Following the trail
https://github.com/kubernetes-sigs/aws-ebs-csi-driver/issues/1139
https://github.com/kubernetes-sigs/aws-ebs-csi-driver/pull/1175
https://github.com/openshift/aws-ebs-csi-driver/pull/206
 
Looks like the fix should be in 4.12, but it still see it being 39 vs ~24 on an m6i instance type.

It seems that the kubelet applies this capacity to the node in 4.11 and earlier and, thus, unlikely to receive this fix for attachable volumes in the upstream CSI driver. 4.12 behavior is currently unknown but it seems that the kubelet might still be setting this capacity.

The actual issue is that kube scheduler schedules pods that require PVs to nodes where those PVs can not be attached.