Google Unveils $80B Data Center Overhaul for Agentic AI Era

Google is investing $80 billion in a complete data center redesign as the company pivots from chat-based AI to what executives call the "agentic era" — a shift demanding fundamentally different compute infrastructure.

Mark Lohmeyer, VP and GM of AI and Computing at Google, outlined the transformation in an interview, describing how AI agents now orchestrate multiple sub-agents working in parallel rather than processing single queries. "Now we're in the agentic era, where you express your intent, agents spin off multiple sub-agents, working in parallel, preserving state," Lohmeyer said. "This is a radical shift in what infrastructure needs to do."

Token Surge Drives Hardware Redesign

The scale is evident in Google's metrics: AI token processing across its data centers has increased sevenfold year-over-year. To handle this load, Google deployed its 8th-generation TPU platform, split into two specialized systems — TPU-8t for training and TPU-8i for inference.

The inference-optimized TPU-8i packs 384MB of on-chip SRAM, triple the prior generation, and increased high-bandwidth memory by 50% to 288GB. These upgrades target KV cache storage, a critical bottleneck for faster inference responses.

Network Architecture Scales to Million-TPU Clusters

Google introduced the Virgo network, a collapsed non-blocking architecture designed to connect over a million TPUs across multiple data center sites, with up to 134,000 TPUs in a single fabric.

Storage received parallel upgrades through Managed Lustre 10T, delivering 10 terabytes per second of bandwidth and supporting 18 petabytes of storage in single clusters.

Cost Per Transaction Under Pressure

Lohmeyer stated that inference transactions for agentic workloads increase 50x to 100x compared to non-agentic workloads. He claimed Google's latest platforms reduce cost per transaction by almost 2x for the same work — forward-looking company projections rather than independently verified metrics.

Full-Stack Control as Competitive Moat

The infrastructure push leverages Google's vertical integration from custom silicon through orchestration layers. The company controls its TPU and Axion CPU designs, Google Kubernetes Engine for workload management, and the Virgo networking fabric. This end-to-end ownership allows Google to optimize the entire stack for agent-native orchestration rather than retrofitting existing cloud architecture.

The $80 billion investment, confirmed by multiple financial outlets including the Wall Street Journal, signals the intensity of the AI infrastructure arms race. As enterprises evaluate whether complex AI agents are commercially viable, the cost and performance of underlying infrastructure will determine which cloud providers capture the next wave of AI workload migration.

Google's bet: specialized hardware for agentic patterns — rather than general-purpose compute scaled horizontally — will define the next generation of cloud infrastructure. The industry is watching whether this architecture can deliver on the promise of making autonomous agent systems economically sustainable at enterprise scale.