I'll quote from the documentation link I referenced:

> For example, the following zips together two lists:

  [ (x, y) | x <- xs | y <- ys ]

That's precisely the difference between a normal list comprehension (one pipe) and a parallel list comprehension (multiple pipes).

For clarity, here's your normal list comprehension (with one pipe) that produces all the combinations instead:

  [ (x, y) | x <- xs, y <- ys ]

And here's the full example from the article converted to Haskell and producing the exact same output:

  {-# LANGUAGE ParallelListComp #-}

  import Control.Monad (mapM)

  elems = [ "water", "earth", "fire", "air" ]
  nats = [ "tribes", "kingdom", "nation", "nomads" ]

  main = mapM putStrLn
      [ show idx ++ " - " ++ e ++ " " ++ n
          | e <- elems
          | n <- nats
          | idx <- [0..]
      ]

EDIT: I suppose an explicit zip with an anonymous function looks more idiomatic though:

  main = forM (zip3 elems nats [0..]) $ \(e, n, idx) ->
      putStrLn (show idx ++ " - " ++ e ++ " " ++ n)

EDIT2: Best of both worlds with the list monad?

  main = mapM putStrLn $ do
      (e, n, idx) <- zip3 elems nats [0..]
      [ show idx ++ " - " ++ e ++ " " ++ n ]