AWS: Can only open a single nodeport on the cluster

Deploying on top of AWS using aws-integrator.

I am seeing a behavior similar to described on: https://github.com/kubernetes/kubernetes/issues/39214
Where only one NodePort can exist per cluster. Seems this is due to the fact k8s does not know if ELBs continue to exist.

On my deployment, I can expose kubernetes-dashboard with:
apiVersion: v1
kind: Service
metadata:
  name: kubernetes-dashboard
  namespace: kube-system
  labels:
    k8s-app: kubernetes-dashboard
    kubernetes.io/cluster-service: "true"
    addonmanager.kubernetes.io/mode: Reconcile
spec:
  type: LoadBalancer
  selector:
    k8s-app: kubernetes-dashboard
  ports:
  - protocol: TCP
    port: 443
    targetPort: 8443

That boots a ELB and configures NodePort correctly.
I need to run open-port on kubernetes-workers as per: https://bugs.launchpad.net/charm-kubernetes-worker/+bug/1842104

Then, I try to create a second service + ELB:
apiVersion: v1
kind: Service
metadata:
  name: tensorboard
  namespace: kubeflow
  labels:
    k8s-app: tensorboard
    kubernetes.io/cluster-service: "true"
    addonmanager.kubernetes.io/mode: Reconcile
spec:
  type: LoadBalancer
  selector:
    k8s-app: tensorboard
  ports:
  - protocol: TCP
    port: 6006
    targetPort: 6006

I see the ELB booting, however k8s do not allocate NodePorts this time.
netstat -tnlp returns empty for the allocated NodePort on kubernetes-workers (first nodeport is there though)

What intrigues me is that solution described on: https://github.com/kubernetes/kubernetes/issues/39214#issuecomment-269236544
Seems not to be working for me. I am using aws-integrator and trusting it (juju trust) with same credentials as Juju's, which should be more than enough to operate ELBs.

Eventually ELB dies out and disappears from my console.