Google plans to merge Gemini and Veo

If you want an artificial intelligence to actually understand gravity, friction, and how objects move, feeding it text books will only get you so far. It needs to watch the world in action.

Google is preparing to do just that by combining two of its most powerful systems. DeepMind CEO Demis Hassabis revealed on the Possible podcast, hosted by LinkedIn co-founder Reid Hoffman, that Google eventually plans to merge its Gemini models with its Veo video-generation models to dramatically improve its understanding of the physical world.

Learning physics from endless YouTube videos

Hassabis explained that Gemini was designed from day one as a multimodal system. The long-term vision is to create a universal digital assistant that can actually help people navigate tasks in the real world.

To achieve that, the model has to grasp basic physical common sense. Hassabis implied that Veo’s massive video training diet comes primarily from YouTube, which Google owns. By simply watching a staggering amount of YouTube videos, Hassabis noted, Veo 2 has essentially started to figure out real-world physics.

That reliance on video data naturally raises old questions around training practices. Google previously told TechCrunch that its models may be trained on some YouTube content in line with creator agreements, and the company reportedly updated its terms of service last year partly to unlock more data for training.

The industry-wide rush toward omni models

Google is not moving in isolation. The frontier AI race is rapidly shifting toward omni models that can ingest, understand, and output virtually any medium.

Google’s newest Gemini models already generate audio, images, and text. OpenAI’s default ChatGPT model can natively generate visual art. Amazon has also announced plans to launch an any-to-any multimodal model later this year.

By folding video generation directly into Gemini, Google wants its flagship model to simulate how the physical environment evolves over time, rather than just predicting text tokens.

Why it matters

We think this combination plays directly into Google’s single biggest unfair advantage: YouTube.

For everyday users and developers, this shift is the missing bridge between chat software and true physical assistance. An assistant that understands spatial depth, trajectory, and basic mechanics could eventually power real-world robotics, troubleshoot hands-on repairs, or navigate complex physical spaces without hallucinating impossible movements.

Against rivals like OpenAI and Amazon, Google holds the ultimate multimodal library. While everyone is scraping the bottom of the barrel for public web text, video represents the next frontier of training data, and nobody owns more of it than Google.

Still, the approach carries real friction. Fusing massive video generation weights with reasoning models will demand immense compute, making inference expensive. Google will also face lingering scrutiny from creators over how their YouTube uploads are being repurposed to build proprietary world models. Google is betting that watching YouTube can teach AI how the universe works, and the rest of the industry will have to prove they can match that data pipeline.