Google + Nvidia tie-up
What happened
- Nvidia and Google Cloud announced tighter collaboration to let companies build end-to-end “AI factories” combining chips, software, and agents. - The stack links Rubin-powered A5X instances, confidential Blackwell GPUs, and Gemini deployed on Google Distributed Cloud. - The partnership deepens Nvidia’s role across cloud, edge and confidential deployments, widening enterprise deployment options (blogs.nvidia.com).
Why it matters
Google Cloud and Nvidia used Google Cloud Next on April 22 to widen a joint AI stack aimed at companies that want to build and run AI systems on one supplier-backed setup. (blogs.nvidia.com) The new package starts with Google’s A5X bare-metal instances, which Nvidia said use Vera Rubin systems and can scale to nearly 1 million Rubin graphics processing units across sites. Google and Nvidia said the design targets large training and inference jobs, the computing step that generates answers after a model is built. (blogs.nvidia.com) The announcement also adds confidential virtual machines with Nvidia Blackwell graphics processing units on Google Cloud, plus a preview of Gemini on Google Distributed Cloud running on Blackwell and Blackwell Ultra hardware. Google said Google Distributed Cloud is meant for customers that need to keep data in specific locations, including on-premises or sovereign environments. (blogs.nvidia.com) (cloud.google.com) An AI “factory” is vendor shorthand for the full chain: chips, networking, software, models and deployment tools bundled so a company can train, serve and manage AI in one system. Google framed Next ’26 around the “Agentic Enterprise,” and launched Gemini Enterprise Agent Platform on April 22 as its control layer for building and governing software agents. (cloud.google.com 1) (cloud.google.com 2) That framing lines up with Nvidia’s push to sell more than chips. In the Google tie-up, Nvidia paired its Nemotron open models and NeMo development tools with Google’s Gemini Enterprise Agent Platform, extending Nvidia further into the software layer that sits above cloud hardware. (blogs.nvidia.com) Google and Nvidia had already been expanding that partnership before this week. At Nvidia GTC in March 2026, Google Cloud said it was adding broader support for Nvidia software, flexible G4 virtual machines using RTX PRO 6000 Blackwell hardware, and upcoming support for Vera Rubin NVL72 systems. (cloud.google.com) The security piece is aimed at regulated buyers that have been slower to move sensitive workloads onto shared cloud infrastructure. Nvidia said the Blackwell announcement is the first confidential computing offering for Blackwell graphics processors in the cloud, while Google has been pitching confidential computing as a way to protect data while it is being processed, not just stored. (blogs.nvidia.com) (cloud.google.com) The hardware claims are also part of a cost fight. Google said at Next ’26 that A5X was co-designed with Nvidia and can scale to 960,000 Rubin graphics processors, while outside coverage of the launch cited Google’s claim of up to 10 times lower inference cost per token and 10 times higher token throughput per megawatt than the prior generation. (msn.com) (storagenewsletter.com) For enterprise buyers, the practical change is more deployment choice inside one Google-Nvidia lane: public cloud for scale, confidential cloud for sensitive workloads, and distributed cloud for on-premises or sovereign setups. The pitch from both companies is that customers can move the same AI work across those environments without rebuilding the whole stack each time. (blogs.nvidia.com) (cloud.google.com)
Key numbers
- The stack links Rubin-powered A5X instances, confidential Blackwell GPUs, and Gemini deployed on Google Distributed Cloud.
- Google Cloud and Nvidia used Google Cloud Next on April 22 to widen a joint AI stack aimed at companies that want to build and run AI systems on one supplier-backed setup.
- (blogs.nvidia.com) The new package starts with Google’s A5X bare-metal instances, which Nvidia said use Vera Rubin systems and can scale to nearly 1 million Rubin graphics processing units across sites.
- Google framed Next ’26 around the “Agentic Enterprise,” and launched Gemini Enterprise Agent Platform on April 22 as its control layer for building and governing software agents.
What happens next
- Google Cloud and Nvidia used Google Cloud Next on April 22 to widen a joint AI stack aimed at companies that want to build and run AI systems on one supplier-backed setup.
- Google and Nvidia said the design targets large training and inference jobs, the computing step that generates answers after a model is built.
- Google framed Next ’26 around the “Agentic Enterprise,” and launched Gemini Enterprise Agent Platform on April 22 as its control layer for building and governing software agents.
Quick answers
What happened in Google + Nvidia tie-up?
Nvidia and Google Cloud announced tighter collaboration to let companies build end-to-end “AI factories” combining chips, software, and agents. The stack links Rubin-powered A5X instances, confidential Blackwell GPUs, and Gemini deployed on Google Distributed Cloud. The partnership deepens Nvidia’s role across cloud, edge and confidential deployments, widening enterprise deployment options (blogs.nvidia.com).
Why does Google + Nvidia tie-up matter?
Google Cloud and Nvidia used Google Cloud Next on April 22 to widen a joint AI stack aimed at companies that want to build and run AI systems on one supplier-backed setup. (blogs.nvidia.com) The new package starts with Google’s A5X bare-metal instances, which Nvidia said use Vera Rubin systems and can scale to nearly 1 million Rubin graphics processing units across sites. Google and Nvidia said the design targets large training and inference jobs, the computing step that generates answers after a model is built. (blogs.nvidia.com) The announcement also adds confidential virtual machines with Nvidia Blackwell graphics processing units on Google Cloud, plus a preview of Gemini on Google Distributed Cloud running on Blackwell and Blackwell Ultra hardware. Google said Google Distributed Cloud is meant for customers that need to keep data in specific locations, including on-premises or sovereign environments. (blogs.nvidia.com) (cloud.google.com) An AI “factory” is vendor shorthand for the full chain: chips, networking, software, models and deployment tools bundled so a company can train, serve and manage AI in one system. Google framed Next ’26 around the “Agentic Enterprise,” and launched Gemini Enterprise Agent Platform on April 22 as its control layer for building and governing software agents. (cloud.google.com 1) (cloud.google.com 2) That framing lines up with Nvidia’s push to sell more than chips. In the Google tie-up, Nvidia paired its Nemotron open models and NeMo development tools with Google’s Gemini Enterprise Agent Platform, extending Nvidia further into the software layer that sits above cloud hardware. (blogs.nvidia.com) Google and Nvidia had already been expanding that partnership before this week. At Nvidia GTC in March 2026, Google Cloud said it was adding broader support for Nvidia software, flexible G4 virtual machines using RTX PRO 6000 Blackwell hardware, and upcoming support for Vera Rubin NVL72 systems. (cloud.google.com) The security piece is aimed at regulated buyers that have been slower to move sensitive workloads onto shared cloud infrastructure. Nvidia said the Blackwell announcement is the first confidential computing offering for Blackwell graphics processors in the cloud, while Google has been pitching confidential computing as a way to protect data while it is being processed, not just stored. (blogs.nvidia.com) (cloud.google.com) The hardware claims are also part of a cost fight. Google said at Next ’26 that A5X was co-designed with Nvidia and can scale to 960,000 Rubin graphics processors, while outside coverage of the launch cited Google’s claim of up to 10 times lower inference cost per token and 10 times higher token throughput per megawatt than the prior generation. (msn.com) (storagenewsletter.com) For enterprise buyers, the practical change is more deployment choice inside one Google-Nvidia lane: public cloud for scale, confidential cloud for sensitive workloads, and distributed cloud for on-premises or sovereign setups. The pitch from both companies is that customers can move the same AI work across those environments without rebuilding the whole stack each time. (blogs.nvidia.com) (cloud.google.com)
