Source: AI Cambrian

It has been a year since the last I/O.

The CEO of Google set the tone right at the beginning: In the past year, the AI industry has entered a new stage, where people are no longer just concerned about the technology itself but want to see AI truly bring value to everyday products. Google's answer is today's press conference.

Token Count: From 480 Trillion to 30 Trillion

Token is an intuitive metric to measure the scale of AI adoption.

Two years ago, the total number of tokens processed monthly by various Google products was 9.7 trillion. By last year's I/O, this number had grown to about 480 trillion. This year, it has jumped directly to over 320 trillion per month, an increase of seven times.

The data from developers and enterprises is equally impressive:

Over 8.5 million developers use Google models to build applications each month.

The model API currently processes about 19 billion tokens per minute.

In the past 12 months, over 375 Google Cloud customers have each processed over 1 trillion tokens.

Product Scale: 13 Products Surpass 1 Billion Users

Google currently has 13 products with more than 1 billion monthly active users, five of which exceed 3 billion.

Search remains the AI product with the broadest coverage. AI Overviews has over 2.5 billion monthly active users. AI Mode, as the largest upgrade in Search's history, surpassed 1 billion monthly active users in just one year, with user engagement evolving from single queries to ongoing conversations.

The Gemini app had 400 million monthly active users last year at I/O, and this year it has exceeded 900 million, more than doubling, with daily request volume increasing by over seven times. The Nano Banana image generation model has generated over 50 billion images.

Conversational AI Enters More Products

Ask YouTube

There is a vast amount of content on YouTube, but finding truly relevant material can be challenging. Ask YouTube redesigns this experience, not just displaying matching videos, but directly jumping to the most relevant segments in the video. It is currently beginning testing and will be fully launched in the U.S. this summer.

Docs Live

Google Docs has introduced a voice feature called Docs Live. Previously, using Gemini to write documents required explicit commands; now users can simply speak their thoughts into a microphone, and Gemini will automatically organize them into a document. In the future, it will also support creating and editing documents directly with voice. Docs Live will roll out to subscribers this summer, along with voice features in Gmail and Keep.

Ask Maps

Maps has undergone its largest upgrade in a decade, including the Ask Maps feature, which supports more complex and longer questions.

Infrastructure: Capital Expenditure from $31 Billion to $180 Billion

Sustaining the operations of these product scales requires significant infrastructure investment.

In 2022, Google’s annual capital expenditure was $31 billion. This year it is expected to reach approximately $180 to $190 billion, about six times that amount.

At the chip level, Google unveiled the eighth generation TPU at Cloud Next, for the first time adopting a dual chip strategy, with separate architectures for training and inference:

TPU 8t is used for large-scale pre-training, with raw computation power approximately three times that of the previous generation. Combined with JAX and Pathways, training is no longer constrained to a single data center and can be distributed across multiple sites, globally accessing over 1 million TPUs to form the largest training cluster in the world. Model training time has been reduced from months to weeks.

TPU 8i is designed specifically for inference, fully optimized for speed. Both chips offer about double the energy efficiency of the previous generation.

New Models: Gemini Omni and Gemini 3.5 Flash

Gemini Omni

AI is shifting from predicting text to simulating reality. Gemini Omni is Google's newly released multimodal world model that can accept inputs of any modality and generate outputs of any modality. The initial version focuses on video output and will later support images and text. Users can start experiencing Gemini Omni Flash today on the Gemini app, Google Flow, and YouTube Shorts, and developer and enterprise clients will gain access via API in the coming weeks.

1) Edit videos with natural language

Omni supports step-by-step video editing through conversation, where each command builds on the previous one, maintaining consistent character and realistic physics, and coherence in scenes.

2) Understanding of physics and world knowledge

Omni has a more accurate intuitive understanding of physical laws like gravity, kinetic energy, and fluid dynamics, resulting in more realistic physical representations in generated scenes. At the same time, it can invoke Gemini's historical, scientific, and cultural background knowledge, linking language, images, and meaning beyond mere pattern matching.

3) Any combination of inputs

Omni allows images, text, video, and audio to be used as reference inputs simultaneously, generating outputs in a unified style.

4) Digital Avatar

Users can create their own digital Avatar using Omni, generating videos that look and sound like themselves. Google has indicated it is still testing modifications to audio and speech in videos.

However, some users have found that the video generation of Omni is quite lacking, falling short compared to seedance 2.0.

Gemini 3.5 Flash

Google today launched Gemini 3.5 Flash, positioned as a next-generation model that combines cutting-edge intelligence with action capabilities.

Compared to 3.1 Pro, 3.5 Flash leads in several benchmark tests, specifically including: Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo), and MCP Atlas (83.6%), reaching 84.2% in multimodal understanding at CharXiv Reasoning. GDPVal specifically measures economically valuable tasks in the real world, showing particularly notable progress. In terms of speed, output tokens per second are four times that of other leading models.

3.5 Flash, in conjunction with Antigravity, can coordinate multiple sub-agents to work together on large-scale complex tasks.

The front-end generation capabilities remain very strong, and on a multimodal basis, 3.5 Flash can also generate more rich interactive web UIs and graphics, such as creating interactive animations for a research paper directly in AI Studio.

In terms of cost, 3.5 Flash is less than half the price of equivalent leading models. Google provided an estimate: leading companies process about 1 trillion tokens daily, and if 80% of the workload is switched from other leading models to 3.5 Flash, over $1 billion can be saved annually.

Gemini 3.5 Flash is now open to all users and APIs. Gemini 3.5 Pro is currently in internal use at Google and will be released next month.

Internal data from Google shows that with the help of the Antigravity platform and 3.5 Flash, the token processing volume for AI development tools has doubled from 500 billion daily in March to now over 3 trillion daily.

Antigravity 2.0: Agent Development Platform

Antigravity was originally an AI programming environment and is now expanding into a complete autonomous AI Agent development and management platform.

Antigravity 2.0 is a new standalone desktop application, serving as the center for Agent interactions, allowing users to coordinate various agents to handle different tasks. This version is optimized for speed and is 12 times faster than other leading models. Antigravity 2.0 and Codex look very similar, 😂

Antigravity users can start experiencing it today. For more details, see the official introduction:

https://deepmind.google/technologies/antigravity/

Gemini Spark: 24/7 Personal AI Agent

The Gemini app is about to launch a personal AI agent—Gemini Spark—that takes action in the digital world on behalf of users with their authorization.

Several key features:

• • Operates on a dedicated virtual machine in Google Cloud, running continuously 24/7 without needing the computer to stay on

• • Driven by Gemini 3.5 and Antigravity, it can easily handle long-term tasks in the background

• • Integrates starting with Google's own tools, with third-party tools being added through MCP in the coming weeks

• • Supports interaction within the Gemini app, and in the future, can also be used through email and instant messaging

• • On Android, users can view real-time progress of the Agent through the new UI space Android Halo, launching later this year

• • Later this summer, Spark will run directly in Chrome, becoming a cross-web Agent browser

Starting this week, Spark will be open to trusted testing users, with the Beta version launching next week for U.S. Google AI Ultra subscribers.

Search Enters the Agent Era

Search is also evolving towards the Agent direction.

Information Agent: Users can set up personalized AI Agents to run continuously in the background, proactively finding essential information at the right time and assisting in actions. This will be launched for Google AI Pro and Ultra subscribers this summer.

Generative UI: Combining Gemini 3.5 Flash and Antigravity, Search will dynamically generate customized interfaces for each question, including personalized layouts and interactive visual content. This will be free for all users this summer.

Persistent Custom Dashboard: For long-term tasks that require continuous tracking, Search can build exclusive dashboards or tracking tools for users, similar to mini-applications customized for specific tasks. This will be available for U.S. Google AI Pro and Ultra subscribers starting in the next few months.

Other Release Contents

Daily Brief: The Gemini app will soon launch a plug-and-play Agent integrating inbox, calendar, and tasks to generate personalized daily summaries, not just aggregating information, but also prioritizing, organizing, and suggesting next steps, presented in a way that facilitates quick browsing.

Google Flow: A new Agent was launched for all users today, capable of planning and handling complex tasks under user participation and control. It supports directly coding various creative tools within Flow, such as video effect design, hand-drawn animation, or text overlay tools.

Google Pics: An AI image creation and editing tool based on the latest Nano Banana model. It treats each element in the image as an independent object rather than a flat image, allowing precise creation, replacement, and adjustment of specific details. It is currently open to trusted testing users and will be launched for Google AI Pro and Ultra subscribers in Workspace later this summer.

Smart Glasses: The AI glasses product unveiled last year has more details, divided into audio glasses (headphone style, voice prompts) and display glasses (capable of showing information), both supporting hands-free use of Gemini. Audio glasses will hit the market first this fall.

Gemini for Science: An AI toolkit for scientific research that integrates Gemini's deep reasoning and research capabilities, Deep Think, and Deep Research, and adds Science Skills that connect platforms like Antigravity to over 30 major life science databases and tools. Users can apply to trial Gemini for Science experimental features in Google Labs, and Science Skills are now directly available in Github and Antigravity.

From TPU 8i to Gemini 3.5, to Antigravity and Spark, what Google presented at this I/O is a comprehensive system evolving towards Agents from chips to applications.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。