Software has a somewhat inverse relationship to scale as manufacturing. For manufacturing the first one costs millions, and each one after costs hundreds for a time. As you get better you winnow away the equipment or maintenance costs and prices drop.

Software use cases experience combinatorics, and almost all useful algorithms have log(n) runtime. Even when Knuth says they are O(1), physics or EE say he’s wrong. There are no economies of scale. Racks don’t get cheaper when you run out of network ports. Cooling doesn’t get cheaper when you run out of roof. Things that failed one time in a million calls now happen every hour instead of twice a month, and actually have to be fixed.

It’s death by a million cuts.