THE SCALE NUMBERS
(up from 9.7T two years ago)
via Gemini API
(doubled in 12 months)
(6x 2022 spend)
on Google models monthly
users
These aren’t vanity metrics. They describe the throughput floor of an infrastructure layer that everything else at I/O was built on top of.
MODELS: FLASH GETS FAST, OMNI GETS PHYSICAL
Gemini 3.5 Flash is live as of May 19. It outperforms 3.1 Pro across nearly all benchmarks, with a notable jump on GDPVal (economically valuable real-world tasks). It runs 4x faster on output tokens per second than comparable frontier models — 12x faster inside Antigravity specifically — at less than half the cost.
Gemini 3.5 Pro is in internal testing, slated for June. Described as combining frontier intelligence with stronger agentic capabilities. No benchmarks shared.
Gemini Omni Flash is the more structurally interesting model. Demis Hassabis introduced it as capable of generating output in any modality from any input. The key distinction: it was specifically trained on physics simulation. Understanding of gravity, kinetic energy, and fluid dynamics is a stated training objective — not a feature bolted on afterward.
TPU 8: THE ARCHITECTURE SPLITS
For the first time, Google has divided TPU into two distinct chips. TPU 8t is optimized for training: JAX and Pathways now distribute training across multiple data centers simultaneously, scaling to over 1 million TPUs globally. TPU 8i is optimized purely for inference latency. Both deliver up to 2x better performance-per-watt versus prior generation.
GEMINI SPARK: A PERSISTENT AGENT OUTSIDE YOUR SESSION
Spark is Google’s personal AI agent. The mechanics matter more than the pitch: it runs on dedicated virtual machines on Google Cloud, operates 24/7 without a laptop open, is powered by Gemini 3.5 Flash and the Antigravity harness for long-horizon background tasks, and is accessible via the Gemini app, email, and chat — meaning Spark is an addressable agent endpoint, not just a UI element.
Beta is live now for AI Ultra US subscribers, rolling out to Gmail, Chat, then Chrome.
SEARCH: THE QUIET TRANSFORMATION
The new search box accepts text, images, files, video, audio, PDFs, and your entire Chrome tab as input. It adapts contextually. Rolling out today in all countries and languages where AI Mode is available.
AI Overviews and AI Mode are merging into one unified surface. The underlying architecture is being consolidated so a single model handles both.
Information Agents are persistent background watchers for topics you define: stock changes, sneaker drops, website changes, job listings. They send alerts. Rolling out this summer.
Generative UI dynamically builds layouts, tables, and interactive widgets on the fly inside search results. Free for everyone.
Personal Intelligence in AI Mode is expanding to nearly 200 countries across 98 languages — free, no subscription required.
UNIVERSAL CART AND THE TWO PROTOCOLS NOBODY DISTINGUISHED
The Shopping blog post describes two architecturally distinct systems that every summary conflated.
UCP vs AP2
Universal Commerce Protocol (UCP) is a shared checkout language across merchants — it allows the cart to read product data, pricing, inventory, and compatibility from any participating retailer.
Agent Payments Protocol (AP2) is separate: a protocol for agent-executed purchases using privacy-preserving technology, creating a tamper-proof mandate with a permanent digital paper trail. AP2 hits Gemini Spark first; UCP is the merchant-side protocol.
Launching US this summer. Compatible retailers at launch: Target, Walmart, Sephora, Wayfair, Nike, and Shopify merchants.
CHROME: THE 15 UPDATES THAT MATTER
The Chrome changes are more architecturally significant than anything in the consumer keynote.
| Feature | What It Does | Status |
|---|---|---|
| WebMCP | Open web standard exposing JS functions as machine-readable tools for browser agents | Chrome 149 origin trial |
| Modern Web Guidance | Expert-vetted agent skills covering 100+ web dev use cases | Live |
| DevTools for Agents | Console logs, network traffic, accessibility trees for 20+ coding tools | Live today |
| HTML-in-Canvas API | Real DOM elements in WebGL/WebGPU canvas | Origin trial |
| Prompt API | Stable multimodal inputs, structured output, reliable JSON | Chrome 148 stable |
| Gemma 197M | Ultra-efficient in-browser model shared across tabs on low-end devices | Live |
| Soft Navigations API | Core Web Vitals for SPAs (Next.js, React, Nuxt) | Live |
| Skills in Chrome | Save and reuse multi-tab AI workflows as one-click macros | Coming |
GEMINI FOR SCIENCE: PEER-REVIEWED, NOT JUST DEMOED
Three experimental tools live in Google Labs: Hypothesis Generation (Co-Scientist), Computational Discovery (AlphaEvolve + ERA), and Literature Insights (NotebookLM).
Named enterprise production partnerships: BASF (AlphaEvolve for supply chain), Klarna (AlphaEvolve enhancing ML models), Daiichi Sankyo and Bayer Crop Science (Co-Scientist), and U.S. DOE National Labs under the Genesis Mission.