We all know CPU performance is much less important than IO performance (file, DB, network), but assuming a large Ruby site and large .NET site both got that part right, and CPU performance does start to matter, what he says sounds reasonable.
Rather than being noticeably faster (lower latency) for a single user, the compiled tier (Java,C#,Go,Haskell,Scala,etc)'s better performance often just allows for far fewer (or cheaper) servers than the interpreted tier (PHP,Perl,Ruby,Python) to be thrown at the problem.
Just to elaborate further on why I think this is a throwaway comment from Atwood rather than getting into the whole "one language is faster than another" dick-measuring competition:
"Fast" in the context of the Internet is incredibly subjective (especially when you factor in global network latency), and has no place in technical discussion unless backed up with benchmarks (and even then, to be taken with a pinch of salt). For example, think about these made-up statements:
1. "Our site is really fast."
2. "That site is faster than the other site."
3. "We made our site 10% faster."
4. "We build our site using language X because it's fast."
I'm sorry, but statements like that raise my bullshit firewall. It's bordering on marketing speak.
Wow, that claim could really do with some qualifying.