Google I/O 2026: Gemini 3.5, World Models, and $190B AI Infrastructure Bet

Gemini 3.5 Flash is 4x faster and 50-67% cheaper than comparable frontier models

Google positions Flash as an efficiency-first model optimized for speed, cost, and agentic tasks. It outperforms the previous flagship Gemini 3.1 Pro on nearly all benchmarks, with particular strength in coding and tool use — making it immediately competitive for developer workloads.

Gemini 3.5 Pro delayed to June pending safety review; will become new flagship

Sundar Pichai said the 3.5 series was specifically designed around tool use, instruction following, long-horizon tasks, and agent decoding. The safety hold on Pro suggests Google is applying more rigorous evaluation before releasing its most capable model publicly.

Gemini Omni is Google's entry into 'world models' — physically accurate multimodal generation

World models simulate real-world physical properties in generated media. Omni accepts and outputs video, images, text, and audio. A demonstrated use case: inserting a user's likeness into an existing video using high-level reasoning. Omni Flash launched at I/O; Omni Pro is in development.

Google expects to spend up to $190 billion on AI infrastructure in 2026

The bulk of this spend goes toward new data centers running Gemini models on Google's proprietary TPU chips. This level of capital expenditure is far beyond what any independent AI lab can sustain and represents Google's core structural advantage in the model race.

Google is now on its eighth generation of tensor processing units (TPUs)

Custom silicon is central to Google's cost and performance edge. Hundreds of thousands of TPUs power Gemini inference across Google's data center fleet, reducing dependence on third-party chip suppliers and enabling tighter hardware-software co-optimization.

Antigravity launched as an agent-first development platform for deploying personal AI agents

Antigravity is aimed at developers building agentic applications, reflecting Google's broader push to make agent deployment a first-class workflow rather than an experimental add-on.

Gemini Spark is an agentic assistant embedded in Google Cloud

Spark brings agentic capabilities directly into the Cloud console, positioning Google Cloud as a platform where AI agents are native rather than integrated through third-party tools.

Google Search updated with agentic coding to generate websites and apps on the fly

This extends Search beyond information retrieval into active creation — users can prompt Search to build functional web artifacts. It represents a direct expansion of Google's core product into AI-native use cases.

Android updated with a built-in AI agent integrated at the OS level

Embedding an AI agent in Android gives Google a distribution channel at massive scale — potentially hundreds of millions of devices — without requiring users to download a separate app. This is a structural advantage over standalone AI assistants.

Google's I/O strategy centers on leveraging search infrastructure advantages against newer AI rivals

Rather than competing solely on model benchmarks, Google is leaning into assets newer entrants lack: decades of search index data, global data center footprint, custom silicon, and distribution through Android and Chrome. The I/O announcements collectively reflect a strategy of making these assets the foundation of its AI products.

AI Mode in Google Search surpassed 1 billion monthly users, doubling every quarter

Google Search VP Liz Reid revealed at I/O 2026 that AI Mode usage has been doubling every quarter. Available at no charge to all Google Search users, removing payment friction and accelerating adoption. Each follow-up query in a conversation counts as a separate search, contributing to Google's reported search volume growth.