The infrastructure commitment: $175-185 billion in capital expenditure for 2026. This isn't theoretical. This is Google betting that the constraint isn't capability anymore. It's control.
ichai opened Cloud Next with numbers. Kurian followed with architecture. Between them, they told a story about why the cloud market is about to reorganize itself around a single operating problem that nobody had solved until now.
That problem isn't building AI agents. It's running hundreds of them in production without losing control of what they're doing.
The Pichai Thesis: Demand Outpaced by Supply
Pichai led with business metrics because the business is what matters. Google Cloud crossed $70 billion in annual revenue and is growing at 48 percent year-over-year. Azure grows at 39 percent. AWS grows at 20 percent. Market share still favors AWS at 32 percent, but Google's growth rate tells you where momentum is moving.
The backlog is the real story. Two hundred and forty billion dollars in contracted but undelivered work. That's three years of current revenue already promised. Customers aren't signing up for experimental AI features. They're signing up for something they need to move production workloads onto, and Google can't build infrastructure fast enough to deliver it.
Pichai quantified Google's own proof point. Seventy-five percent of new code at Google is now AI-generated and approved by engineers. That wasn't a feature announcement. That was Pichai saying: we're not just selling you this. We're running it at massive scale inside Google. A complex code migration that took engineers a year to complete twelve months ago now takes agents and engineers together six weeks. That speed advantage compounds.
But here's the constraint Pichai didn't say explicitly: none of that matters if you can't prove to your auditor that your agents aren't doing something dangerous without supervision. A 48 percent growth rate means nothing if your next quarter brings a compliance incident because an agent modified access permissions in a way your security team didn't see coming.
That single sentence is the pivot point. Pichai wasn't celebrating capability. He was acknowledging the operational crisis that everyone in the room was already facing.
The Kurian Architecture: Making Fleet Management Real
Kurian answered the question Pichai posed. He walked through not a product roadmap, but an operating system architecture for running agent fleets. The specificity matters, because this is where Google moves beyond "we built something cool" to "here's how you control it."
Agent Identity. Every agent gets a unique cryptographic ID and authorization policy. The agent cannot act outside that policy. You can trace every action back to a specific agent, in a specific workflow, with specific permissions. That's table stakes for regulated environments. That's the baseline.
Agent Gateway. Kurian called it "air traffic control for your agent ecosystem." Every agent-to-agent interaction, every agent-to-data handoff, every time an agent calls an external system, it goes through the gateway. The gateway enforces real-time policy. If an agent tries to do something outside its authorization scope, the gateway stops it before the action executes. Not after. Before.
Agent Observability. You can see every decision, every step in an agent's reasoning, every tool it called, every data it accessed, and every output it generated. Not in a log file you review after an incident. In a dashboard you can interrogate in real time. The observability is standardized on open telemetry, so you can integrate it with your existing security infrastructure.
Agent Anomaly Detection. The system learns what normal agent behavior looks like. If an agent starts acting abnormally—accessing data it typically doesn't touch, calling tools it doesn't usually invoke—the system flags it automatically. Before the anomaly becomes an incident.
Long-running agents. Most agent discussions assume agents run for minutes or hours. Kurian introduced agents that can work autonomously for days, orchestrating complex multi-step business processes—reconciliation activities, sales prospecting workflows, supply chain orchestration. The catch: they do this in secure sandboxes, with governance applied at every step.
Kurian then showed how this architecture works in the real world. GE Appliances deployed 800 agents across manufacturing, logistics, and supply chain operations. Not 800 experiments. Not 800 pilots. Eight hundred agents running production workloads. Macquarie Bank reclaimed 100,000 hours of employee time by automating workflows with agents. NASA used agents to power flight readiness checks for Artemis II. These aren't proof of concepts. They're operational at scale.
The Competitive Constraint: Cloud Agnosticism
Here's where the business model gets interesting. Eighty percent of enterprises now run workloads on two or more public clouds. Not because they love complexity. Because single-vendor dependency has become too risky. AWS for one set of workloads. Azure for another. Google for AI and data analytics. No single vendor owns the entire stack.
This creates a fundamental constraint on Google's otherwise commanding position. Customers want Kurian's agent platform. They want the governance layer, the observability, the control. But they don't want to be locked in by it. They want to be able to run agents on AWS or Azure infrastructure while using Google's Agent Platform governance. Or they want to migrate agents between clouds without rewriting their governance rules.
Kurian's answer is integration. The Agent Platform integrates with Model Context Protocol, the open standard for agent-to-agent communication. ServiceNow can integrate with the Agent Platform using MCP. Partner agents can plug in. The architecture is intentionally open because the market demands it.
But here's the asymmetry: Google owns the data layer (BigQuery, the Agentic Data Cloud), the governance layer (Agent Gateway, Agent Identity), and the observability layer (Agent Observability). AWS doesn't offer that integrated control plane. Azure doesn't either. You can run agents on AWS or Azure, but you have to stitch together governance from multiple vendors. Google gives you unified control.
That's the competitive position. Not "we lock you in." But "we solve the governance problem so thoroughly that you trust us enough to build your agent infrastructure on top of our platform, even if you run workloads across multiple clouds."
What This Means for Your Infrastructure Investment
If you're a CIO evaluating cloud platforms in 2026, the traditional vendor selection criteria don't apply anymore. The question isn't "Which cloud is cheapest?" or "Which cloud has the most services?" The question is: "Which cloud vendor gives me enough visibility and control over agent operations that I can run hundreds of agents in production without creating uncontrollable risk?"
Google Cloud, based on what Kurian announced, has an answer: Agent Identity for authorization, Agent Gateway for policy enforcement, Agent Observability for visibility, and Agent Anomaly Detection for risk flagging. It's a coherent, integrated architecture built specifically for the problem you're actually solving.
AWS will compete on scale, breadth, and integration with existing infrastructure. Azure will compete on Microsoft 365 integration and hybrid cloud capabilities. But if your next five years look like "we're going to deploy hundreds of agents across our business," the vendor that gives you architectural control matters more than the vendor with the most services.
Pichai's backlog proves demand is real. Kurian's architecture proves Google is solving the actual problem. Together, they're repositioning cloud from a commodity compute service to a specialized operating system for autonomous business operations.
The vendors who compete on that basis will win. The vendors who try to add agent governance to existing cloud services as an afterthought will lose.
CIO/CTO Viability Question
In the next eighteen months, if your organization deploys 200 or more AI agents, can your cloud provider show you—in real time, before an agent takes action—whether that agent is operating within its authorization scope? Can you see every decision the agent is making? Can you automatically flag when an agent starts behaving abnormally? And if you need to migrate agents to a different cloud provider, can you move the governance rules without rewriting them? If the answer to any of these is "we'll have to integrate multiple vendors," you're not evaluating a cloud platform built for the agentic era. You're evaluating yesterday's infrastructure with governance added on top.
