Google has unveiled its next-generation network architecture, purpose-built to meet the unprecedented demands of AI at global scale. As outlined by Bikash Koley, Google’s VP of Global Networking and Infrastructure, at Google Cloud Next 2025 in Las Vegas, the company is entering a new networking era—shifting from the internet, streaming, and cloud-centric designs toward an architecture optimized for distributed AI workloads. With a bold rethinking of how wide area networks (WANs) should function—treating continents as data centers—Google is reimagining reliability, scalability, and programmability to support the next wave of AI innovation.
This next-gen network is guided by four key design principles: exponential scalability, beyond-9s reliability, intent-driven programmability, and autonomous operation. Google’s multi-shard network architecture, inspired by hyperscale application patterns, supports horizontal growth, regional isolation, and advanced rerouting capabilities. With machine learning at its operational core, the network can proactively detect and resolve issues with minimal human intervention—boosting resilience, reducing outages, and enabling real-time performance for AI-intensive apps across Google Cloud.
Google’s New Network Vision:
• Three eras of evolution: From the internet (Search, Gmail), to streaming (YouTube), to cloud (multi-tenant security & resiliency)—Google’s network has constantly evolved.
• AI-first design: WAN is now the new LAN. AI workloads like Gemini span campuses, metros, and continents, driving demand for ultra-scalable bandwidth and ultra-low latency.
• Multi-shard architecture: Independently operated network shards support elastic growth and regional isolation—enabling 7x bandwidth growth from 2020–2025.
• Protective ReRoute: Reduces cumulative outage minutes by up to 93% by quickly detecting and routing around failures.
• Autonomous operations: ML-powered digital twin (inspired by DeepMind’s GNNs) allows for real-time fault prediction, faster root cause analysis, and smarter capacity planning.
• Intent-driven programmability: Standard APIs and universal network models (like MALT) let customers control compliance, security, and performance dynamically.
• Global footprint: Over 2 million miles of lit fiber, 33 subsea cables, 202 edge sites, 3,000+ CDN locations, connecting 42 cloud regions and 127 zones.
• Customer benefit: Always-on, zero-trust secure, AI-optimized global network available today via Cloud WAN for enterprise customers worldwide.
Watch the full keynote and explore sessions on Google Cloud Next here.
