AI Gateway

Three separate engineering organizations run something called «AI Gateway», and they do not mean the same component. At Uber, six engineers (Matt Mathew and co-authors, 21 May 2026) place it inside a Zero Trust architecture for AI agents, where it mediates external LLM calls and redacts sensitive data through AI Guard, sitting beside an «MCP» Gateway that enforces policy on tool access. At «Bloomberg», Lei Zhang's platform team offers a gateway as one lane of a "Paved Path" for 9,000 engineers, next to simplified PaaS deployment and an «MCP» hub for shared connectors. Cloudflare's «VibeSDK» (23 September 2025) wires «AI Gateway» to unified access across OpenAI, Anthropic and Google, automatic caching of popular responses, and tracking of requests, tokens, response times and cross-model costs.

The gateway is also where traffic becomes a dataset. The «service de fine-tuning RL» announced by Michelle Chen, Alex Reneau and Kevin Flansburg on 1 October 2026 draws its training data from «AI Gateway», with Workers AI running rollouts, Containers sandboxing the scoring, and a new Trainer updating weights.

That is the pattern worth watching. A component installed for observability, caching and redaction ends up holding the record of every model call an organization has made, in a form ready to specialize a model on it. Who may read that record, and how long it is retained, goes unstated.