Let's assume that beside the need of URIs, to be linked data you need a way to decompose it to triples, well, to map it to the RDF abstract syntax (you don't actually need a particular serialisation to manifest somewhere in the process).
Then I would say that if the transformation to RDF is not explicitly defined, it is not linked data, as you need to "sniff" the content and "know" how to transform/extract it.
But, once somebody puts a GRDDL "hint" in the Atom namespace document, all the atom feeds would become automatically linked data, right?
To answer the question, a particular Atom document is not linked data if there is explicitly defined mapping/transformation to RDF (in the document itself or in the namespace document); but if Atom becomes an agreed/standardised (controlled?) serialisation, then it would became linked data as the mapping to the RDF's abstract syntax
would be "known".
Then I would push your question to even web pages with microformats (a html page with RDFa is linked data already, I hope) — if you have the proper html @profile defined, that means that you can follow links towards how to GRDDL the data? is it linked-data?
I would say yes. As long that you can follow an explicitly defined trail towards the
data embedded there (via its extraction service).
Then what is not linked data? well, when you have to guess it, to use heuristic methods to "extract" it, that would not be linked data.
Twilight zone: a web page with a link rel="alternate" type="application/rdf+xml" towards its equivalent data, is it linked data? I would say not quite, as it points to a different document and since they were separated at birth, they may evolve independently.