Google built I/O 2026 with its own generative stack — and the hard part was the glue

What Google actually built

For the "TPU Training Day" short film, Nexus Studios captured performances with puppetry and 3D animation, then used Nano Banana to generate stylized first frames through a custom tool built in Google AI Studio to enforce pixel-perfect consistency, before Gemini Omni merged the base animation with the stylized frames. The event's brand identity came from feeding Gemini five years of past I/O recaps and iterating imagery in Nano Banana.

"Jellectronica," built with the Monterey Bay Aquarium, trained a YOLO8 model in Google Colab and deployed it on a Coral NPU to track moon-jellyfish movement, driving Google Flow Music and Lyria — with a mass stem generator built in Antigravity to automate bass, chords, melody and drums. The Antigravity Coffee Co. pop-up used Flutter with the A2UI protocol for real-time adaptive interfaces, wiring Firebase to Nano Banana so attendees could design and order custom latte art.

The hard problem was consistency, not capability

Across every project, the difficult work was consistency: a bespoke AI Studio tool to keep generated frames pixel-stable, a stem generator to make music production repeatable, and an adaptive-UI protocol to make generative interfaces predictable. The general-purpose models were the starting point; the production-grade output came from custom orchestration built on top of them.

Consistency tooling is the transferable lesson

When done right, the event is amazing on its own, and, as a viewer, you stop thinking about how AI was used.Google

Teams wiring several generative models into one product hit exactly this wall: any single model call is straightforward, but stitching many into a repeatable pipeline that does not drift is where the real work lives. The reusable takeaway from I/O is not the model list — it is that shipping generative output at quality means investing in consistency tooling and human review at every seam.