When I interviewed Koray Kavukcuoglu of Google last week, I had a strong sense of déjà vu.

This was Kavukcuoglu's first in-depth interview since becoming CEO of Google DeepMind. In the conversation, he gave his most detailed remarks to date on the direction of the company's robotics business and how Google's approach differs from other players in the industry.

Figure, backed by OpenAI, and Tesla Motors (TSLA), led by Elon Musk, are both going all out to build a flagship humanoid robot; but Kavukcuoglu said Google's core thinking is essentially a bet on software and intelligence capabilities.

I pressed him on whether Google plans to build its own humanoid robots. He did not completely rule out the possibility, but made clear that Google's competitive advantage lies in the model layer, not in robot body hardware.

This scene instantly took me back to 2007 or 2008, when I interviewed Google co-founder Sergey Brin and asked whether Google would make phones. At the time, Google was working with T-Mobile and HTC to build Android devices, but would Google make its own phone? Brin, rollerblading through the Google cafeteria (no exaggeration), told me: only if absolutely necessary. That remark laid the groundwork for Google's own phones. Later, Google first partnered with HTC, then gradually launched the Pixel series of phones on its own. If I had to bet, I believe that far in the future, Google will likewise move toward vertical integration and launch its own humanoid-class robots.

Now Google is replicating the first step of the Android model: external partnerships. Google launched Gemini Robotics, a version based on the core Gemini large model specifically adapted for controlling physical devices. Google's strategy is to deploy this intelligence capability onto various robots produced by third-party manufacturers such as Boston Dynamics, rather than placing all its chips on a single self-developed hardware platform.

Speaking at The Information's AI Agenda live event last week, Kavukcuoglu said: "We want to work with other robotics companies so they can use this capability to empower all kinds of robots with AI intelligence in a reliable and safe way." (You can watch the full in-depth interview.)

Kavukcuoglu also tied the robotics business layout to Google's grand goal of artificial general intelligence: this multimodal Gemini architecture, which handles text, images, and audio, must also have generalization ability in the real physical world. In other words, pedestrian recognition, translation, and robot control would all be handled by one system, rather than developing a separate specialized model for each task.

He also cited Waymo as the most powerful current example, calling autonomous driving a "very safe proving ground" that can validate physical-world artificial intelligence capabilities at scale.

Speaking of Gemini, Google opened Gemini 4 Argon to a small group of customers on Wednesday, mainly for network defense; a full-scale public release is still some time away. This means we will soon see whether Google has narrowed the gap with the two leading AI companies, Anthropic and OpenAI, in overall general large-model strength.