# NeuralMesh on Google Cloud Z4M, Get More Work From Every Resource **Author:** Phil Curran **Published:** November 5, 2026 ![Black and white view of rows of server racks with bundled cables in a data center.](https://cdn.sanity.io/images/ult5g8gw/production/c78688b4acd2fb100860216c2703e59eb2069d70-1376x768.jpg) ## TL;DR Google built Z4M for AI/ML workloads and distributed storage at scale: the largest local storage capacity of any Compute Engine machine. We worked with Google through the preview, and NeuralMesh™ supports Z4M at GA. - Up to 168 TiB local SSD, 400 Gbps network, RDMA, accelerator options. - NeuralMesh delivered ~2X sequential read throughput and 76–79% more IOPS than our best prior-gen config. - ~94% of rated network bandwidth delivered as throughput. - Metadata create/remove improved ~35–50%. AI workloads keep expanding. Datasets grow faster than anyone plans for, checkpoints accumulate through every training run, and the context behind production inference has to live somewhere fast. These workloads need more capacity and more performance. The systems supporting them need balance. Adding capacity alone does not help, and neither does adding bandwidth alone. The system has to stay in proportion, so that the capacity you buy is capacity your workloads can actually reach. When it falls out of proportion, the cost shows up first as idle resources, then as long training pipelines and inference sessions waiting on data, and then as over-provisioned infrastructure you are paying for and not using. That is what Google Cloud did with Z4M. Alongside a large increase in capacity, up to 168 TiB of local SSD, the most on a single node in the Compute Engine family, Google Cloud sized the rest of the machine to match, with up to 400 Gbps of network bandwidth, RDMA support, bare-metal shapes, and the option to be colocated with accelerators for fast, low-latency access to data. It is a machine built to sustain performance under load by avoiding bottlenecks, and balanced for the demands of large-scale AI workloads. Designing balanced systems ensures you don't end up with idle capacity on one side, or over-provisioning to work around a bottleneck on the other. That is exactly what we designed NeuralMesh to do: get the most out of every part of an AI infrastructure, rather than the one part that happens to be easiest to grow. ## **Collaborating With Google Cloud on Z4M** WEKA has been collaborating with Google Cloud on performance-intensive workloads for years. NeuralMesh, available on [Google Cloud Marketplace](https://console.cloud.google.com/marketplace/product/wekaio-public/weka-data-platform_vm), deploys inside customers' own Google Cloud projects. Customers run their most demanding AI work on it, from training pipelines through long-context, multi-turn inference. They rely on the same combination for chip design and verification pipelines, where small-file and metadata behavior punishes conventional storage, and for genomic sequencing and analysis, where research moves only as fast as the data underneath it. So we were glad to preview Google Cloud's Z4M, and excited about what it enables for our customers. We spent the preview putting NeuralMesh under load on Z4M and comparing it against the strongest prior-generation configuration we had tested. Here is what customers can expect. **WEKA found roughly twice the read throughput and about 76–79% more IOPS, generation over generation.** In an AI pipeline, sustained read throughput is what keeps accelerators fed during training and what loads a model quickly when an inference service scales up or recovers from a failure. The write side matters just as much and gets discussed less: checkpoints land fast, and a checkpoint that lands fast is a training run that spends less of its life paused. Performance also scaled nearly linearly across the shapes we tested, so capacity, throughput, and resiliency grow together as a cluster grows. **NeuralMesh converted roughly 94% of the rated network into delivered throughput.** That is architectural rather than a tuning result: every node in the NeuralMesh cluster serves data and metadata. There is no metadata tier to size, no controller pair to saturate, and no separate cache layer to manage. It matters most in the environments our customers actually run, where training, checkpointing, and inference land on the same infrastructure at the same time, and one phase cannot be allowed to starve another. **Metadata create and remove performance improved approximately 35–50% over the prior generation at matched core counts.** Metadata performance has emerged as the new storage performance standard to watch with AI workloads. A training job begins with a metadata storm as thousands of ranks come online at once, and modern datasets are made of enormous numbers of small files. It is usually what decides whether a verification run or a simulation job finishes inside the window it was given. "We've been collaborating with Google Cloud for years on some of the most demanding AI and machine learning workloads. In our testing, we found the new Z4M machines deliver roughly twice the read throughput and about 76–79% more IOPS, generation over generation (Z3 family). This means more data, more tokens, and more finished work. We’re excited to see what our joint customers build with these powerful machines." — Ajay Singh, Chief Product Officer, WEKA ## **What This Means for AI Teams** AI infrastructure is judged on finished work: training runs that complete on schedule, checkpoints that don't stall the job, and inference that stays responsive as context and concurrency grow. None of that is decided by any single specification. It is decided by whether the whole system stays in proportion under load. Z4M brings the capacity and keeps it balanced; NeuralMesh turns it into work. ## **Getting Started** NeuralMesh on Google Cloud’s Z4M is available in the regions where Z4M launches. Supported instance types are listed in the [WEKA documentation](https://docs.weka.io/), and you can learn more on our [Google Cloud page](/deployment/google-cloud-platform) or [talk to our team](/contactPage/contact).