#kubernetes
3 posts tagged “kubernetes”.
-
What changes when AI work is throughput rather than conversation — the parser is the swappable seam instead of the model, queues replace serving, and residency gets enforced by queue topology rather than application code.
-
Hybrid AI platforms fail at the network boundary, not the model layer — outbound-only architecture, CIDR planning, choosing between Cilium Cluster Mesh, Istio and no mesh at all, and portable identity across the edge.
-
What you actually depend on if an AI platform has to run on two clouds and in your own datacenter — the eight rules, what portability costs, and why the gateway comes first and Kubernetes second.