The history and future of AI at Google, with Sundar Pichai

Sundar Pichai has now spent a decade as CEO of Google, and Alphabet is planning roughly $180 billion in capital expenditure in 2026 alone. In a wide-ranging conversation, he addressed the most persistent criticism of Google’s AI strategy—that the company invented the transformer architecture but let others productize it—and argued that the story is more nuanced than the conventional narrative suggests.

The transformer was not a pure research project. It was developed to solve a specific product need: improving translation quality at scale. And it was immediately productized inside Google. The biggest jumps in search quality during that period came from BERT and MUM, both transformer-based models. Google also had an internal product called LaMDA—an early version of what became ChatGPT—but held it back because the company had a higher quality bar and hadn't yet solved for RLHF. Pichai points out that even OpenAI's ChatGPT launch was a "buried launch" the week of Thanksgiving, not a grand unveiling. His broader point: in consumer internet, surprises happen. Google acquired YouTube after Google Video failed. Facebook bought Instagram. These moments are not signs of strategic failure; they are the nature of the business.

Speed as a design principle