IBM's Mainframe Bet: The Solution No One Chose in a 48-Hour Hackathon

Last week at an enterprise AI Hackathon in Toronto, my teammate and I were arguing about inference infrastructure selection at 3 AM. There was a former IBM engineer in our group who insisted on using Watson inference services on z/OS, saying it was "stable." We glanced at the API docs—we couldn't even get the environment running after four hours. In the end, we knocked out the demo with three lines of FastAPI plus Llama.cpp and won the prize. IBM mainframes? That stuff can't even run an MVP within a 48-hour lifecycle.