The hashgraph algorithm requires that nodes add two hashes, a timestamp, and the node's signature to the fundamental payload of business transactions within an event. It is the two hashes that allow all nodes to create a consistent representation of the history of messages - the hashgraph. The hashes are 48 bytes in size, the timestamp 12 bytes, the signature 384 bytes so the total overhead for each event is approximately 500 bytes. The payload transactions vary in size - a simple transfer of hbars is 150 bytes. If there was a single transaction within an event, then the overhead would be significant. But only in exceptional circumstances would this happen - typically events will have many transactions. The number of transactions within an event will determine the relative contribution of the overhead of the hashes and timestamp to the size of the event (and so bandwidth usage). The number of transactions within an event will itself depend on the number of transaction submitted to the network, the number of nodes, and how many events each node creates per second. As an example, for 15 nodes, a throughput of 10,000 hbar tps with each node creating 10 events per second, the overhead is approximately 0.5% = 500/(500 + 10,000/(15*10)*150).
The hashgraph is something you get almost for free, by adding a small amount of overhead to the fundamental payload of business transactions within the gossip messages. If there are no transactions at all for a few seconds, then the entire network goes quiet, and stops creating new events. So no new hashes are generated or communicated. But if there is a steady stream of transactions flowing into and through the network, then the algorithm adds only a small overhead to each of the communications that are happening anyway.