Google funds Nvidia expansion

Google Cloud is putting $750 million behind the consultants and hardware needed to push more corporate artificial intelligence projects into production. (cloud.google.com) The fund, announced April 22 at Cloud Next 2026 in Las Vegas, is aimed at global consulting firms, software partners and channel partners building “agentic” systems — software that can plan and carry out multistep tasks with limited human input. Google said the program will support deployments through firms including Accenture, Deloitte, KPMG, McKinsey and PwC. (cloud.google.com, bloomberg.com) Google paired that services push with more spending on compute. At the same event, it rolled out new Google Cloud capacity built around Nvidia’s Blackwell chips and introduced eighth-generation Tensor Processing Units, with one version for training large models and another for serving answers to users. (blog.google, blog.google, marketwatch.com) That pairing reflects how the cloud market is working in 2026. The biggest providers still need Nvidia’s graphics processors to satisfy near-term demand, but they are also building in-house chips to lower costs, control supply and tune performance for their own software stacks. (blog.google, blogs.nvidia.com) For customers, the bottleneck is no longer just buying a model. Companies now need consultants to redesign workflows, connect data systems and govern deployments, which is why Google is subsidizing partners that can turn cloud capacity into billable enterprise projects. (cloud.google.com, thenextweb.com) Google said Accenture has already built more than 450 agents on Google Cloud. KPMG committed $100 million to its Google Cloud partnership, while PwC committed $400 million and NTT DATA said it would dedicate 5,000 engineers to agentic artificial intelligence work on the platform. (thenextweb.com, world-today-journal.com) On the infrastructure side, Google has been widening AI Hypercomputer, its package of chips, networking and software for training and inference. In 2025 it introduced Ironwood as a seventh-generation TPU built specifically for inference, and this week it split the next generation into TPU 8t for training and TPU 8i for inference. (blog.google, blog.google) Google and Nvidia have also been tightening their ties outside the public cloud. At Cloud Next 2025, the companies said enterprises would be able to run Google’s Gemini models on-premises through Google Distributed Cloud on Nvidia Blackwell systems, a pitch aimed at customers with security and data-sovereignty requirements. (blogs.nvidia.com, fierce-network.com) The result is a two-track strategy: buy more Nvidia capacity where demand is immediate, and build more Google silicon where control and margins matter most. Google’s latest fund suggests it thinks the next cloud sale will depend as much on consultants shipping working agents as on chips in the data center. (cloud.google.com, blog.google)