<-- Back to All News

Storage-optimized Z4M machine series are GA

Publish Date: November 9, 2026

Executive Overview

Modern enterprise computing environments are experiencing a fundamental inflection point driven by the computational appetite of generative artificial intelligence training, high-throughput inference fleets, and real-time distributed data systems. As organizations scale complex machine learning pipelines, traditional cloud infrastructure architectures frequently encounter severe I/O bottlenecks where high-cost graphics processing units and tensor processing units sit idle, starved for training batches and model weights. Google Cloud has formally addressed this persistent systemic constraint by announcing the General Availability (GA) of the Z4M storage-optimized machine series within Google Compute Engine.

Engineered specifically for demanding artificial intelligence and machine learning workloads, distributed key-value caches, and parallel file systems, Z4M redefines local storage density and throughput metrics in cloud virtualization. By pairing fifth-generation Intel Xeon Scalable processors with next-generation Local Solid-State Drives (LSSD), high-speed Cloud Remote Direct Memory Access (RDMA), and a balanced ratio of local storage to networking capacity, the Z4M series establishes an infrastructure foundation designed to eliminate data movement latency across demanding enterprise environments. Available in both virtual machine shapes and bare-metal configurations, Z4M provides systems architects with flexible, high-density computing platforms tailored to process massive datasets locally while sustaining full network line-rate client transfers.

Key Technical Specifications

Feature Metric Z4M Series Capability Enterprise Benefit
Max Local Storage Up to 88 TB Local SSD Hosts massive working sets in flash tiers
Max Random Read IOPS Up to 15.6 Million IOPS Eliminates bottlenecks for transactional systems
Sequential Read Bandwidth Up to 75,600 MiB/s Sustains high-throughput batch loading for GPUs
Networking Architecture Sub-millisecond Cloud RDMA Direct memory-to-memory transfers bypassing OS stack
Throughput Balance 1.4:1 LSSD to Network ratio Prevents internal pipeline contentions
vCPU to LSSD Ratio 1:875 capacity ratio Optimizes cost per terabyte of local storage

Features

The Z4M machine series incorporates a cohesive suite of hardware innovations and architectural primitives engineered to maximize throughput while minimizing data access latency.

  • Unprecedented Local Storage Performance and Capacity: Z4M delivers up to 15.6 million random-read I/O operations per second (IOPS) and up to 75,600 MiB/s of sequential-read throughput. This represents an improvement in Local SSD performance and sequential throughput of up to 70% and 110% respectively when compared against the predecessor Z3 machine series. Each instance supports up to 88 TB of integrated high-performance Local SSD storage, accommodating massive distributed caches and working datasets entirely in local memory and flash storage tiers.
  • Balanced Compute, Storage, and Network Ratios: The platform is architected around a purpose-built 1.4:1 ratio of Local SSD throughput to network throughput, supported by an exceptionally high vCPU-to-Local-SSD capacity ratio of 1:875. This balance ensures that storage controllers and network interfaces operate in harmony, allowing instances to process data locally at maximum speeds while concurrently streaming data across high-speed fabric interconnects at full line rate without contention.
  • Sub-Millisecond Cloud RDMA Networking: Integrating Cloud RDMA enables direct memory-to-memory data transfers between interconnected instances, bypassing kernel software stacks, operating system context switches, and traditional TCP protocol overhead. Cloud RDMA reduces data-access latency across distributed nodes by up to 60% compared to standard TCP networking, accelerating checkpointing sequences and parallel file synchronizations.
  • Dual Deployment Modalities (Virtual Machines and Bare-Metal): Z4M is deployed as both managed Compute Engine virtual machines and dedicated bare-metal instances. Bare-metal nodes provide direct physical access to underlying hardware execution layers, eliminating hypervisor overhead and supporting specialized enterprise hypervisors, custom operating system kernels, and proprietary licensing models such as Nutanix Cloud Clusters.
  • Flexible Maintenance and Availability Domains: To support continuous operations across enterprise fleets, Z4M virtual machines equipped with 42,000 GiB or less of Local SSD support transparent live migration during planned cloud platform maintenance windows. For larger configurations (84,000 GiB or more of LSSD and bare-metal nodes), Google Cloud provides structured maintenance orchestration via Cluster Director, enabling infrastructure teams to stagger maintenance cycles across distinct availability domains while preserving localized data integrity.
  • Deep Integration with Hyperdisk Storage: In addition to internal high-speed Local SSD arrays, Z4M instances support seamless attachment to Google Cloud Hyperdisk block storage tiers up to 512 TiB per instance. Support for Hyperdisk Extreme delivers ultra-low block latency with up to 500,000 IOPS and 12,500 MiB/s of dedicated throughput, satisfying demanding enterprise transactional database requirements.

Benefits

Deploying the Z4M machine series provides immediate technical and economic advantages across enterprise IT and cloud operations.

Maximized GPU and TPU Accelerator Saturation

The primary operational cost in AI clusters stems from idle compute time during training and inference phases. By delivering 75,600 MiB/s of local sequential read bandwidth alongside sub-millisecond Cloud RDMA networking, Z4M continuously streams multi-gigabyte training batches and complex KV cache states to adjacent accelerator clusters, keeping hardware accelerators saturated and shortening training timelines.

Substantial Infrastructure Footprint and Cost Reduction

The high storage density per node—characterized by the 1:875 vCPU-to-storage ratio—enables organizations deploying distributed parallel file systems to consolidate their physical storage footprint. Enterprises can achieve required petabyte-scale capacity targets utilizing significantly fewer compute nodes, directly reducing per-core software licensing expenses and baseline VM rental fees.

Predictable Performance for Transactional Workloads

Enterprise database systems often suffer from I/O degradation during periods of peak transaction concurrency. Sustained high random IOPS (up to 15.6M IOPS) ensures that transactional SQL engines, NoSQL key-value stores, and vector search indices maintain steady, low-latency transaction processing regardless of concurrent query volumes.

Seamless Enterprise Hybrid Migration

With bare-metal Z4M support and validated partner platforms like Nutanix Cloud Clusters, enterprises can migrate legacy on-premises I/O-intensive workloads to the cloud without refactoring architecture, maintaining existing operational playbooks while gaining cloud elasticity.

Use Cases

The distinct balance of local storage bandwidth, processor performance, and low-latency networking positions Z4M across several critical enterprise applications.

  • Large-Scale AI Model Checkpointing and Distributed KV Caching: Distributed training runs for large frontier language models require frequent state snapshots (checkpoints) to prevent loss of progress in the event of hardware failure. Z4M instances act as high-speed checkpoint repositories, ingesting and writing terabytes of model state in seconds via Cloud RDMA. Simultaneously, inference serving platforms leverage Z4M’s massive LSSD capacity as a distributed key-value cache layer, retrieving pre-computed attention keys instantly across long conversational context windows.
  • High-Performance Parallel File Systems: Leading enterprise storage vendors—including WEKA, Sycomp, Silk, and VDURA—deploy distributed file system architectures on Z4M infrastructure. These parallel storage solutions serve aggregated datasets to thousands of client nodes at full line rate, creating high-speed unified storage namespaces for genomic sequencing, seismic processing, and financial algorithmic simulation.
  • Business-Critical Enterprise and Vector Databases: High-volume relational database clusters (PostgreSQL, MySQL, SQL Server), distributed NoSQL platforms (Cassandra, ScyllaDB), and dedicated vector search engines utilize Z4M’s local drives to execute billions of transactional reads and vector similarity searches with sub-millisecond query response profiles.
  • Dedicated Sandboxing for Autonomous Agent Frameworks: As organizations deploy autonomous software agents requiring dedicated, isolated execution environments, bare-metal Z4M instances provide hardware-level sandboxes integrated with platforms like Agent Substrate. Autonomous code execution, compilation cycles, and temporary container builds occur on ultra-fast local storage without hypervisor jitter or shared-tenant resource contention.

Alternatives

Technology leaders evaluating high-performance storage-optimized infrastructure must compare several public cloud and on-premises deployment approaches.

  • Amazon Web Services EC2 I4i and I8g Instances: AWS provides storage-optimized compute through the I4i series (powered by Intel Ice Lake processors) and the newer I8g series (powered by AWS Graviton4 processors). While AWS storage-optimized instances deliver competitive local NVMe drive density, Z4M differentiates itself through native Cloud RDMA interconnects capable of reducing network latency by 60%, alongside higher aggregate sequential read throughput (up to 75,600 MiB/s) and a higher vCPU-to-storage ratio.
  • Microsoft Azure Lsv3 and Lsv4 Virtual Machines: Azure’s storage-optimized Lsv3 and Lsv4 series offer high-throughput local NVMe storage directly mapped to virtual machines for data-intensive workloads. However, Azure’s standard L-series instances are primarily architected around conventional TCP networking fabrics, lacking the tightly integrated, co-located RDMA director capabilities that allow Google Cloud customers to place Z4M storage clusters in the same physical fabric zones as TPU and GPU superclusters.
  • Dedicated On-Premises All-Flash NVMe Appliances: Enterprise data centers frequently rely on dedicated physical storage appliances (such as Pure Storage FlashBlade or Dell PowerStore) interconnected via InfiniBand. While physical appliances eliminate public cloud networking charges, they impose massive capital expenditure budgets, rigid multi-year capacity planning, physical data center space limitations, and lengthy procurement lead times that fail to match the dynamic provisioning agility of cloud-native Z4M fleets.
  • Cloud-Native Network-Attached Block Storage (Google Hyperdisk Extreme / AWS io2 Block Express): Rather than relying on local NVMe storage, organizations can configure standard compute instances attached to extreme-tier network block storage. While network-attached volumes offer effortless independent scaling and snapshot flexibility, they incur higher per-gigabyte operational costs at massive scale and remain bounded by network attachment limits, unable to match the 15.6M random-read IOPS achieved by Z4M’s integrated physical bus.

An Alternative Perspective

A comprehensive architectural analysis requires evaluating the practical engineering trade-offs, operational liabilities, and governance challenges associated with storage-optimized cloud instances.

The primary operational consideration when utilizing Local SSD storage is ephemeral data persistence. Unlike persistent network-attached block storage (such as Google Hyperdisk or Persistent Disk), Local SSDs are physically tethered to the hosting hypervisor hardware. While Z4M configurations under 42,000 GiB support live migration during routine infrastructure maintenance, instances with larger storage configurations (84,000 GiB or more) and bare-metal nodes cannot undergo live migration. During unplanned hardware faults or host reboots, data residing solely on local drives risks loss unless the application layer actively manages synchronous replication across independent availability zones. Engineering teams must treat Local SSDs as a high-speed transient tier, backing data with robust replication topologies or asynchronous persistent snapshots to avoid catastrophic data loss.

Furthermore, deploying high-density storage nodes can inadvertently amplify enterprise failure domains. Because a single Z4M node can host up to 88 TB of local storage, the unexpected failure of an individual physical server removes a substantial percentage of cluster storage capacity. In distributed parallel file systems and database clusters, reconstructing 88 TB of data across the network during node replacement incurs massive network transit overhead that can temporarily degrade cluster-wide query throughput.

Finally, organizations must account for regional availability constraints and capacity commitments. Due to the high-density hardware footprint of fifth-generation Intel Xeon processors and enterprise LSSD arrays, Z4M instances are rolled out selectively across specific global cloud regions. Multinationals requiring uniform global latency may find Z4M unavailable in secondary data centers, necessitating hybrid architectures that balance Z4M in core regions with standard storage-optimized machines elsewhere.

Final Thoughts

The General Availability of the Z4M machine series represents a strategic leap forward in cloud compute design, directly eliminating the I/O starvation bottleneck that has historically plagued large-scale artificial intelligence, distributed caching, and parallel computing architectures. By pairing up to 15.6 million IOPS and 75,600 MiB/s of sequential throughput with sub-millisecond Cloud RDMA interconnects, Google Cloud has delivered an infrastructure platform capable of matching the relentless pace of modern accelerator clusters.

For enterprise infrastructure architects, chief technology officers, and engineering directors, Z4M provides a compelling blueprint to consolidate parallel storage footprints, accelerate critical AI checkpointing cycles, and optimize cloud infrastructure expenditure. By pairing these powerful machines with disciplined data replication and availability domain planning, organizations can build resilient, high-velocity data platforms prepared for the computational demands of the agentic era.

Source

Google Cloud Blog: Storage-optimized Z4M machine series are GA