Google Cloud is now offering instances built on the Nvidia GB300 NVL72 system.

The A4X Max instances each feature 72 Blackwell Ultra GPUs and 36 Nvidia Grace CPUs connected by Nvidia NVLink.

The instances also use Google's Titanium ML adapter and Google Cloud's Jupiter network fabric, with the instances built to be able to scale to tens of thousands of GPUs.

According to Google, the A4X Max delivers 2x the network bandwidth of the A4X, which was based on the Nvidia GB200 NVL72.

Each GB300 NVL72 offers 1.4 exaflops of compute, and a four times increase in LLM training and serving performance compared to VMs with Nvidia H100 GPUs.

Google's launch of GB300-powered instances comes just a few weeks after Microsoft revealed it had deployed a cluster of 4,600 GB300 NVL72s. Microsoft claims it is the first to offer a cluster "at scale."

AI cloud provider Lambda revealed last month that it had deployed at least one GB300 NVL72 system.

Google recently published its quarterly earnings, with cloud revenues of $15.2bn, up 34 percent Year-over-Year.

Capex for the quarter was $24 billion, of which around 60 percent went on servers.