As great as content addressed IDs are, they only allow you to build DAGs. Many domains require a graph to be properly modelled, a collision free string of random entropy is the best thing you can get, in terms of distributed-ness, for these scenarios.
There is a certain duality to unique identifiers and content addressed identifiers, both are best utilised together.
If you want a graph then you can just encode it in the data by having `n` content addresses and ca. `n^2` bits for the adjacency matrix.
The DAG of content addressess is only in the abstract when you have "unflattened" data and all you can do is chase the references to earlier content addresses. If you have some semantics for relating content addresses then you can flatten stuff into graph structures (or whatever you desire) by providing the relevant metadata (that relates the data in question).
Now you have more canonical data that can be content addressed.. This is the essence of what I am doing with datalisp.