AI Gateway

Uber's platform team lists AI Gateway as one of six components in the zero trust architecture that Matt Mathew and five colleagues published on 21 May 2026, alongside the Agent Registry, the AI Agent Mesh, the STS, the MCP Gateway, and SPIRE. Its role in that stack is narrow and specific: it mediates outbound LLM calls and redacts sensitive data through AI Guard. The gateway sits at the boundary where an internal agent stops talking to Uber systems and starts talking to an external model, which is exactly where the actor chain and the single-hop token stop protecting anything.

The same component name recurs in two very different places. Lei Zhang, who runs technology infrastructure at Bloomberg, describes a central AI gateway as part of the "paved path" platform serving 9,000 engineers, sitting next to a simplified PaaS deployment and an MCP hub for sharing connectors. Cloudflare ships AI Gateway inside VibeSDK, where it handles unified access to OpenAI, Anthropic and Google, response caching to cut inference costs, and centralized observability on requests, tokens and response times.

Three organizations, three motives: Uber wants security enforcement, Bloomberg wants a governable default path, Cloudflare wants multi-model routing and cost control. The convergence is on the chokepoint itself. Once agents call models from thousands of places, someone has to own the single door those calls pass through, and the question of what that door is for stays unsettled.