A former $AMZN employee who worked in data center planning shares his insights o

A former $AMZN employee who worked in data center planning shares his insights on AI data centers.

- He thinks that with AI workloads, data centers can actually be smaller with higher power densities. However, cooling should be considered. He also believes that training workloads will lead to more data centers in nontraditional and affordable locations, rather than building data centers in areas available to meet existing customer needs.

- Training workloads give up some of the reliability, elasticity and latency required for customer-facing applications. 99.9% uptime might be enough instead of 99.999%. Maybe you don't need diesel backup generators, etc.

- Liquid cooling can be a very interesting topic, especially if it is constrained by power at the utility level, the low PUE immersion design can maximize the utilization of that power.

- In his view, finding the right combination of power-backed land is a challenge and will continue to be so. The problem with renewable energy is that it is less reliable 24/7.

He thinks that small nuclear solutions combined with data centers can be an interesting solution. There are also synergies such as location security.

————————————————

Why I bought SMCI and Virtiv.