Google's TPU & Agent Push
What happened
- Google unveiled new AI chips and doubled down on enterprise AI agents at its Cloud Next event. - The company split its TPU line into training and inference processors to boost efficiency and lower serving costs. - Google is pitching chips plus agents as a full-stack monetisation strategy while still supporting Nvidia hardware, per Reuters (reuters.com).
Why it matters
Google used its Cloud Next conference on April 22 to sell companies a bundled AI pitch: custom chips to run models and software agents to do work. (usnews.com) The event opened in Las Vegas on Wednesday, April 22, with Chief Executive Sundar Pichai and Google Cloud chief Thomas Kurian framing enterprise customers as the clearest source of AI revenue. Reuters reported that Alphabet still plans $175 billion to $185 billion in capital spending this year, and Pichai said just over half of machine-learning compute investment in 2026 is expected to go to the cloud business. (usnews.com) Google also reorganized its AI software under the Gemini Enterprise name and introduced a Gemini Enterprise Agent Platform for building, governing and optimizing large fleets of agents. Kurian said the question has shifted from whether companies can build an agent to how they manage thousands of them. (blog.google) An AI agent is software that can plan steps, use tools and carry out tasks with less human prompting than a chatbot. Reuters said Google is betting those agents will move from experiments into production systems for large businesses, even as safety, reliability and oversight remain open concerns. (usnews.com) The chip side of the strategy is about separating two expensive jobs: training a model and serving it after it is built. Google said its eighth-generation Tensor Processing Units now come as TPU 8t for training and TPU 8i for inference, or the fast response work needed when users or agents make requests. (cloud.google.com) Google said TPU 8t is built for frontier-model training, while TPU 8i is tuned for large-scale inference and reinforcement learning. In its product posts, the company said TPU 8t can scale to 9,600 chips in a superpod, while TPU 8i is designed for lower latency and higher memory bandwidth for serving many agent requests at once. (cloud.google.com) Pichai said Google’s first-party models now process more than 16 billion tokens per minute through direct customer API use, up from 10 billion last quarter. Google also said nearly 75% of Google Cloud customers use its AI products, and 330 customers processed more than 1 trillion tokens each over the past 12 months. (blog.google) Google is not asking customers to abandon Nvidia hardware. Reuters said the company presented its own chips as part of a broader cloud stack while continuing to support Nvidia’s processors, a practical stance in a market where many buyers want flexibility across models and infrastructure. (usnews.com) The company spent part of the conference showing how those agents could plug into existing business software instead of sitting in a demo. On April 22, Salesforce and Google Cloud said expanded integrations would let agents work across Slack, Google Workspace and Salesforce workflows with shared context. (salesforce.com) That leaves Google making a full-stack argument at the same moment OpenAI, Anthropic, Microsoft and Amazon are all chasing enterprise AI budgets. The closing message from Cloud Next was that Google wants to sell the model, the chip, the security layer and the agent that sits on top of all of them. (blog.google)
Key numbers
- Google used its Cloud Next conference on April 22 to sell companies a bundled AI pitch: custom chips to run models and software agents to do work.
- (usnews.com) The event opened in Las Vegas on Wednesday, April 22, with Chief Executive Sundar Pichai and Google Cloud chief Thomas Kurian framing enterprise customers as the clearest source of AI revenue.
- Reuters reported that Alphabet still plans $175 billion to $185 billion in capital spending this year, and Pichai said just over half of machine-learning compute investment in 2026 is expected to go to the cloud business.
- Google said its eighth-generation Tensor Processing Units now come as TPU 8t for training and TPU 8i for inference, or the fast response work needed when users or agents make requests.
What happens next
- Google used its Cloud Next conference on April 22 to sell companies a bundled AI pitch: custom chips to run models and software agents to do work.
- Reuters reported that Alphabet still plans $175 billion to $185 billion in capital spending this year, and Pichai said just over half of machine-learning compute investment in 2026 is expected to go to the cloud business.
- (blog.google) An AI agent is software that can plan steps, use tools and carry out tasks with less human prompting than a chatbot.
Quick answers
What happened in Google's TPU & Agent Push?
Google unveiled new AI chips and doubled down on enterprise AI agents at its Cloud Next event. The company split its TPU line into training and inference processors to boost efficiency and lower serving costs. Google is pitching chips plus agents as a full-stack monetisation strategy while still supporting Nvidia hardware, per Reuters (reuters.com).
Why does Google's TPU & Agent Push matter?
Google used its Cloud Next conference on April 22 to sell companies a bundled AI pitch: custom chips to run models and software agents to do work. (usnews.com) The event opened in Las Vegas on Wednesday, April 22, with Chief Executive Sundar Pichai and Google Cloud chief Thomas Kurian framing enterprise customers as the clearest source of AI revenue. Reuters reported that Alphabet still plans $175 billion to $185 billion in capital spending this year, and Pichai said just over half of machine-learning compute investment in 2026 is expected to go to the cloud business. (usnews.com) Google also reorganized its AI software under the Gemini Enterprise name and introduced a Gemini Enterprise Agent Platform for building, governing and optimizing large fleets of agents. Kurian said the question has shifted from whether companies can build an agent to how they manage thousands of them. (blog.google) An AI agent is software that can plan steps, use tools and carry out tasks with less human prompting than a chatbot. Reuters said Google is betting those agents will move from experiments into production systems for large businesses, even as safety, reliability and oversight remain open concerns. (usnews.com) The chip side of the strategy is about separating two expensive jobs: training a model and serving it after it is built. Google said its eighth-generation Tensor Processing Units now come as TPU 8t for training and TPU 8i for inference, or the fast response work needed when users or agents make requests. (cloud.google.com) Google said TPU 8t is built for frontier-model training, while TPU 8i is tuned for large-scale inference and reinforcement learning. In its product posts, the company said TPU 8t can scale to 9,600 chips in a superpod, while TPU 8i is designed for lower latency and higher memory bandwidth for serving many agent requests at once. (cloud.google.com) Pichai said Google’s first-party models now process more than 16 billion tokens per minute through direct customer API use, up from 10 billion last quarter. Google also said nearly 75% of Google Cloud customers use its AI products, and 330 customers processed more than 1 trillion tokens each over the past 12 months. (blog.google) Google is not asking customers to abandon Nvidia hardware. Reuters said the company presented its own chips as part of a broader cloud stack while continuing to support Nvidia’s processors, a practical stance in a market where many buyers want flexibility across models and infrastructure. (usnews.com) The company spent part of the conference showing how those agents could plug into existing business software instead of sitting in a demo. On April 22, Salesforce and Google Cloud said expanded integrations would let agents work across Slack, Google Workspace and Salesforce workflows with shared context. (salesforce.com) That leaves Google making a full-stack argument at the same moment OpenAI, Anthropic, Microsoft and Amazon are all chasing enterprise AI budgets. The closing message from Cloud Next was that Google wants to sell the model, the chip, the security layer and the agent that sits on top of all of them. (blog.google)
