Google Cloud launched Z4M virtual-machine and bare-metal instances, offering up to 168,000 GiB local storage and 400-Gbps networking.

Z4M is a generally available storage-optimized machine series in Google Compute Engine. It uses fifth-generation Intel Xeon Scalable processors and Titanium enhancements, with virtual-machine and bare-metal configurations offering up to 192 virtual CPUs and 1,536 GiB of memory. Local solid-state-drive (SSD) capacity reaches 168,000 GiB, with the virtual-machine portfolio providing 875 GiB of local SSD capacity per virtual CPU.

Google reports substantial local-storage throughput. The announcement specifies up to 15.6 million random-read input/output operations per second (IOPS) and 75,600 MiB/s of sequential-read throughput, with improvements over Z3 of up to 70% and 110%, respectively. Titanium SSDs offload local-storage processing from CPU resources.

Networking supports direct access to storage for accelerator workloads. Z4M provides up to 400 Gbps of Cloud Remote Direct Memory Access (RDMA) networking, which Google says reduces data-access latency by up to 60% compared with Transmission Control Protocol (TCP) networking. It supports NVIDIA GPUDirect Storage and TPUDirect Storage, and Google reports a local-SSD-to-network-throughput ratio of 1.4:1 while transferring data at line rate and processing locally.

The intended workloads include AI training, inference, and parallel file systems. Google specifically identifies checkpointing and distributed key-value caches, while bare-metal access serves custom hypervisors and workloads with particular licensing requirements. It also identifies isolated agentic-AI sandboxes, including Agent Substrate, as a bare-metal use case.