Sundar Pichai Opening Remarks | I/O 2026 Keynote

TL;DR: Sundar Pichai at I/O 2026 announced AI monthly token processing reached 32 quadrillion, unveiled next-gen TPUs, the Gemini Omni model, and multiple product updates, emphasizing that AI is driving scale through full-stack innovation. ## Opening: A Year Driven by AI Sundar Pichai kicked off with a humorous space capsule segment, noting the past year was an "intense, non-stop launch period." He reviewed Google's decade-long transformation into an AI-first company, highlighting its differentiated full-stack AI innovation approach: from custom chips and security foundations to world-class research and models, to product platforms reaching billions of users. This approach allows the company to "iterate and innovate faster, and illuminate every corner." ## AI Adoption Scale: From Tokens to Users ### Explosive Token Growth Pichai showed concrete evidence of widespread AI adoption — token processing volumes: - Two years ago: 9.7 trillion tokens per month - Last I/O: ~480 trillion tokens - Now: **32 quadrillion tokens/month** (7x growth) He remarked: "I never thought I'd say 'quadrillion' in an I/O keynote, but that's the reality." ### Developer and Product Demand - **8.5 million developers** build apps with Google models each month - Model API processes **19 billion tokens per second** - Over the past 12 months, **more than 375 customers** each processed over 1 trillion tokens - Google has **13 products** with over 1 billion users, 5 of which exceed 3 billion ### Search and Gemini App - **AI Overviews**: over 2.5 billion users monthly - **AI Mode**: surpassed 1 billion monthly active users within a year, "the biggest upgrade to Search ever" - **Gemini App**: MAU grew from 400 million at last I/O to **900 million**, doubling in a year; daily requests grew 7x+ - **Nano banana model**: generated over 50 billion images ## Conversational AI Deeper into Products ### Maps: Biggest Upgrade in a Decade Added "Maps" feature supporting complex long queries. Pichai gave a real example: "My child just fell into a duck pond and we have a wedding in 30 minutes. Where can I buy her a new dress?" ### Ask YouTube Addressing user confusion with massive video libraries, a new feature can answer questions like "How do I teach a three-year-old to ride a bike?" It provides an overview, practical tips, related videos, and automatically jumps to the most relevant parts; also supports follow-ups (e.g., "Should I get a hand brake or foot brake?") and presents comparison info in a table. **Rolling out widely in the US this summer.** ### Docs Live: Voice-Created Documents Users can voice their ideas, and Gemini automatically creates the document. In a demo, a user prepared points for a high school career day speech, asked to pull a resume from Drive, extract school emails, add an analogy table, and personal story. Pichai emphasized all demos were live, unaccelerated. **Available to Pro and Ultra subscribers this summer**, also arriving in Gmail and Google Keep. ## Infrastructure Investment: Custom Chips and Global Training ### Capex Surge - 2022: $31 billion - 2026 estimated: **$180-190 billion** (~6x) ### 8th Gen TPU: TPU 80 and TPU 80i - Dual-chip approach, optimized for training and inference respectively - **TPU 80**: for large-scale pre-training, raw compute nearly **3x** previous generation - **Jackson Pathways**: training can scale to **100k+ TPUs** across multiple sites, enabling the world's largest training clusters, reducing training cycles from months to weeks - **TPU 80i**: inference-specific, with drastically reduced latency. Live demo of Flash model generating Chrome Dino game at **nearly 1500 tokens per second** - Energy efficiency: up to **2x** performance per watt ### Humorous TPU Behind-the-Scenes A personified dialogue showed TPUs' "daily work" — processing protein folding, simulating climate data, generating pug images, and joking about "trillions of tasks to do." ## World Model: Gemini Omni ### From Predicting Text to Simulating Reality Demis Hassabis appeared, noting that last year's I/O vision of a world model has entered a new phase. He announced **Gemini Omni**: a model that "creates any output from any input," combining Gemini intelligence with generative media models like Veo, Nano banana, and Genie, achieving "step changes in simulating kinetic energy and gravity." ### Key Features - Generates highly accurate videos (e.g., "claymation explaining protein folding") - Supports conversational, language-based iterative video editing (input selfie, freely adjust style, add elements) - "Anything can become a canvas for creating entirely new realities" - First models: **Gemini Omni Flash** (already in products); Omni Pro coming soon ## AI Transparency: SynthID and Content Credentials ### Challenge of Identifying Deepfakes Research shows people correctly identify high-quality deepfake videos only about a quarter of the time. ### Google's Solutions - **SynthID**: invisible watermark, already applied to **over 100 billion images and videos**, equivalent to **60,000 years of audio** - **SynthID Detector**: millions use it in Gemini apps - **Content Credentials verification**: shows if content is AI-generated, camera-captured, or GenAI-edited - **Expanded to Search and Chrome**: users can circle or right-click to ask "Is this AI-generated?" - **Industry collaboration**: NVIDIA joined last year; **OpenAI, Kakao, ElevenLabs** are adopting SynthID 2 ## Looking Ahead Pichai concluded: "That's the progress we've made on world models. Now let's talk about what's next for our Gemini 3 series." The talk ended after mentioning Gemini 3 is already released, hinting at more details to come. --- Source: Sundar Pichai Opening Remarks | I/O 2026 Keynote (https://www.youtube.com/watch?v=duHhImuaZGU)