Google Cloud details July AI infrastructure updates, including Managed Lustre GA

Google Cloud used its August 1 AI infrastructure roundup to frame July as a month of production-oriented updates for organizations moving agentic AI and large model workloads out of pilots. The post is not a single product launch, but it is a useful marker for how cloud providers are packaging compute, networking, storage, and orchestration into one stack for AI teams.

The most concrete availability change is Google Cloud Managed Lustre reaching general availability. Google says the managed file-system service is now offered in four performance tiers: 125 MB/s, 250 MB/s, 500 MB/s, and 1000 MB/s per TiB of capacity, with scaling up to 8 PB. That matters for training, fine-tuning, simulation, and inference workflows that need very high-throughput access to shared data without each customer operating its own parallel file-system deployment.

Why it matters

The update also points to a broader infrastructure pattern. Google listed the C4N virtual machine series as generally available, describing it as a network- and block-storage-optimized VM family built on 5th Gen Intel Xeon processors and Google's Titanium offload hardware. The company says C4N can reach 400 Gbps of network bandwidth and up to 25 GiB/s of block storage throughput when paired with Hyperdisk Extreme. Those numbers are aimed at the bottlenecks that appear when AI systems are split across many accelerators, services, and datasets.

For Kubernetes users, Google highlighted GKE Dataplane V2 support for standard clusters of up to 15,000 nodes while maintaining active Network Policy enforcement. It also pointed to co-operative time-slicing in llm-d, a technique the company says can raise accelerator duty cycles for reinforcement-learning workloads from roughly 40% to as much as 70% without harming convergence or accuracy.

• Managed Lustre is now generally available with four throughput tiers.
• C4N VMs are now generally available for high-throughput networking and block storage.
• GKE Dataplane V2 now supports very large clusters with Network Policies enabled.
• Google open-sourced k8s-aibom to generate machine-learning bills of materials for AI runtimes on Kubernetes.

The common theme is that AI infrastructure is becoming less about renting accelerators in isolation and more about reducing the friction around data movement, cluster policy, accelerator utilization, and supply-chain visibility. For enterprises, those details often decide whether a model workload stays experimental or becomes a dependable production service.

Sources

• cloud.google.com

Cover image: Christoph Scholz, source, licensed under BY-SA.