We're currently investigating an elevated level of machine start failures. Workaround in the meantime would be to replace affected machines with a new one: fly machine clone <machineID>