I was invited to speak at the AI Infra Summit 2026 in Santa Clara, California, USA. Back from the AI Infra Summit, where I joined Bryan Liu, Hagay Lupesko, and David Murray on a panel about the cost of AI. My key takeaway from the panel is that optimizing one layer in isolation usually just moves the bottleneck somewhere you weren't measuring.
I also presented "What to measure and optimize when every GPU counts?", covering several approaches we've taken at Booking.com. The main lesson is that the goal is not fewer GPUs, it is the right compute for the result. Inference must be useful, reliable, and at the lowest sustainable unit cost.
It's striking, in person, how fast the lower layers of the AI stack are moving. I've noticed that it's easy to lose that sense of pace when you're heads-down on your own systems. Hoping to be back next year!
Linkedn post announcement: https://lnkd.in/p/e7JJ6FjQ
