That sounds like a really bad idea!

In general, you should aim to get as much intelligence as possible out of the remote device.

To put it another way, you should always aim to push the "intelligence" as far "up" the tree (ie, away from the remote devices) as possible.

The reasons should be obvious:

You have far more remote devices than "servers"^[1], so managing them becomes becomes more complex.

precisely because they are remote, managing them, debugging them, deploying them, etc, etc, are all far more difficult than dealing with a single, central server located in some nice, clean environment.

They told me this the right way to do in a client-server system because the server can't handle complex sequences.

Rubbish.
In general, the "server" is a far more powerful system than any remote device.

Even if it were true, it says nothing about which end should be considered the "server"...

there will be potentially hundred thousands of devices located all around Europe.

Absolutely - so you want to make them as cheap and simple as possible!

This means the server must be able to handle thousands of concurrent connections

Which is why the "server" needs to be the master.
That way, the server can manage its resources, rather than always having to be prepared for anything that the devices can throw at it!

and is not possible to handle a conmplex sequence for each connection (maybe this will eat too much resources).

Nonsense - how do they suppose that major e-commerce sites cope - like Amazon.com ...?!

Of course "the server" doesn't have to be a single computer - large-scale systems will be distributed across many computers!

[1] Note that I am using "server" here in the commonest sense of being the "Big Box" in your system;
In HTTP, it could be the other way around - see: http://www.8052.com/forum/read.phtml?id=114032