Google Cloud released three concrete changes in August 2026 that directly affect AI‑focused workloads: Filestore now runs on a Colossus‑backed storage layer that lets you provision IOPS independently of capacity, gVisor sandboxes are exposed for distributed Ray clusters on GKE, and Cloud Run instances provide dedicated, always‑on runtimes for low‑cost personal agents. Each change touches performance, isolation, or cost predictability, which are core concerns for AI engineers, platform teams, and SREs.
Decoupled IOPS in Filestore
Filestore’s new backend leverages Google’s Colossus distributed storage system. The key technical shift is the ability to set IOPS separate from the amount of storage allocated. Because the service is tightly integrated with GKE, workloads that rely on shared NFS mounts—such as large agentic swarms that read and write a common dataset—can now scale I/O throughput without inflating storage size. For practitioners, this means you can tune storage performance to match the I/O profile of your AI pipelines while keeping storage costs under control.
Operationally, the independent IOPS setting introduces a new configuration dimension in your Filestore provision scripts. Monitoring should be extended to track IOPS consumption versus the provisioned quota to avoid throttling. Since the backend is built on Colossus, you can expect the same durability and regional replication characteristics as other Google storage services, but you should verify that any existing backup or snapshot processes remain compatible with the updated service.
gVisor Sandboxing for Ray on GKE
In partnership with Anyscale, Google released an experimental library that injects gVisor—Google’s open‑source application kernel—into Ray clusters running on GKE. gVisor provides lightweight sandboxing with stronger isolation than standard containers while keeping startup latency low and memory overhead modest. The feature is presented as a library, so enabling it requires adding the library to your Ray cluster definition and configuring the GKE pod spec to use the gVisor runtime.
From a security perspective, the sandbox adds a boundary that limits the impact of a compromised worker process, which is valuable in multi‑tenant or untrusted model execution scenarios. For DevOps and SRE teams, the experimental status suggests you should test the integration in a staging environment before production rollout, and monitor for any compatibility issues with Ray extensions or custom operators.
Always‑On Cloud Run Instances
Cloud Run instances are now offered as dedicated, singleton compute runtimes that remain active even when idle. The pricing detail disclosed—$5.70 per month for a 1 vCPU, 1 GiB instance—makes the offering attractive for personal AI agents that need instant responsiveness without the latency of cold starts. Because the instance does not scale down automatically, you retain a predictable performance baseline at a modest cost.
Implementation requires specifying the instance mode in your Cloud Run service configuration. Operators should incorporate health‑checking and logging to ensure the singleton remains healthy over long idle periods. The cost model also encourages budgeting for always‑on services, which may affect capacity planning for larger fleets of agents.
Stateless Model Context Protocol (MCP) Update
The MCP specification released on 2026‑07‑28 removed the handshake and session identifier headers, making every request self‑describing and independent. This shift eliminates stateful session management on the server side, simplifying the request flow for AI model serving endpoints.
Practically, developers must adjust client libraries to include all required context in each request, and server implementations need to drop any session‑tracking logic. The change may reduce memory pressure on serving nodes, but it also means that any caching strategy must be re‑evaluated because the protocol no longer guarantees session continuity.
Related CloudNinjas coverage: Google Cloud.
What This Means For Practitioners
Evaluate Filestore configurations to match IOPS to your AI swarm workloads and update monitoring accordingly. Test the gVisor‑Ray integration in a controlled environment to verify isolation benefits without disrupting existing pipelines. Consider Cloud Run instances for low‑latency personal agents, and factor the always‑on cost into your budgeting. Finally, revise your MCP client and server code to accommodate the stateless request model, and review any caching or session‑based logic that may no longer apply.
