Google introduces Z4M: storage-optimized compute for AI/ML and parallel

Google Cloud announced the general availability of its storage-optimized Z4M machine series in Google Compute Engine on November 7, 2026. The series arrives in virtual machine and bare-metal formats, engineered for high-capacity data processing, parallel file systems, and machine learning workloads. According to Google Cloud Blog's report, Z4M combines fifth-generation Intel Xeon Scalable processors with custom Titanium silicon architecture to deliver dense local storage alongside 400 Gbps networking.

This breakdown details the architectural specifications, hardware configurations, and operational limits for infrastructure architects, storage administrators, and machine learning engineers. Organizations running distributed key-value caches, deep learning training checkpoints, high-performance computing clusters, or transactional databases can evaluate how Z4M scales throughput across physical nodes. We examine the networking and storage performance metrics, bare-metal deployment targets, maintenance procedures, and regional availability guidelines released by Google.

Google Z4M: the announcement

Google Cloud has made its storage-optimized Z4M instance family generally available across Google Compute Engine, providing compute nodes designed to eliminate storage bottlenecks in data pipelines. Product management lead Bob Napaa stated that the series offers virtualized instances and direct bare-metal hardware. Powered by fifth-generation Intel Xeon Scalable processors, Z4M scales up to 192 vCPUs and 1,536 GiB of memory. These instances pair central compute with up to 168,000 GiB of Local SSD storage and 400 Gbps Cloud Remote Direct Memory Access networking.

The storage subsystem relies on Titanium SSDs, which offload input and output processing from host processors to maintain low latency and data security during heavy operations. In raw performance ceilings, Z4M delivers up to 15.6 million random-read IOPS and sustains sequential-read throughput reaching 75,600 MiB/s. Compared to the previous Z3 generation, these additions represent a 70 percent increase in Local SSD input-output performance and a 110 percent boost in sequential read bandwidth, allowing clusters to process massive datasets without stalling processors.

What actually changed with Google Z4M

A major shift in the Z4M architecture is the standard integration of 400 Gbps Cloud Remote Direct Memory Access networking, cutting data-access latency by up to 60 percent compared to TCP stacks. This network configuration enables direct memory transfers using NVIDIA GPUDirect Storage and TPUDirect Storage protocols, allowing storage nodes to deliver inputs at line rate directly to accelerator memory without CPU mediation. Z4M maintains a balanced 1.4:1 ratio between Local SSD throughput and network bandwidth, alongside 875 GiB of storage per vCPU.

Partner testing demonstrates the impact of these architectural changes across real-world workloads. Silk reported that Cloud RDMA on Z4M achieved up to 3.7 times lower latency and a sixfold reduction in CPU consumption compared to standard TCP. Weka chief product officer Ajay Singh noted that benchmark testing showed roughly double the sequential read throughput and up to 79 percent more IOPS over Z3. Nutanix confirmed support for Nutanix Cloud Clusters on bare-metal Z4M, utilizing Intel Advanced Matrix Extensions to accelerate enterprise virtualization and artificial intelligence operations.

Who should care about Google Z4M

Engineering teams running large-scale machine learning training clusters and inference fleets represent the primary audience for the Z4M release. Fast model checkpointing and distributed key-value cache operations require immense local storage density alongside rapid network synchronization so that expensive accelerators remain continuously saturated. Organizations implementing parallel file systems like Weka, Sycomp, Silk, and VDURA can capitalize on the balanced 1:875 vCPU-to-storage ratio to serve aggregated files directly from local drives at full network line rate, significantly reducing physical node counts and hardware cluster footprints.

Database administrators overseeing high-transaction SQL, NoSQL, vector search engines, and real-time enterprise indexing systems benefit from sustained local read-write input rates. Furthermore, teams developing autonomous agent platforms and custom virtualization stacks should examine the bare-metal Z4M offering. Bare-metal nodes eliminate hypervisor overhead, satisfy software licensing structures, and provide isolated physical execution sandboxes at scale. Google highlighted that these bare-metal servers integrate with frameworks like Agent Substrate, allowing developers to execute secure, high-throughput autonomous agents within dedicated hardware boundaries that require rapid local caching.

How to try Google Z4M

Cloud administrators can provision Z4M virtual machines and bare-metal nodes through the Google Cloud console by choosing Z4M under the Storage-Optimized machine family when creating Compute Engine instances or Google Kubernetes Engine node pools. Because instances are currently deployed in select global regions, organizations must verify geographical coverage through their Google Cloud account representatives. Beyond internal Local SSD drives, Z4M instances can attach Hyperdisk block storage tiers up to 512 TiB per node, including Hyperdisk Balanced with 320,000 IOPS and Hyperdisk Extreme reaching 500,000 IOPS and 12,500 MiB/s throughput.

Cluster orchestration can be handled using Google Cloud Cluster Director alongside deployment scripts in the Cluster Toolkit. Infrastructure teams can configure clusters using a compact topology to minimize physical distance and maximize input speeds, or adopt a spread topology targeting distinct availability domains for fault tolerance. Cluster Director enables administrators to co-locate Z4M storage instances directly inside physical network zones housing GPU and TPU superclusters. This tight proximity ensures that distributed accelerators receive continuous multi-terabyte dataset streams without crossing broad transit fabrics or incurring external transit charges.

What to watch after Google Z4M

Operating high-capacity storage nodes introduces specific lifecycle considerations that infrastructure engineers must monitor closely during production operations. Google provides automated notices several days prior to required hardware maintenance, but operational behavior diverges depending on allocated storage capacities. Z4M virtual machines equipped with 42,000 GiB or less of Local SSD support transparent live migration during scheduled host updates. In contrast, instances provisioned with 84,000 GiB or more of local storage, as well as all bare-metal nodes, cannot live-migrate; they require scheduled termination and restart cycles while preserving underlying data.

To prevent unexpected downtime across critical workloads, administrators should use Cluster Director to schedule or stagger maintenance windows across separate availability domains. This controlled rotation prevents simultaneous host restarts and maintains quorum in parallel storage systems like Silk and Weka. Operators should also monitor storage utilization ratios, adjusting Hyperdisk throughput and capacity independent of core compute resources as operational demands evolve. Managing these maintenance boundaries and disk attachments ensures that distributed AI caches, vector search systems, and high-transaction enterprise databases achieve continuous availability despite physical servicing requirements.