Google Cloud has made the Storage-optimized Z4M machine series generally available in Compute Engine, on both virtual machines and bare-metal instances. Z4M is built for AI/ML training and inference workloads that handle very large data volumes and require parallel data processing, such as checkpointing, distributed key-value (KV) cache, and parallel file systems. The platform combines fifth-generation Intel Xeon Scalable processors with the latest Titanium enhancements, providing up to 168,000 GiB of Local SSD (LSSD) and up to 400 Gbps of Cloud Remote Direct Memory Access (RDMA) networking. RDMA reduces data-access latency by up to 60% compared with TCP, which is important for keeping GPUs and TPUs busy while right-sizing cloud investment.

The Z4M VM portfolio supports up to 192 vCPUs, 1,536 GiB of memory, and 875 GiB of LSSD per vCPU, so clusters can be scaled to workload requirements. Bare-metal instances provide direct access to physical hardware without a virtualization layer, lowering latency for custom hypervisors and licensing-specific workloads, and they are positioned for agentic AI workloads, including isolated sandboxes at scale with Agent Substrate. Storage is handled by Titanium SSDs, which offload local storage processing from CPU resources. Compared with the previous Z3 family, Z4M delivers up to 15.6M random-read IOPS and 75,600 MiB/s sequential-read throughput, with LSSD performance improving by up to 70% and throughput by up to 110%. The 400 Gbps RDMA network supports NVIDIA GPUDirect Storage and TPUDirect Storage so massive data transfers can happen at line rate while GPUs and TPUs process large multimodal datasets. Z4M also maintains a balanced LSSD-to-networking throughput ratio of 1.4:1.

The source lists three workload categories with associated benefits. For AI/ML training and inference, the large LSSD capacity plus low-latency RDMA with GPUDirect Storage and TPUDirect Storage helps keep accelerators fully utilized. For parallel file systems and enterprise-class storage, the high vCPU:LSSD ratio of 1:875 and balanced throughput ratio improve application performance by processing data locally and transferring to clients at line rate. For business-critical databases, sustained LSSD I/O performance supports high transaction rates. Partner testing reported in the announcement adds quantitative detail: Silk measured up to 3.7x lower latency and up to 6x lower CPU usage with Cloud RDMA versus TCP; WEKA measured roughly twice the read throughput and up to 79% more IOPS against the Z3 family. Nutanix and Sycomp also provided statements on density, scale, and storage-platform support in the launch context.

For fleet management, Cluster Director can configure, deploy, and manage large Z4M clusters, with compact or spread topologies targeting availability domains, and the ability to combine or stagger maintenance. Z4M instances can be co-located with GPU and TPU clusters for real-time, at-scale processing. Maintenance is designed to be more predictable: the system gives several days of notice, VMs with up to 42,000 GiB of LSSD can live-migrate, and larger VMs or bare-metal instances must be terminated and restarted while preserving data. Z4M also supports Hyperdisk, Google Cloud’s workload-optimized block storage, including Hyperdisk Balanced, Hyperdisk Throughput, and Extreme Hyperdisk. Storage capacity can scale to 512 TiB per instance, with Hyperdisk Balanced providing up to 320K IOPS and Hyperdisk Extreme reaching 500K IOPS and 12,500 MiB/s per instance for demanding databases. Z4M VMs and bare-metal instances are available today in select regions worldwide, and are selectable under the Storage-Optimized machine family in the Google Cloud console.