Gateways as control planes

Portkey on March 24, 2026 announced a newly unified Gateway and made it fully open source, saying the Gateway now runs in production paths that process "1T+ tokens" and "120M+ AI requests" per day while managing "$180M+ in annualized AI spend" for "24,000+ organizations." (financialcontent.com) The project's GitHub repo describes the AI Gateway as a lightweight, open-source proxy that routes to "1,600+ language, vision, audio, and image models," advertises a sub-millisecond (<1ms) response target and a tiny footprint (~122KB), and provides a two-minute quickstart for integrations. (github.com) Portkey's Integrations and virtual-key model let platform teams store provider credentials once, share them across workspaces, and apply per-workspace budgets and rate limits; Portkey added a dedicated "virtual key" budget feature that auto-expires keys when a USD limit is reached. (portkey.ai) The MCP Gateway release positions a control plane for agent-tool access with session-aware routing, an adapter/registry model for MCP servers, and enterprise auth hooks (OAuth 2.1 plus IdP integrations such as Okta and Entra) to avoid distributing credentials to agents. (portkey.ai) Observability promises include unified traces across LLM calls and tool invocations, searchable audit logs for every tool call, and analytics that surface request volume, latency, and error rates per server/tool/workspace, plus runtime policy enforcement features (PII redaction and content filtering). (portkey.ai) Portkey closed a $15M Series A led by Elevation Capital announced Feb. 19, 2026 to scale the control plane, and the company cited earlier scale metrics in that raise (reporting ~500B LLM tokens and ~125M requests per day at the time of the round). (portkey.ai)