Google opens Z4M servers with 168,000 GiB of fast local storage

Google made its Z4M server series available to all customers in Google Compute Engine, its cloud platform. Z4M holds and moves the large data of AI training and inference. A maximum of 168,000 GiB of local solid-state storage is in one server. It is for teams that make and operate AI models and must have fast data. Google gave no price.

How it works

Z4M comes as a virtual machine or as a bare-metal server. A bare-metal server gives the customer the hardware with no virtualization layer. It uses Intel Xeon processors and 400 Gbps RDMA networking. RDMA lets one computer read the memory of a different computer with less latency.

Google tells that bare-metal servers are also good for agent sandboxes. These are isolated spaces for AI agents.

The numbers

All numbers are from Google. It gave no external test.

• Latency is as much as 60% less than with TCP networking.
• Random reads are 15.6M in one second.
• Storage speed is 70% more, and throughput is 110% more, than with the Z3 series.
• One server has a maximum of 192 vCPUs and 1,536 GiB of memory.

What is not known

Google gave no price and no test from a customer. These values are the maximum. Google does not tell where Z4M is open.

Insight: Google now gives storage with its AI processors

On 21 October 2026, Google Cloud told that its Rubin GPU servers open to all customers. On 5 November, it released a runtime, software that operates a model, as open software. Z4M adds storage that gives data to those processors. This reads as Google gives all the parts, from processors to storage to software, and each part has one vendor. Leaders must get the price of each part independently.