We've run nearly 500 benchmark runs: provider selection matters but region does too big time.

Still early (3 runs each), so we'll keep adding depth with temporal buffering...