Google AI Updates: April 25, 2026

1. Google Commits Up to $40 Billion to Anthropic in Two-Tranche Deal

Google. Google committed to invest up to $40 billion in Anthropic, structured as $10 billion in cash now at a $350 billion valuation plus $30 billion contingent on the Claude maker hitting performance targets. The deal includes five gigawatts of compute over five years and access to up to one million seventh-generation Ironwood TPUs for Anthropic. The move arrives four days after Amazon raised its own Anthropic commitment to roughly $33 billion total, putting the two largest cloud providers in a bidding war to keep the same flagship AI customer on their infrastructure. Source

2. DeepMind Publishes Decoupled DiLoCo for Resilient Multi-Datacenter Training

Google. DeepMind released Decoupled DiLoCo, a distributed-training architecture that splits frontier-model training across loosely coupled “islands” of compute exchanging asynchronous updates. The technique is designed to train across geographically distant data centers with much lower bandwidth requirements and survive hardware failures gracefully — DeepMind reports 88% system availability versus 27% for traditional synchronous training in high-failure scenarios. It is the production follow-up to the original DiLoCo work and is squarely aimed at the multi-region training problem hyperscalers run into above a certain cluster size. Source

3. Cloud Next ‘26 Closes With Seven Headline Announcements Recap

Google. Google published a Cloud Next wrap-up that consolidates the week’s seven biggest moves: the Gemini Enterprise Agent Platform, the Gemini Enterprise consumer-facing app with no-code Agent Designer, the 8th-gen TPU 8t/8i split, the Agentic Data Cloud (Knowledge Catalog plus Cross-Cloud Lakehouse with zero-copy federation onto AWS), AI-powered security with Wiz, Workspace Intelligence across Docs/Drive/Meet/Gmail, and named customer deployments at Home Depot, Papa John’s, Mars, and Unilever. The platform now offers Anthropic’s Claude Opus 4.7 alongside Gemini 3.1 Pro and Nano Banana 2 — Google explicitly framing the platform as model-agnostic. Source

4. Google Open-Sources DESIGN.md as a Portable Brand Spec for AI Coding Agents

Google. Google Labs released the DESIGN.md spec under Apache 2.0: a markdown file with a YAML front-matter block of design tokens (colors, typography, spacing, radii) plus prose explaining rationale. Drop one in a project root and Claude Code, Cursor, Copilot, or Stitch can generate brand-consistent UIs without retelling the system. A community list of DESIGN.md exemplars (VoltAgent/awesome-design-md) is already growing, positioning the file as the AI-agent equivalent of package.json for design intent. Source

5. Google Publishes Spring Productivity Playbook for Gemini Agent Mode and Nano Banana

Google. Google published a tip-sheet showing eight concrete consumer use cases for Gemini’s newer features: personalized cleaning checklists, photo-upload clutter audits, fridge-content meal planning via Gemini Live, camera-assisted home repair troubleshooting, errand planning through Ask Maps, room-redesign visualizations with Nano Banana, plant care diagnostics, and Gmail decluttering. AI Ultra subscribers get Agent Mode for autonomous email organization with smart archiving and task extraction. The post is the clearest single artifact yet of how Google wants consumers to use Gemini Live, Nano Banana, and Agent Mode together rather than as separate products. Source