Google I/O 2026: Tokenmaxxing, Capex, and the Rise of Always-On Agents

What is the Scale of Google's Tokenmaxxing?

Sundar Pichai opened the Google I/O developer conference by framing the company's recent trajectory as a celebration of both data throughput and financial commitment. He introduced the term tokenmaxxing to describe the exponential growth in how Google handles tokens, which are the fundamental units of data exchange for artificial intelligence models.

The numbers presented during the keynote illustrate a staggering acceleration in demand. Two years ago, Google processed 9.7 trillion tokens per month. By last year, that figure jumped to 480 trillion. Currently, the company reports handling 3.2 quadrillion tokens monthly. Pichai acknowledged that critics might view this metric as vanity, but he argued it reflects genuine product utility and developer adoption.

Over 8.5 million developers are now building applications using Google's Gemini model family on a monthly basis. These developers consume approximately 19 billion tokens per minute through API calls. Furthermore, more than 375 enterprise customers have consumed over one trillion tokens each in the past twelve months, signaling robust commercial demand for AI inference capabilities.

Why Does Massive Capex Matter for Infrastructure?

Supporting this volume of token processing requires substantial physical and computational resources. Pichai emphasized that Google has been investing heavily in data centers, compute capacity, and Tensor Processing Unit hardware to meet both user and enterprise needs.

In 2022, Google's annual capital expenditure stood at $31 billion. For the current year, the company expects this figure to reach approximately 180 to 190 billion dollars. This sixfold increase underscores the financial intensity required to maintain leadership in AI infrastructure. The scale of investment is comparable to other major technology firms racing to build compute capacity.

While competitors like SpaceX are also expanding their technological ambitions, Google's focus remains squarely on the internal and external demand for AI inference. The capital expenditure strategy is designed to ensure that the company can serve millions of developers and enterprises simultaneously without bottlenecking performance.

How Does Gemini 3.5 Flash Change Performance?

Demis Hassabis, co-founder and CEO of Google DeepMind, provided an update on progress toward artificial general intelligence. He introduced Gemini Omni as a step in that direction, describing it as a model capable of creating anything from any input. This includes digital content rather than atomic replication.

Gemini Omni combines the intelligence of the Gemini family with generative media models like Veo and Genie. It incorporates physics modeling to accurately depict object interactions involving kinetic energy and gravity. The first model in this family, Gemini Omni Flash, is now available for use.

Pichai returned to the stage to announce the next generation of the core model family: Gemini 3.5 Flash. Compared to Gemini 3.1 Pro, Flash offers better performance across almost all benchmarks. A major selling point is its speed. The model processes approximately 289 tokens per second, which Google claims is four times faster than other frontier models.

For developers using the Antigravity coding harness, the gains are even more significant. DeepMind engineer Varun Mohan stated that Flash is twelve times faster within Antigravity version 2.0. This optimization targets one of the most remunerative use cases for AI: software development and code generation.

Price efficiency is another critical factor. Pichai noted that top companies in Google Cloud process about one trillion tokens daily. If these firms shift eighty percent of their workloads to Gemini 3.5 Flash, they could save over one billion dollars annually. This cost reduction makes the model highly attractive for enterprise adoption.

What Are the Implications of Always-On Agents?

Google is expanding its agentic capabilities through Gemini Spark, an agent service integrated into the Google Gemini app and Search. Pichai described Spark as a personal AI agent that helps users navigate their digital life by taking action on their behalf under direction.

Spark runs on dedicated virtual machines in Google Cloud and operates twenty-four hours a day. Based on Gemini 3.5 Flash and assisted by Antigravity, it can perform long-running tasks in the background without incurring excessive token costs. The agent will initially connect to Google apps like Gmail and Chat, with third-party tool integration via Model Context Protocol planned for later.

Liz Reid, VP of Search, detailed further AI incursions into the Search service. Gemini 3.5 Flash is now the default model for AI Mode. The Search interface has been redesigned to surface AI-based suggestions and facilitate inputs from images, files, videos, and Chrome tabs.

The most significant change is Search Agents. These agents run while users are away from their keyboards, finding information exactly when needed. Users can spin up multiple agents simultaneously to make progress on various tasks. Google is also introducing code-based interactive widgets, allowing users to create dynamic layouts and charts through a containerized environment.

How Does SynthID Protect Content Authenticity?

To address concerns about AI-generated content, Google announced an expansion of SynthID, its AI watermarking technology. The company will support C2PA content credentials verification across its products to help distinguish between AI-created and camera-captured content.

Pichai stated that SynthID and content credentials verification are being expanded to Search and Chrome. Users can circle an image in Search or right-click in Chrome to ask if the content was generated with AI. The response will include helpful context regarding edits made via Google Photos.

Google is collaborating with other industry players to make this technology broadly useful. OpenAI, Kakao, and ElevenLabs have decided to adopt SynthID. This cross-industry adoption aims to establish a standard for content provenance in an era where digital media creation is increasingly automated.

What Is the Future of AI Subscription Models?

The launch of Gemini Spark coincides with changes to Google's subscription tiers. A new Ultra plan tier costs one hundred dollars per month. The top Ultra tier has been deflated from two hundred fifty dollars to two hundred dollars per month.

Pichai acknowledged that making agents easy to use, secure, and helpful remains a challenge in the early days of this technology. However, the rollout of Spark to trusted testers and Google AI Ultra subscribers in the United States marks a significant step toward always-on personal assistance.

Expect Google's token expenditures to continue growing as agentic labor becomes more prevalent. The pressure to purchase subscriptions will likely increase as users rely on these tools for continuous digital navigation and task management. The integration of agents into Search and Chrome suggests that AI is moving from a reactive tool to an proactive partner in daily workflows.

Conclusion

Google I/O 2026 highlighted the company's commitment to scaling infrastructure while introducing more capable and cost-effective models. The shift toward always-on agents and expanded content verification tools reflects a broader industry trend toward integrating AI deeply into daily digital experiences.