The complex solution to this type of problem as alluded to is to implement a hierarchy within the structure of the pathfinding so that A* doesn't search through that many nodes. If you take the globe and partition it into a lower resolution polar grid, and then associate each of the high detail nodes into those 'buckets', you can create a hierarchy for the pathfinder that will dramatically reduce the complexity of calculating the path.
As an alternative, depending on your requirements, for that many nodes, it would only be about 25MB to store a lookup table for pathfinding, if the pathfinding requirements don't need to include dynamic weighting. If weighting needs to change dynamically it gets complex with having to update the lookup table.
As another alternative. If the goal node changes infrequently, or especially if there are many units moving to shared destinations, it might be faster and easier to implement if you flood fill from the destination node to generate an influence map that the ships can simply follow the decreasing weighted edges until they reach the destination. I've experimented with good results of essentially pathfinding on influence maps like this by flooding weight values. Since the flood isn't so branch and condition heavy as a full path query, building and updating the layers are actually very fast and cache friendly, and once you have the influence map, pathfinding is just a lookup and a follow into the best weighted neighbor.