Pichai Prioritizes Compute for Advanced AI

Google CEO Sundar Pichai said in an interview with the Big Technology Podcast, published on the 20th local time, that Google prioritizes computing for advanced AI models even if it gives up some near-term cloud revenue. He said ensuring researchers have enough computing capacity is non-negotiable, and Google does adjust resources between its cloud business and AI research.

Google first ensures Google DeepMind has the computing needed for research and next-generation model development, then plans resources for other businesses. Pichai said AI supports Google businesses including Search and Cloud. He said Google shifted to an “AI first” approach in 2015 and considers both revenue from selling computing capacity to external customers and the competitiveness of models used across its businesses.

Google plans and secures cloud computing capacity over the long term based on customer demand, existing contracts and business forecasts. When resources are scarce, it can be difficult to meet both advanced model development needs and demand from other businesses, so the company may give up some near-term revenue opportunities. Google builds its own chips and data centers, creating a gap between investment decisions and when capacity becomes operational. Pichai said the computing resources Google is using now are the result of a decision he made in 2024.

Pichai said that around the same time last year, the rapid growth of coding agents had not yet begun, making demand several years ahead difficult to predict. “Not just us, every company is short on computing,” he said, adding that everyone would wish they had secured twice what they originally planned. Google continues investing based on estimates of its needs 2–3 years ahead, but technological progress can make current decisions turn out not to have been the best later. He said Google will maintain its principle of providing DeepMind researchers with top-tier computing despite that uncertainty.

Pichai said the cloud business expands the scale of Google’s operations by bringing a range of customers and applications onto its infrastructure, and also lets the company validate technology in real-world settings. His approach is to prioritize resources for research on Google’s own models while improving AI technology from chips to software through Cloud.

Pichai said Gemini 4 Argon processed about 25 trillion tokens in its first 72 hours after launch, three times the amount processed at the launch of Gemini 3.0. For internal development work, he said productivity rose 30% during the first 4 weeks of adoption, while rollbacks—reverting code to an earlier state because of errors or other issues—fell by one-third.

Pichai described recursive self-improvement (AI helping develop next-generation AI) as still being at an early stage. Researchers and AI agents work together on various model-development tasks, he said, but this remains far from a fully self-directed process in which AI continues development on its own.

Google CEO Sundar Pichai said in an interview with the Big Technology Podcast, published on the 20th local time, that Google prioritizes computing for advanced AI models even if it gives up some near-term cloud revenue. He said ensuring researchers have enough computing capacity is non-negotiable, and Google does adjust resources between its cloud business and AI research.

Google first ensures Google DeepMind has the computing needed for research and next-generation model development, then plans resources for other businesses. Pichai said AI supports Google businesses including Search and Cloud. He said Google shifted to an “AI first” approach in 2015 and considers both revenue from selling computing capacity to external customers and the competitiveness of models used across its businesses.

Google plans and secures cloud computing capacity over the long term based on customer demand, existing contracts and business forecasts. When resources are scarce, it can be difficult to meet both advanced model development needs and demand from other businesses, so the company may give up some near-term revenue opportunities. Google builds its own chips and data centers, creating a gap between investment decisions and when capacity becomes operational. Pichai said the computing resources Google is using now are the result of a decision he made in 2024.

Pichai said that around the same time last year, the rapid growth of coding agents had not yet begun, making demand several years ahead difficult to predict. “Not just us, every company is short on computing,” he said, adding that everyone would wish they had secured twice what they originally planned. Google continues investing based on estimates of its needs 2–3 years ahead, but technological progress can make current decisions turn out not to have been the best later. He said Google will maintain its principle of providing DeepMind researchers with top-tier computing despite that uncertainty.

Pichai said the cloud business expands the scale of Google’s operations by bringing a range of customers and applications onto its infrastructure, and also lets the company validate technology in real-world settings. His approach is to prioritize resources for research on Google’s own models while improving AI technology from chips to software through Cloud.

Pichai said Gemini 4 Argon processed about 25 trillion tokens in its first 72 hours after launch, three times the amount processed at the launch of Gemini 3.0. For internal development work, he said productivity rose 30% during the first 4 weeks of adoption, while rollbacks—reverting code to an earlier state because of errors or other issues—fell by one-third.

Pichai described recursive self-improvement (AI helping develop next-generation AI) as still being at an early stage. Researchers and AI agents work together on various model-development tasks, he said, but this remains far from a fully self-directed process in which AI continues development on its own.