In a technical blog post dated September 3, 2026, NVIDIA engineers Bhagat Khemchandani and Rohan Somvanshi describe a central identity gateway pattern that propagates user identity across federated Kubernetes clusters and AI platforms. The architecture separates session ownership from request enforcement, allowing regional gateways to validate user identity through a shared session store instead of managing independent OIDC flows.

NVIDIA says this approach cut repeated login events by 55% across its internal developer platforms spanning Kubernetes clusters in AWS and OCI, while enabling unified platform shells and AI assistants with delegated user identity. The pattern uses standard OpenID Connect, a minimal /gateway/userinfo validation endpoint, stateless regional gateways, and a shared Redis-backed session store with explicit TTLs.

Confirmed

  • The post describes a central identity gateway pattern that separates session creation and validation from per-cluster enforcement: regional gateways validate sessions via a lightweight /gateway/userinfo side-call, with sessions keyed by an opaque session ID, stored in a shared Redis store with explicit TTLs, and scoped to the platform domain through an HTTP-only cookie.
  • The pattern is built on standard OpenID Connect: the central gateway runs the OIDC authorization code flow, refreshes access tokens from stored refresh tokens, and makes logout immediate and platform-wide by deleting the shared session record.
  • NVIDIA reports a 55% reduction in repeated login events across its internal developer platforms spanning Kubernetes clusters in AWS and OCI — a vendor-reported result, not an independently verified outcome.
  • The post references reusable building blocks: an OAuth2 Proxy local environment for OIDC login and cookie handling, an Istio external authorization sample, an OPA Envoy Istio example, and Authorino for JWT and API-key validation.

What's new

The post contrasts two identity patterns for federated platforms. In distributed session ownership, each service gateway runs its own login flow, session store, token refresh logic, and logout behavior. That model keeps clusters independent but forces users to authenticate per tool and leaves logout, refresh, and audit records inconsistent across the platform.

The centralized session ownership pattern moves session creation, validation, and lifecycle management to a dedicated identity gateway. Regional gateways remain in place to enforce per-cluster policy and inject identity headers, but they delegate session validation to the central gateway via a lightweight /gateway/userinfo side-call. The session is keyed by an opaque session ID and stored in Redis with a defined time-to-live, scoped to the platform domain through an HTTP-only cookie.

Three primary flows define the pattern:

  • Login: A user without a valid platform session is redirected to the central identity gateway, which runs the OIDC authorization code flow, exchanges the code server-side, stores the session in Redis, and sets the platform cookie.
  • Per-request validation: On subsequent requests, the regional gateway sends the session cookie to /gateway/userinfo. The central gateway performs a session lookup and returns identity claims (user ID, email, groups, roles, session metadata). The regional gateway injects standardized identity headers before forwarding the request downstream.
  • Token refresh and logout: The central gateway refreshes access tokens using stored refresh tokens and updates the shared session record. Logout deletes the session record, making revocation immediate and platform-wide.

Why it matters

AI and data platforms increasingly run workloads in regional clusters, separate cloud accounts, on-premises environments, or specialized execution planes. Users expect a single platform experience across notebooks, catalogs, query tools, dashboards, and AI assistants. The central identity gateway pattern reduces upstream identity-provider load to scale with active users rather than user-tool combinations, standardizes identity headers for downstream services, and creates a reusable foundation for AI assistants that can act with delegated user identity across data planes.

Security guardrails in the post include mutual TLS or workload identity between gateways, stripping inbound identity headers before injecting trusted ones, and defining explicit failure behavior for session-store unavailability.

Unknown

  • Independent verification of the 55% figure is not available; the number comes from NVIDIA's own post.
  • The latency and capacity cost of the per-request session-validation side-call at production scale is not quantified in the post.
  • Whether the pattern has been adopted outside NVIDIA is not addressed in the post.

Our take

The pattern solves a real SSO pain for multi-cluster AI workbenches. The less marketed tradeoff is concentration: session ownership and Redis become a platform-wide dependency — the same design that cuts login friction also widens the blast radius when the central path fails. That is the decision, not a footnote.

Sources