He compares several metrics to come up to this conclusion, such as peak throughput, throughput/memory ratio and overall memory usage. He later explains why JIT supposed gains (constant tuning) do not always pay off in practice.
That's pretty much in opposition to what I heard for the last 10 years. His case his built on a single, likely not representative of all loads, example. But the guy has been working on a JIT java VM for the last 15 years, I guess he knows what he's saying to some extent.
I'm by no means an expert in JITs, I can't determine whether this case makes sense or not yet.
@Ninjatrappeur I think "I can do it live" is a slightly different nuance from "it can do it live", but I admit I didn't watch yet, maybe the talk addresses both?