.

We run bare-metal GPU servers for our customers and ourselves, including 8× H200 NVL, 4× H100, and more, with a total value in the tens of millions of CZK. In this talk, we’ll show what it actually takes to run an AI cluster at this scale, what size models it can handle, and, most importantly, what the cost-to-performance ratio and ROI look like.

Everything is based on our own data and hands-on experience. As with our other talks, expect a cocktail of enthusiasm and skepticism, plenty of strong opinions, and a few well-intentioned unsolicited pieces of advice.