I had a product that had mostly stopped growing in usage. It was running on say fifty machines. I had put considerable effort into some memory optimizations, which was the scaling point for new hardware, so I talked Ops into bunching the active traffic load onto fewer machines. All of the active traffic load. Started hitting the memory limits (32 bits linux) and our server framework exited on malloc failure, so lots of exiting of long lived processes and loss of expensive state, delayed alerts etc.

I still think it was over-provisioned, but they told Ops to stop listening to me unless someone else agreed. Probably ran on the 50 machines till it was discontinued 10 or 15 years later, but I left so who knows.