According to a review published by Google Cloud on innovations in AI infrastructure and orchestration, the company showcased a series of product updates, technical guides, research reports, and customer stories across AI compute, networking, storage, frameworks, and orchestration software covering the months of May through August 2026.
August 2026 Updates: Storage, Runtimes, and Cloud Run
In the storage domain, Google Cloud's Filestore service received a new backend storage layer built directly on Colossus, Google's distributed storage system. The new layer allows IOPS performance to be provisioned independently from storage capacity and integrates with GKE, and in AI environments may help service large groups of agents (agentic swarms) that read and write to the same dataset without performance degradation.
Additionally, gVisor isolation environments were introduced within distributed Ray clusters on GKE in partnership with Anyscale. This experimental library integrates Google's open-source application kernel, providing stronger isolation than ordinary containers alongside fast startup times and low memory overhead. Furthermore, new dedicated singleton compute runtimes were launched on Cloud Run, designed to run personal AI agents without shutting down when the agent is idle. According to the publication, running such an instance with 1 vCPU and 1 GiB of memory continuously for 30 days costs $5.70.
Implementation Guides and the MCP Protocol in August 2026
The technical guides highlighted an update to the Model Context Protocol (MCP) specification dated 2026-07-28, in which the protocol core became stateless. The handshake was eliminated, and the initialization process defined in SEP-2575 along with the Mcp-Session-Id header defined in SEP-2567 were removed, making every request self-describing and independent.
Another guide addressed real-time load management for AI agents and voice systems managing a continuous, live bidirectional stream of audio chunks, transcripts, model outputs, and synthesized speech, which must update context or trigger tools without dropping the connection. Additional guides presented an architecture for an elastic LLM inference platform on GKE combining Capacity Advisor and Compute Advisor with RunAI:model streamer or GCPFuse; instructions for updating GPU and TPU hosts without live migration; creating Advanced Compute Images (ACIs) templates using the CLI, console, or Slurm; and three approaches to dynamic capacity management in Google Cloud: scheduling capacity for planned downtime, maintaining automated fallback capacity for unplanned downtime, and using GKE orchestration capabilities.
Among customer updates in August: UiPath moved to a shared GPU fleet on Google Cloud featuring A3 instances with NVIDIA H100 chips for training and G4 instances with NVIDIA RTX PRO 6000 Blackwell Server Edition chips for inference; Mirendil AI lab is using AI Hypercomputer with TPU and GPU processors; Replenit built a decision engine using BigQuery, the Gemini Enterprise Agent Platform, and open-source Gemma models on Cloud TPUs, and reported that the latter combination provided it with 90% lower pipeline costs compared to its previous cloud provider; and Malachyte architected an e-commerce recommendation platform on top of Bigtable, Managed Service for Apache Kafka, Pub/Sub, Compute Engine, and GKE.
July 2026 Updates: Lustre, C4N Machines, and Supply Chain Security
In July 2026, Google Cloud Managed Lustre became generally available (GA) in four performance tiers delivering throughput of 125, 250, 500, and 1000 MB/s per TiB of capacity, scaling up to 8 PB of storage, powered by DDN EXAScaler. Additionally, C4N virtual machines powered by 5th Gen Intel Xeon processors and Titanium offloading hardware were launched, delivering 400 Gbps network bandwidth, 95 million packets per second (MPPS), and block storage throughput of up to 25 GiB/s when paired with Hyperdisk Extreme.
Other July updates include GKE Dataplane V2 support for up to 15,000 nodes with active Network Policy enforcement, a co-operative time-slicing feature in the llm-d tool that enables raising accelerator utilization from a baseline of about 40% up to 70% in reinforcement learning (RL) workloads, and the open-source k8s-aibom security tool, which monitors container clusters, detects AI runtimes such as vLLM and Triton, and generates CycloneDX-formatted ML-BOM files.
July guides included: Day 0 support for Moonshot AI's 2.8-trillion-parameter Kimi K3 model; using GKE managed DRANET with GPUs and TPUs for Gemma 4; running Ray on TPUs; a microbenchmark suite for assessing TPU performance; and optimizing the Mistral 3 Large MoE model on Ironwood (TPU v7x), which achieved a 1.5x performance gain and an increase of up to 48% in throughput. Additionally, Google was reported as named a Leader in the Gartner Magic Quadrant for AI Infrastructure, and a survey of over 1,400 IT leaders found that 83% of organizations need infrastructure upgrades to support production-grade agentic AI applications.
June and May 2026 Updates: Confidential Computing, Inference Gateway, and Cloud Storage Rapid
In June 2026, Confidential Computing capabilities were launched to cryptographically protect data in use within TEE environments for G4 machines featuring Blackwell chips; the TPU Developer Hub was launched; and an OpenTelemetry-based telemetry collection agent was added for TPUs. According to an independent benchmark report, GKE Inference Gateway achieved 15.7% higher throughput, 92.8% shorter wait times, and 62.6% lower inter-token latency compared to the next leading managed Kubernetes service, thanks to prefix caching that stores KV cache states of repetitive prompts. Guides were also published on connecting agents to unstructured data in Cloud Storage via MCP and addressing cold starts using Run:ai Model Streamer, alongside customer stories from Pager Health and Trustpilot.
In May 2026, GKE Agent Sandbox reached general availability (GA), the Agent Substrate open-source project was launched to increase agent infrastructure density, and on-device language model evaluation support was added to Google AI Edge Portal. Furthermore, the Cloud Storage Rapid family was introduced—comprising Rapid Bucket and Rapid Cache—and architecture reviews were published on Google network upgrades and a cluster-level reliability model for frontier model development on TPUs, alongside Imgix using AI Hypercomputer with G4 instances to serve over 8 billion images and videos.