OK. So this hasn't changed.
I really gave the cloud a try, but it is soooo sloooow for many calls, little computation scenarios. The test I did, in my application environment, was simply x^2, and it was probably half a second per call, which is not bad at all, considering that just an intercontinental ping takes at least 0.1 seconds, and then there's the get in line, etc. But not enough if we want to do 1000 calls of x^2, and there's no algorithmic way of knowing it in advance, so to ask for all computations in a single call. This 1000 computations take almost 10 minutes to complete, for something that locally takes 1/10 000th of the ping time itself.
So, I definitely have to go through a local computation, but the cost break point between programming it in a lower level environment, like .net, vs the cost of multiple full Mathematica licenses, doesn't compensate going the Mathematica way. I only which we could have an equivalent to WR competitors' MCR.
Thank you anyway for the prompt response.