Cost in developer time, yes. Cost in latency and performance? No! That's the whole point of async I/O: to improve latency, performance, and scalability.
It's much harder to rewrite software to use async I/O than it is to write it to use async I/O in the first place. It's also much harder to write software using async I/O than sync I/O. Therefore it's all about economics and forecasting. If you can forecast needing to scale soon enough, then start with async I/O, otherwise increase revenue and profits first then rewrite. For many companies developer time is a larger item on their budgets than the cost of extra servers and power and all that, which is why we have so much thread-per-client software -- it's a natural outcome.
It's much harder to rewrite software to use async I/O than it is to write it to use async I/O in the first place. It's also much harder to write software using async I/O than sync I/O. Therefore it's all about economics and forecasting. If you can forecast needing to scale soon enough, then start with async I/O, otherwise increase revenue and profits first then rewrite. For many companies developer time is a larger item on their budgets than the cost of extra servers and power and all that, which is why we have so much thread-per-client software -- it's a natural outcome.