Google Cloud doubles down with NVIDIA

Google Cloud previewed flexible, fractional G4 VMs using NVIDIA vGPU for the RTX PRO 6000 Server Edition — presented as the first fractional-vGPU preview for that specific GPU on a major cloud. (cloud.google.com) Google Cloud said it will integrate NVIDIA’s Dynamo inference stack into GKE Inference Gateway and expand NVIDIA support across Vertex AI Training and Model Garden to tie Rubin hardware into its managed ML tooling. (cloud.google.com) NVIDIA’s NVL72 “Vera Rubin” rack pairs 72 Rubin GPUs with 36 Vera CPUs and is specified at roughly 3.6 exaFLOPS NVFP4 inference and 2.5 exaFLOPS NVFP4 training, with 20.7 TB of HBM4 and 54 TB of LPDDR5X memory per rack and ~260 TB/s scale-up bandwidth via NVLink 6. (nvidia.com)(videocardz.com) NVIDIA’s slides claim Rubin can cut inference token costs by up to 10x versus Blackwell and reduce GPU counts for MoE training by about 4x, positioning NVL72 as a cost/efficiency pivot for production inference and sparse-model training. (videocardz.com) Microsoft Azure reported being the first cloud to bring up and begin validating an NVL72 system — a validation Microsoft executives said demonstrates early cloud availability of Rubin racks — while NVIDIA says Rubin-based products will be available from partners starting in the second half of 2026. (msn.com)(nvidianews.nvidia.com) OEMs and system builders are shipping liquid-cooled and rack-scale designs to support NVL72 deployments, with Supermicro expanding liquid-cooling production capacity and vendors like ASUS announcing liquid-cooled Rubin-based systems to meet the thermal and power demands of exascale inference racks. (rallies.ai)(marketwatch.com)