Google Cloud Launches Next-Generation Storage-Optimized Z4M Machine Series to Accelerate AI Workloads and High-Performance Computing

Google Cloud has officially announced the general availability of its next-generation storage-optimized Z4M machine series within Google Compute Engine, marking a significant milestone in cloud infrastructure designed for massive data-intensive environments. Available immediately as both virtual machines (VMs) and bare-metal instances, the Z4M portfolio has been engineered specifically to address the demanding computational and throughput requirements of modern artificial intelligence (AI) and machine learning (ML) lifecycles. As organizations increasingly adopt large multimodal models, generative AI frameworks, and agentic workflows, the underlying infrastructure must scale accordingly. The Z4M series arrives as a direct response to these industry bottlenecks, bridging the historical gap between ultra-fast local storage and high-speed cloud networking.

The rollout of the Z4M series represents the culmination of years of iterative architecture development at Google Cloud, succeeding previous generations such as the Z3 family. Over the past decade, cloud providers have witnessed an exponential surge in data volume requirements driven by deep learning, large-scale data analytics, and distributed enterprise databases. Historically, the primary constraint in AI training and inference was compute capability, specifically the speed of GPUs and TPUs. However, as accelerators have grown exponentially faster, data starvation—the delay in feeding data from storage to the processing units—has emerged as a critical performance inhibitor. Google’s introduction of the Z4M architecture directly targets this bottleneck by integrating advanced hardware acceleration, fifth-generation Intel Xeon Scalable processors, and state-of-the-art networking capabilities to keep accelerators fully saturated.

At the core of the Z4M architecture is an unprecedented integration of hardware density and low-latency connectivity. Powered by fifth-generation Intel Xeon Scalable processors and enhanced by Google’s custom Titanium security and offload technology, the Z4M instances provide up to 168,000 GiB of Local SSD (LSSD) storage. This massive storage capacity is coupled with up to 400 Gbps of Cloud Remote Direct Memory Access (RDMA) networking. By leveraging RDMA instead of traditional TCP networking, the Z4M series achieves up to a 60% reduction in data-access latency. This dramatic reduction allows enterprises to maximize application throughput while maintaining precise control over cloud expenditure. Furthermore, the inclusion of Titanium SSDs offloads local storage processing from the main CPU resources, ensuring real-time data handling, enhanced storage security, and robust performance consistency.

The hardware specifications of the Z4M portfolio are categorized into flexible virtual machine configurations and uncompromising bare-metal instances. The Z4M VMs scale up to 192 vCPUs and 1,536 GiB of memory, offering an exceptional ratio of 875 GiB of LSSD capacity per vCPU and a total LSSD capacity reaching 168,000 GiB. For organizations requiring absolute, unmediated access to physical hardware—such as those operating custom hypervisors, strict software licensing models, or large-scale agentic AI sandboxes via frameworks like Agent Substrate—the Z4M bare-metal instances eliminate the virtualization layer entirely. Benchmark testing underscores these architectural improvements: compared to the preceding Z3 generation, Z4M delivers up to a 70% increase in random-read IOPS—reaching up to 15.6 million random-read IOPS—and a staggering 110% improvement in sequential-read throughput, hitting up to 75,600 MiB/s.

Cloud RDMA and Direct Storage Access Mechanisms

A defining feature of the Z4M series is its implementation of 400 Gbps RDMA-enabled standard networking, which facilitates direct memory access across distributed nodes. In the context of large-scale AI workloads, this capability is vital for operations such as checkpointing, distributed Key-Value (KV) caching, and parallel file system synchronization. By utilizing NVIDIA GPUDirect Storage and TPUDirect Storage, Z4M bypasses traditional CPU and operating system bottlenecks, allowing massive multimodal datasets to transfer directly between local storage and accelerator memory at full line rate. The balanced LSSD-to-networking-throughput ratio of 1.4:1 ensures that data processed locally can be streamed to clients instantaneously, preventing any queueing delays that might otherwise leave expensive accelerators idle.

Industry Applications: AI/ML, Parallel File Systems, and Enterprise Databases

The versatility of the Z4M machine series allows it to cater to three distinct, mission-critical workload categories: AI and machine learning training and inference, parallel file systems, and business-critical enterprise databases.

For AI and ML applications, the combination of high-density LSSD storage and ultra-low latency RDMA networks addresses the perennial challenge of rapid model checkpointing and distributed KV cache management. In enterprise storage environments, the Z4M series serves as an ideal foundation for high-performance parallel file systems and software-defined storage solutions provided by ecosystem partners. For relational and non-relational databases, including SQL, NoSQL, and modern vector databases essential for semantic search and retrieval-augmented generation (RAG), the sustained high I/O performance ensures low transaction latencies and high transactions-per-second (TPS) metrics.

Ecosystem Reactions and Partner Validations

The release of the Z4M series has drawn strong endorsements from leading enterprise software and storage vendors within the Google Cloud ecosystem, who have conducted rigorous early testing and integration of the new hardware.

Nutanix highlighted the synergy between its cloud platform and Google’s latest hardware. Saveen Pakala, Vice President of Product Management at Nutanix, noted, "Nutanix is committed to helping organizations run their applications and data anywhere. Bringing Nutanix Cloud Clusters (NC2) to Google Cloud’s new Z4M metal instances reflects that mission and our continued partnership with Google Cloud. By combining Z4M’s exceptional I/O performance and high storage density along with Intel’s advanced compute and AMX-driven efficiency, NC2 on Z4M provides the scale, density, and performance required to power everything from dense VDI clusters to demanding, data-intensive AI workloads."

Storage innovators also reported dramatic performance gains during evaluation phases. Adik Sokolovski, Chief R&D Officer at Silk, stated, "Google’s Z4M VMs bring Cloud RDMA to a storage-optimized VM with much higher local SSD density than previous generations. In our testing, Cloud RDMA on Z4M demonstrated outstanding performance with up to 3.7x lower latency and up to 6x lower CPU usage than TCP. For our customers, that means faster, more predictable performance across a wide variety of use cases, from I/O-intensive databases to analytics and AI workloads, with the storage density to scale capacity efficiently."

High-performance computing specialists emphasized the generational leaps achieved in throughput and capacity. Scott Fadden, Senior HPC Solutions Architect at Sycomp, remarked, "We’re excited to support Google Cloud’s new Z4M platform as part of the Sycomp Intelligent Data Storage Platform. With greater capacity per node and the continued evolution of the Z-series architecture, Z4M helps customers achieve new levels of performance and efficiency. Supporting innovations like Z4M reinforces Sycomp’s commitment to delivering high-performance storage solutions across the Google Cloud ecosystem."

Similarly, Ajay Singh, Chief Product Officer at WEKA, pointed to measurable performance metrics from joint trials. "We’ve worked alongside Google Cloud for years on the most demanding AI and machine learning workloads. In our testing, we found the new Z4M machines deliver roughly twice the read throughput and up to 79% more IOPS, generation over generation (Z3 family). This means more data, more tokens, and more finished work. We’re excited to see what our joint customers build with these powerful machines."

Orchestration, Management, and Cluster Topology

Managing infrastructure at the scale required by Z4M necessitates sophisticated orchestration tooling. Google Cloud addresses this through Cluster Director, which offers advanced placement and maintenance capabilities designed to work seamlessly with the Cluster Toolkit for rapid deployment. Administrators can configure cluster topologies to be compact—maximizing intra-cluster bandwidth and minimizing latency for tightly coupled computational tasks—or spread across specific availability domains to ensure high availability and fault tolerance. Furthermore, Z4M instances can be co-located directly alongside dedicated GPU and TPU clusters, establishing an optimized physical layout for real-time, large-scale data processing pipelines.

Maintenance, Migration, and Hyperdisk Integration

Operational planning is a critical consideration for enterprise-grade infrastructure. The Z4M series incorporates structured maintenance notification protocols, providing advance warning several days prior to scheduled interventions. For Z4M virtual machines configured with 42,000 GiB or less of LSSD storage, Google Cloud supports live-migration capabilities during maintenance events, ensuring zero downtime for running applications. For larger configurations featuring 84,000 GiB or more of LSSD storage, as well as all bare-metal instances, planned maintenance requires a standard termination and restart cycle, during which user data is securely preserved.

In addition to Local SSD storage, Z4M VMs and bare-metal instances offer full compatibility with Hyperdisk, Google Cloud’s workload-optimized block storage service. This integration allows users to independently tune storage performance and capacity per instance. Specifically, Z4M supports Hyperdisk Balanced, Hyperdisk Throughput, and Hyperdisk Extreme storage classes. For general-purpose applications, Hyperdisk Balanced delivers up to 320,000 IOPS per instance with an optimal balance of cost and performance. For ultra-demanding database architectures, Hyperdisk Extreme provides sub-millisecond latencies and supports up to 500,000 IOPS and 12,500 MiB/s of throughput per instance, scaling network-attached storage capacity up to 512 TiB.

Broader Market Implications and Strategic Analysis

The general availability of the Z4M series arrives at a pivotal time in the cloud computing market, where infrastructure providers are intensely competing to capture enterprise AI workloads. By bridging the performance divide between local flash storage and high-speed remote networking, Google Cloud is positioning its compute engine as a premier destination for enterprises seeking to train foundational models or execute complex real-time inference at scale.

The integration of Cloud RDMA and Titanium SSD technology reflects a broader industry trend toward hardware offloading and specialized silicon design. As datasets swell into petabyte scales, traditional hypervisor-managed storage architectures introduce latency overheads that degrade the return on investment for expensive AI accelerators like TPUs and GPUs. By eliminating these bottlenecks, Z4M not only enhances raw throughput but also improves the economic efficiency of cloud deployments. Enterprises can process more data per node, reduce the total number of instances required for heavy computational pipelines, and accelerate time-to-market for AI-driven products.

Deployment and Regional Availability

The Z4M machine series is available immediately in select Google Cloud regions worldwide. Enterprise customers and developers can provision Z4M instances by selecting the machine family within the Storage-Optimized category when configuring new virtual machines or Google Kubernetes Engine (GKE) node pools via the Google Cloud Console. Organizations interested in specific regional availability or custom architectural advisory services are advised to consult the official documentation or contact a Google Cloud sales representative to map out their migration and scaling strategies.