Google I/O '26 Keynote

**TL;DR:** Google I/O '26 keynote showcased AI acceleration across the board: 3.2 quintillion tokens processed monthly, Gemini monthly active users surpassing 900 million, new TPU chips and world model Gemini Omni announced, along with conversational AI features like Ask YouTube and Docs Live. ## Opening & 10-Year AI Vision Sundar Pichai kicked off the keynote with a personal anecdote. He noted that this year marks the tenth anniversary of the company's AI-first strategy, and that AI is becoming the ultimate tool for solving the most complex scientific problems. Google takes a differentiated, full-stack approach to AI innovation—covering everything from custom chips and secure infrastructure to models, products, and platforms. This approach accelerates innovation and illuminates every corner of the company. ## Scale Growth: Tokens, Developers, and Users Over the past two years, the token volume processed by Google services has surged from 9.7 trillion per month to 480 trillion, and then multiplied sevenfold again to 3.2 quintillion per month. Currently, over 8.5 million developers use Google's models each month to build new applications, with the model API processing roughly 19 billion tokens per minute. In the past 12 months, more than 375 customers each processed over 1 trillion tokens. On the product side, Google has 13 products each with over 1 billion users, 5 of which have over 3 billion users. Monthly active users of the Gemini app grew from 400 million last year to 900 million, with daily requests increasing more than sevenfold. Additionally, over 50 billion images have been generated using the Nano Banana model. ## Search & Gemini App Upgrades AI Overviews have over 2.5 billion monthly active users, while AI Mode is described as the biggest upgrade to Search ever, with over 1 billion monthly active users. When users engage with AI features, they search more frequently, and the search experience shifts from single queries to ongoing conversations. The Gemini app now includes personalized intelligence features, making responses more tailored. Sundar mentioned using Gemini personally to understand his parents' medical records. ## New Conversational AI Features Three major products received conversational AI upgrades: - **Ask Maps**: Supports complex long-form questions, e.g., "My kid just fell into a duck pond, wedding starts in half an hour, where can I walk to buy her a new dress?" - **Ask YouTube**: Supports natural language queries, e.g., "Teach a 3-year-old to ride a balance bike when they already know how to pedal." The system provides an overview, tips, and jumps directly to the most relevant part of the video, while remembering context for follow-ups and table comparisons. Rolling out broadly in the US this summer. - **Docs Live**: Users can dictate ideas by voice; Gemini automatically pulls resumes from Drive, extracts email details, and generates document drafts. In a demo, a user requested talking points for a career day speech—Gemini not only pulled the resume but also extracted info from a "Career Day Logistics" email, formatted comparisons as a table, and even added personal story notes automatically. Docs Live launches this summer for Pro and Ultra subscribers, with similar features coming to Gmail and Google Keep afterward. ## Infrastructure Investment & TPU Chips Google's 2022 capital expenditure was $31 billion; this year it is expected to reach $180–190 billion. A key investment is custom silicon. Following the eighth-generation TPU, Google has adopted a dual-chip design for the first time: - **TPU 8t**: Optimized for large-scale pre-training, offering nearly three times the raw compute of the previous generation. With JAX and Pathways, training is no longer limited to a single data center—it can be distributed across over 1 million TPUs globally, enabling training of large models in weeks rather than months. - **TPU 8i**: Optimized for inference with significantly reduced latency. In a live demo, a Flash model running on 8i generated a Chrome dinosaur game, with token generation speed approaching 1,500 tokens per second. Both chips deliver up to 2x power efficiency improvement. The keynote featured a short animated TPU video showing the chips at work in tasks like protein folding and climate simulation. ## Model Breakthrough: Gemini Omni Demis Hassabis announced the Gemini Omni series, a world model that combines Gemini's multimodal capabilities with generative media models (Veo, Nano Banana, Genie). Omni can generate any output from any input, achieving a step-change in simulating physical concepts like motion and gravity. In a demo, the input "Make a claymation video explaining protein folding" resulted in an animation showing amino acid chains, alpha helices, and beta sheets. More importantly, Omni supports conversational video editing: users can modify details in a selfie, change styles, or add elements. A simple circle can morph into a black hole; an evening walk scene can be adjusted in real time. The first model, Gemini Omni Flash, is available in products starting today. More details on Omni Pro will be shared in the future. ## Transparency & SynthID In response to deepfake risks, Google's SynthID watermarking has marked 100 billion images and videos, plus audio assets equivalent to 60,000 years. Millions of users verify AI-generated content through detectors in the Gemini app. Google is now also adding content credential verification across products, showing whether content is AI-generated and its provenance. --- Source: https://www.youtube.com/live/wYSncx9zLIU