But the logic you use seems to imply a long life for the asset. In fact, Internet gear gets used up so quickly (as bandwidth demand rises) that in fact it typically has a life-in-place of two or three years. So, while it's charming to assert that the variable costs are low, it's also irrelevant.
Put it another way. To build an infrastructure that serves several million households and businesses will cost several billions of dollars. To make that network still useful in, say, three years' time, the operator has to again spend billions of dollars.
So the most useful way to compute the effective variable cost is NOT to assert "it's low" but to actually divide the entire cost by the traffic throughput (current peak offered load). A few years' back, doing the calculation this way suggests a variable cost for transmitting a DVD-quality movie to be about $2, and for transmitting an MP3 song about $0.05. I am sure these estimates are off now, perhaps by a factor of 10. But not by a factor of 100.
Also: the Comcast figures are obviously nonsense. If the variable cost to CMCSA was so low, they'd have deployed it everywhere. In fact, the figure the Comcast guy cites is the cost to upgrade the shared headend. Unfortunately, they also have to upgrade the taps ($100 plus labor) and the in-home terminal ($100 plus labor). Often, also the coax into the home needs to be replaced. And the hub near the home.