It's the cost of all those pesky humans, experts and programmers, that are curating and tuning the datasets. What they need to cut costs is replace them with AI.
After all it's working for the Japanese purchasers of IBM kit?
I'd love to see a break down of HW costs vs ongoing "development costs" for each Watson sale. Or are Watson systems like a sale of a Adobe Creative Cloud or Oracle DB*, a trojan to ensure annual income from support, so it's losing money as it's still at the infiltrate stage (like inkjet printers in supermarkets which have to have only partially filled cartridges as they are so cheap).
(*Though those have an almost zero additional cost per sale unlike Watson sales.)