These currently have 8GB assigned, but on average use significantly less, it should be fine to reduce them 4 GB at least.
I would go down to 4G with (on ganeti1001):
sudo gnt-instance modify -B memory=4g kafkamon1001.eqiad.wmnet
Same thing for the codfw instance. From grafana it seems that we could go down even more, but since these hosts run Burrow (and we rely on it for Kafka lag metrics etc..) I'd be more cautious with the first step.