NVIDIA–Google Cloud Expansion

NVIDIA and Google Cloud widened their Blackwell lineup on April 22, adding new virtual machines for training, serving, and smaller slice-by-slice GPU rentals. (blogs.nvidia.com) The new menu runs from A4 virtual machines built on NVIDIA HGX B200 systems to rack-scale A4X machines with GB200 NVL72, A4X Max systems with GB300 NVL72, and fractional G4 instances using RTX PRO 6000 Blackwell Server Edition GPUs. Google Cloud said the additions are aimed at frontier models, agentic artificial intelligence systems, and physical artificial intelligence workloads such as robotics and simulation. (blogs.nvidia.com) Google had already been rolling out parts of that stack. In March, the company said G4 virtual machines with RTX PRO 6000 Blackwell GPUs were gaining traction, and in November it said A4X Max was shipping in production with 72 Blackwell Ultra GPUs and 36 Grace central processors in each GB300 NVL72 system. (cloud.google.com, cloud.google.com) The product names map to different jobs. A4 and A4X are built for large training runs and high-volume inference, while G4 is pitched for latency-sensitive applications, visual computing, and cases where customers want less than a full server’s worth of accelerator capacity. (docs.cloud.google.com, cloud.google.com) That matters because cloud customers are no longer buying only one kind of artificial intelligence compute. They are mixing giant clusters for model training with smaller pools for inference, simulation, and software agents that need fast responses at lower cost. (blogs.nvidia.com, cloud.google.com) The catch is supply. Jensen Huang said this week that NVIDIA does not raise prices and does not sell chips to the highest bidder during shortages, a sign that allocation still matters even as Blackwell reaches volume production. (247wallst.com, blogs.nvidia.com) Google’s own documentation reflects that spread of hardware rather than a single default machine. Its accelerator-optimized catalog now spans A2, A3, A4, A4X, A4X Max, and G4 families, with different recommendations for pre-training, fine-tuning, inference, and graphics-heavy work. (docs.cloud.google.com, docs.cloud.google.com) NVIDIA and Google have been building toward this for more than a year. At NVIDIA’s March 2025 GTC announcements, Google Cloud said it would be among the first providers to offer GB300 NVL72 and RTX PRO 6000 Blackwell systems, and NVIDIA later said Google was first to offer both HGX B200 and GB200 NVL72 through A4 and A4X virtual machines. (nvidianews.nvidia.com, blogs.nvidia.com) For customers, the expansion means more ways to get Blackwell on Google Cloud, but not a world where every workload lands on the same scarce chip. The practical plan is a mixed fleet: big systems where scale is essential, and smaller or fractional instances where availability and cost decide the job. (blogs.nvidia.com, docs.cloud.google.com)