Lifestyle
Google Cloud launches Z4M storage-optimized machines: large local SSD plus 400 Gbps networking, aimed at AI training and big data workloads
Google Cloud has announced general availability of its Z4M storage-optimized machine series, rentable as virtual machines or whole physical servers (bare metal). Z4M offers large local SSD capacity and 400 Gbps Cloud RDMA networking, affecting businesses and developers running AI training, large file systems or databases on Google Cloud. Performance figures are self-reported by Google Cloud and its partners.
About 6 min read
What happened
Google Cloud announced that its new-generation storage-optimized Z4M machine series is now generally available (GA) on Google Compute Engine. Customers can choose a virtual machine (VM), a virtual computer carved out of a physical server, or a bare-metal instance, which uses the physical hardware directly with no virtualization layer.
According to Google Cloud, Z4M is designed for three kinds of AI/ML training and inference (running a trained model) workloads. The first is checkpointing, which periodically saves progress during training. The second is distributed key-value (KV) caching, which spreads frequently used data across multiple machines for fast access. The third is parallel file systems, storage systems that let many servers read and write the same data at once. Google Cloud says Z4M uses 5th-generation Intel Xeon Scalable processors and the latest Titanium enhancements, with up to 168,000 GiB of local SSD (LSSD) and up to 400 Gbps of Cloud RDMA networking.
RDMA (remote direct memory access) is a networking technology that lets servers exchange data more directly. TCP is the transport commonly used on ordinary networks. Google Cloud states that, compared with TCP networking, RDMA can cut data access latency by up to 60%.
Specifications and performance highlights
Google Cloud says Z4M VMs and bare-metal instances offer up to 192 vCPUs (virtual processor cores) and 1,536 GiB of memory. VMs provide 875 GiB of LSSD capacity per vCPU. IOPS means read/write operations completed per second. The company says Z4M reaches up to 15.6 million random-read IOPS (15.6M) and up to 75,600 MiB per second of sequential-read throughput. Compared with the previous-generation Z3, LSSD performance and throughput improve by up to 70% and 110% respectively.
On networking, Google Cloud says Z4M moves data at line rate through NVIDIA GPUDirect Storage and TPUDirect Storage. Line rate means transferring at the maximum network bandwidth, intended to help keep GPUs and TPUs highly utilized. Google Cloud also says Z4M's LSSD-to-network throughput ratio is 1.4:1, so it can process data locally while sending data out at full speed.
Google Cloud's article also cites partners' own test results. Adik Sokolovski of Silk says Cloud RDMA on Z4M delivers up to 3.7x lower latency and up to 6x lower CPU usage than TCP. Ajay Singh of WEKA says Z4M's read throughput is roughly double that of the Z3 series, with up to 79% more IOPS. These are vendor-reported tests, not independently verified.
| Item | Z4M VM | Z4M bare-metal instance |
|---|---|---|
| Features described by Google Cloud | Large total LSSD capacity, 875 GiB LSSD per vCPU | Direct access to physical hardware with no virtualization layer; suited to custom hypervisors (software that manages VMs) and specific licensing needs |
| Maximum specs | Up to 192 vCPUs, 1,536 GiB memory, 168,000 GiB LSSD | Also up to 192 vCPUs, 1,536 GiB memory, 168,000 GiB LSSD |
| Behavior during maintenance | LSSD of 42,000 GiB or less can live-migrate; 84,000 GiB or more must stop and restart (data preserved) | Must stop and restart (data preserved during planned maintenance) |
| Use cases cited by Google Cloud | AI/ML, parallel file systems, enterprise databases | Isolated sandboxes for agentic AI (AI that carries out tasks on its own), including Agent Substrate; custom hypervisors |
Who is affected and why it matters
When training AI models, expensive GPUs and TPUs sit idle if data does not arrive in time. Google Cloud's pitch for Z4M is to deliver data to these accelerators faster so they stay highly utilized.
Google Cloud lists three target scenarios. The first is AI/ML training and inference. The second is parallel file systems and enterprise-grade storage, naming Weka, Sycomp, Silk and VDURA. The third is business-critical enterprise databases, including SQL, NoSQL and vector databases. Hyperdisk is Google Cloud's network-attached block storage, meaning disks reached over the network rather than built into the server. Google Cloud says Z4M also supports Hyperdisk Balanced, Hyperdisk Throughput and Hyperdisk Extreme, up to 512 TiB per instance. It cites up to 320K IOPS for Hyperdisk Balanced and up to 500K IOPS and 12,500 MiB/s for Hyperdisk Extreme.
For management, Google Cloud says Cluster Director and Cluster Toolkit can be used to configure, deploy and manage large Z4M clusters. Z4M can also be co-located with GPU and TPU clusters. Google Cloud also says customers are notified several days before maintenance is needed.
How to get started
According to Google Cloud, when creating a new VM or GKE node pool in the Google Cloud console, choose Z4M under the Storage-Optimized machine series. A GKE node pool is a group of servers in Google's container service. Confirm regional availability with a Google Cloud sales representative.
FAQ
What is Z4M?
According to Google Cloud, Z4M is Compute Engine's new-generation storage-optimized machine series, rentable as VMs or bare metal. It offers large local SSD capacity and 400 Gbps Cloud RDMA networking, for data-intensive work such as AI/ML and parallel file systems.
Are the performance figures reliable?
The figures come from Google Cloud's announcement, which also cites test results run by partners Silk and WEKA. No independent third party has confirmed them, so treat them as vendor-reported.
Where is Z4M available?
Google Cloud says Z4M is available in some regions worldwide; contact its sales representatives for the actual regions.
What happens to my instance during maintenance?
Google Cloud says VMs with 42,000 GiB or less of LSSD can live-migrate during maintenance without shutting down. VMs with 84,000 GiB or more of LSSD and bare-metal instances must stop and restart, with data preserved during planned maintenance. Notice is given several days ahead.
Browse the latest news in this topic
Lifestyle
Meta Launches Spaces in the Muse App: Describe an Idea and Muse Builds a Personalized Feature for You
Meta announced on November 6, 2026 that Muse users can now describe an idea and have Muse build a custom feature called a space, such as a gift tracker or trip planner. Users can choose where spaces appear, and once they connect the apps they use, spaces update automatically in the background.
Lifestyle
Cloudflare: Quantum Computers Could Forge Privacy Pass Tokens but Not Break User Privacy, and MOLE Is Its Next Step
In a research post on November 6, 2026, Cloudflare says future quantum computers could let attackers forge Privacy Pass tokens and bypass sites' anti-abuse protections, while user privacy remains intact. The post also introduces MOLE, a new design for anonymous credentials (digital passes that let a site check access without learning who holds them), with early cost estimates.
Lifestyle
ASUS Unveils XG Core External GPU: USB4 Connection, 16GB Memory for Gaming and Local AI
On November 6, 2026, ASUS announced the XG Core Radeon RX 9060 XT LP 16GB GDDR6 external graphics card. It connects over USB4 and adds stronger graphics to compatible laptops, handhelds and mini PCs without opening them up. Here is what ASUS has announced and what to check before use.
Lifestyle
GitHub Unveils Rubicon, a New Git Storage Architecture Built to Let Thousands of AI Agents Push to One Repository at Once
GitHub's engineering blog describes a new Git storage architecture code-named Rubicon. GitHub says internal tests reached 7,200 pushes per second, and the way developers use Git stays the same. It is in limited private preview, with a broader rollout expected early next year.