"just" is always a weird modifier in these kinds of conversations. Most things can be reduced to "just" some basic operation(s). What is amazing is the complexity that emerges from those simple operations at scale.
- Distributed computing
- Feature selection
- Budgeting / accounting of experimental design (these TPU clusters are not cheap)
- ML architecture
- Involvement of domain experts (multidisciplinary teams)
- Storage
The whole list is much longer, but just some topics to think about. For me the term "AI", although overly broad, come to mean the engineering and organisation needed to pull these kind of projects.