Google Cloud Details September AI Infrastructure Push for Agentic Workloads
Google Cloud has declared September "scalability month" and rolled out a set of updates across GKE, compute, storage and orchestration aimed at the expected surge of agentic fleets. The list includes the open-source GKE Agent Substrate, built-in scale-to-zero in GKE, and two generally available VM families.
- Google Cloud says the open-source GKE Agent Substrate runs millions of sandboxes with 10x higher density than standard container runtimes.
- GKE now includes built-in scale-to-zero and custom PromQL metric scaling in the HPA.
- GKE Pod snapshots can cut AI inference start-up by up to 89%, according to Google's internal tests.
- M4N and Z4D VMs are generally available, while Filestore agent volumes are in preview.
- Google cites Forrester, Gartner and SemiAnalysis rankings, and says SeaVerse cut infrastructure costs by 60%.
GKE gets denser sandboxes, scale-to-zero and snapshots
The Google Kubernetes Engine team focused on scalability. According to Google Cloud, the new open-source GKE Agent Substrate is built to run millions of sandboxes with 10x higher density than standard container runtimes. The company also says it delivers sub-500ms resume operations at over 500 suspend/resume activations per second, with a native zero-trust kernel and network isolation.
GKE now has scale-to-zero capabilities built in. Google says users no longer need to configure complex components to scale workloads down, because the HPA with the Autoscaling Metric and support for KEP-2021 handle it out of the box. The same HPA can now scale up and down based on custom PromQL metrics, in addition to standard metrics.
GKE Pod snapshots let users save the running state of a workload, including CPU and GPU memory, and restore it on demand. According to Google's internal tests, the feature can reduce AI inference start-up by as much as 89%.
For reinforcement learning, Google points to bottlenecks that appear when scaling to tens of thousands of parallel rollouts, such as idle accelerators, image cardinality and a saturated control plane. It has made GKE Agent Sandbox optimized for RL, an Agent Sandbox RL orchestration SDK and native integrations for popular RL gyms and harnesses generally available.
Google also launched GKE agentic migration, an agent plugin for moving container workloads from AWS EKS to GKE. It is built as a compilation of agent skills and a local Model Context Protocol (MCP) server, and uses AI to translate AWS EKS IaC and Kubernetes manifests into GKE landing zones, protected by deterministic guardrails.
Storage and compute: Filestore, M4N, Z4D and AlloyDB
Filestore agent volumes offer elastic, persistent file storage for agentic workloads. Through integration with GKE Agent Substrate and GKE Agent Sandbox, Google says the service allocates and attaches a dedicated, isolated file workspace to agent sandboxes in milliseconds. It is available as a preview, and access must be requested.
The M4N VM family is now generally available and targets generative AI and RAG data layers such as Milvus, Pinecone, Qdrant, Vespa and Redis. Google claims the highest per-core IOPS and throughput of leading hyperscalers. Paired with the Titanium offload architecture and Hyperdisk Extreme, M4N instances deliver up to 25,000 MiB/s (25 GiB/s) of aggregate host storage performance and up to 1 million IOPS, according to the company.
Z4D, another Compute Engine product, is also generally available. As a bare metal instance, it provides the high local SSD capacity and low latency that agentic microVMs need, so users can run thousands of isolated sandboxes per host, Google says.
A new version of Storage Intelligence advisor helps answer "What’s in my buckets?" and spot unexpected changes, while enhanced batch operations automate bulk changes such as moving storage classes, deleting stale data or applying metadata, tagging, retention or encryption changes. A new AlloyDB for PostgreSQL version combines the Colossus distributed file system and the Jupiter network for what Google calls a database architecture for the agentic era.
Guides, customer results and analyst recognition
Two practitioner guides accompany the release. One shows how to make GPUs and TPUs across the globe behave as a single pool, using the multi-cluster GKE Inference Gateway for global traffic distribution and the LLM-d router for memory-aware scheduling. Google says routing through the gateway added less than 1% overhead in tests. The other explains workload lifecycle metrics in GKE for teams running AI training or inference on TPUs.
On the customer side, Google says gaming startup SeaVerse uses GKE Agent Sandbox for multi-tenant workloads and decreased its infrastructure costs by 60%.
Google also cites third-party validation. It says it was named a leader in The Forrester Wave: Public Cloud Platforms, Q3 2026, with the highest overall score of any cloud provider. In the 2026 Gartner Magic Quadrant for Strategic Cloud Platform Services it was positioned furthest for "Completeness of Vision", and in the Magic Quadrant for Container Management highest in "Ability to Execute". Google Cloud also received a Gold rating in the SemiAnalysis ClusterMax 3.0 report on GPU providers.
What it means for users
The updates are aimed at teams preparing to run large numbers of AI agents, where density, fast resume times, isolation and cost control matter. Several items, including Agent Substrate, scale-to-zero and the RL sandbox tooling, are described by Google as available now, while Filestore agent volumes remain in preview. Most performance figures come from Google itself or from its internal tests, so results in production may differ.
Google Cloud: What’s new in AI infrastructure and orchestration in September
