Sundar Pichai sold the agentic future and then named the bill. In Google's edited I/O 2026 keynote transcript, he said the company expects this year's capital expenditure to be about six times its 2022 level of $31 billion, or approximately $180 billion to $190 billion. [1] That number is the business story behind every pleasant demo.

Wednesday's paper treated Google's agent instructions as platform files, because AGENTS.md and SKILL.md-style definitions made the agent more than a chatbot. Thursday's follow-up puts those files under a balance sheet. A managed agent is not cheap magic. It is a claim on data centers, custom silicon, cloud sandboxes, developer tooling and inference capacity.

Pichai framed the spending as full-stack advantage. Google, he said, has custom silicon, a secure foundation, research, models, products and platforms that touch billions of people. [1] The agent story is therefore not a startup pitch grafted onto a search company. It is Google's preferred explanation for why scale still matters.

The usage numbers do the sales work. Pichai said Google processed 9.7 trillion tokens a month two years ago, roughly 480 trillion last year and more than 3.2 quadrillion per month now. [1] He said more than 8.5 million developers build monthly with Google's models, that model APIs process roughly 19 billion tokens per minute, and that more than 375 Google Cloud customers each processed over one trillion tokens in the past 12 months. [1]

That is a wonderful story if you sell compute. It is a harder story if you must finance it. The same keynote that celebrates adoption says supporting users, enterprises and developers requires massive infrastructure investment. [1] In other words, the agentic era is not merely a software cycle. It is a capital cycle with a product interface.

Managed Agents in the Gemini API show why the money must be spent before the revenue is obvious. Google's developer post says a single call can spin up an agent that reasons, uses tools and executes code in an isolated, ephemeral Linux environment. [2] It can call tools, manage files, browse the web and preserve state for follow-up interactions. [2] That is convenient for developers precisely because Google absorbs the operational complexity.

The file detail is the strategic detail. Google says developers can define custom agents in markdown files like AGENTS.md and SKILL.md and register them as managed agents. [2] A standard begins as a habit. If enough developers learn that agents live in these files and run in Google's sandboxes, Google gains leverage that does not appear in a keynote slide.

Pichai's I/O transcript makes the consumer version of the same argument. Gemini Spark, he said, runs on dedicated virtual machines on Google Cloud, works 24/7, uses the Google Antigravity harness and will integrate with tools, including third-party tools through MCP. [1] Search will also get information agents that work in the background, with custom dashboards and trackers coming later. [1] The agent is becoming an ambient cloud worker.

X will mock the branding and compare screenshots. That is useful theater, but it underprices the capex claim. The real fight is not whether a demo looks slick. It is whether Google can make agents routine enough to justify a $180 billion to $190 billion infrastructure year and then route enough tasks through its stack to keep that infrastructure busy. [1]

The mainstream event frame usually treats I/O as a product cascade. That misses the operating model. Google is turning agents into demand generators for TPUs, cloud sandboxes and model APIs. It is also turning infrastructure into an argument that smaller rivals cannot match at the same reliability, latency or cost. The company is saying, in effect: trust our agent because we own the factory.

There is risk in that argument. A capital plan at this scale narrows patience. If tokens grow but margins do not, the agentic era becomes an expensive habit. If developers use managed agents lightly, the sandboxes sit as a cost center. If consumers treat Gemini Spark as a novelty, the 24/7 worker metaphor becomes a data-center invoice with a mascot.

Still, Google's position is coherent. It has products with billions of users, developers already spending tokens, enterprise customers with trillion-token workloads and a distribution surface that can make agents feel normal. [1] The capex bill does not undermine the agent story. It reveals its wager. Google is not selling agents because they are cheap to build. It is selling them because only companies willing to spend like infrastructure sovereigns may be able to make them ordinary.

-- THEO KAPLAN, San Francisco