The demand variability of AI is dramatic - nobody knows the requirements of tomorrow as AI/ML is still fundamentally a research project - and this aspect of AI development took up a lot of talk time.
The pace of innovation is startling, and opens up a lot of challenges related to technical obsolecence. Chips now have a shelflife as short as 12 months, and data centre designs themselves are also in danger of not meeting client needs even before the concrete has set.
That said, it's important not to get carried away. Companies don't change their architecture overnight, and database structures and email servers are not going anywhere. This will probably mean hybrid architecture between traditional services and HPC workloads.