Google AI Updates: April 20, 2026

1. Google Releases A2UI 0.9, a Cross-Platform Generative UI Protocol for Agents

Google. Google published version 0.9 of A2UI, a framework-agnostic protocol that lets agents render UI elements on the fly by composing an application’s existing components. The release ships renderers for React, Flutter, Lit, and Angular plus a shared web core, an Agent SDK in Python (Go and Kotlin coming), client-defined functions, client/server data sync, and improved error handling. Google highlights interop with AG2, A2A 1.0, Vercel’s json-renderer, and Oracle’s Agent Spec, with reference apps including a Personal Health Companion and Life Goal Simulator. For practitioners building agentic frontends, A2UI joins MCP (tools) and A2A (agent-to-agent) as the third leg of the emerging standards stack — worth evaluating before rolling another bespoke renderer. Source

2. Google Adds Marvell as Third TPU Partner for Memory Processing Unit and Inference TPU

Google. Per The Information’s Qianer Liu (April 19), Google is in talks with Marvell to co-design two custom chips: a memory processing unit (MPU) that offloads bandwidth-bound work alongside existing TPUs, and a new inference-optimized TPU. Google plans to produce nearly 2 million MPUs, and the MPU design is targeted to lock by next year before test production. Marvell joins Broadcom and MediaTek as the third design partner — a deliberate multi-supplier split that prevents any single vendor from owning the TPU roadmap, mirroring how automakers manage component suppliers. The move comes as custom-ASIC sales are projected to grow 45% in 2026 and reinforces Google’s bet that TPU economics, not GPU parity, are the route to undercutting NVIDIA on inference. Source