you're just incorrect. You probably missed 2 points :

- battery storage

- and in the article

"However, AI labs and some hyperscalers have relaxed those requirements as there is now a lower uptime tolerance applied to both inference and training, not just training. Many of Meta’s self-built AI datacenters, for example, target just two nines of uptime and forgo backup generators entirely, as detailed in our Industrials Model."