Thomas Kurian walked onto the Mandalay Bay stage in Las Vegas in April 2026 and opened with a number that reframed everything that followed.
$185 billion.
That is Google’s capital expenditure plan for 2026. Not a five-year commitment. One year. A nearly sixfold increase from the $31 billion they spent in 2022. And more than half of that machine learning compute, Kurian told the audience at Google Cloud Next, would go directly to Google Cloud customers.
In a room full of enterprise CIOs and CTOs, numbers like these produce a particular kind of attention. Not the attention of watching a competitor announce a product. The attention of understanding that the infrastructure beneath an industry is shifting in ways that will not be reversible.
What Google described at Cloud Next 2026 was not a product roadmap. It was the architecture for something they are calling the agentic enterprise: a model of business operations where autonomous AI agents execute tasks, make decisions, and orchestrate workflows across entire organizations without requiring constant human instruction. The question Sundar Pichai framed from the stage captured the moment precisely. “The conversation has gone from ‘can we build an agent’ to ‘how do we manage thousands of them.'”
Thousands. Not pilot programs. Not experimental deployments in a single department. Thousands of autonomous agents operating in production across the full scope of a large enterprise’s operations.
The Physics Problem That Explains the $185 Billion
Before the agents, the platforms, and the customer case studies, there is a constraint that most enterprise technology discussions never reach: the fundamental physics of training and serving AI at scale.
Amin Vahdat, Google’s SVP and Chief Technologist for AI and Infrastructure, addressed it directly. “It is a constant struggle against the limits of software architecture, hardware architecture, power, cooling, and even the speed of light.”
Speed of light is not a figure of speech. When training frontier AI models across tens of thousands of chips in a single cluster, the time required for signals to travel between processors becomes a real performance constraint. The topology of how chips connect matters. Power delivery to the cluster matters. Heat dissipation across the facility matters. These are engineering problems, not software problems, and they cannot be solved by renting capacity from a standard cloud provider.
This is why Google built what they call the AI Hypercomputer: not just better chips but an entire data center architecture designed from first principles around the specific physics of AI training and inference at massive scale.
The eighth generation of Google’s TPUs, announced at the event, marked something new. For the first time, the company split into two specialized platforms rather than one general-purpose accelerator. The TPU v8-T, optimized for training, delivers 121 exaflops of FP4 compute per pod and connects 9,600 processors in a 3D torus topology with twice the interconnect bandwidth of the previous generation. The TPU v8-I, optimized for inference and reinforcement learning, broke the memory wall by hosting the entire memory cache on silicon, deploying 1,152 chips into a single pod using a dragonfly topology specifically designed for running millions of concurrent agents with near-zero latency. Performance increased 9.8x over the previous generation for inference workloads.
The divergence between training-optimized and inference-optimized hardware reflects the divergence in what those workloads actually require. Training wants raw compute power and massive memory bandwidth. Inference wants low latency and the ability to handle concurrent requests at a scale that training hardware was never designed for. Trying to optimize a single platform for both was, in retrospect, a constraint.
Google also announced Virgo, a new networking architecture that links 134,000 chips with up to 47 petabits per second of non-blocking bandwidth, supporting up to 960,000 GPUs and delivering 1.7 million exaflops of compute. Training runs that previously took months now take weeks. The company is also among the first to offer Nvidia’s Vera Rubin NVL72 GPUs, achieving what they describe as 10x performance efficiency for high-interactivity and long-context workloads.
Why does any of this matter to an enterprise executive who is not going to build their own TPU cluster?
Because the availability of this infrastructure at cloud scale changes what is economically feasible to build and run. Citadel Securities, which trades more than half a trillion dollars per day, partnered with Google Cloud to build what their team describes as the preeminent scalable cloud-based research environment. Results: workloads that previously took weeks now execute in hours or minutes. The operation runs 2x to 4x faster at 30 percent lower cost compared to their previous infrastructure.
The constraint on Citadel’s quantitative researchers is no longer the compute platform’s capacity or economics. It is their own imagination. That shift, from infrastructure-limited to imagination-limited, is what $185 billion in annual capital expenditure is actually buying.
Gemini and the Agent Architecture
The hardware story explains the possible. The software architecture explains the actual.
Google’s Gemini model family, which now ranges from Gemini 2.0 Flash for high-volume lightweight tasks to Gemini 2.5 Ultra for the most demanding reasoning workloads, functions as the reasoning layer for the company’s agentic architecture. But the more consequential announcement for enterprise deployments was around the infrastructure surrounding models: how agents access data, how they interact with external systems, and how organizations govern what agents can and cannot do.
Agentspace is Google’s response to the practical challenge of deploying agents in enterprises with messy, distributed, and heterogeneous data environments. Most organizations have data spread across dozens of systems. Email, documents, CRM records, ERP data, proprietary databases, third-party services. Getting an AI agent to reason about business context requires giving it coherent access to all of that without requiring the organization to first consolidate it into a clean unified data store, which no large enterprise has ever successfully done.
Google’s approach connects agents to data in place rather than requiring migration. The Unified Data Plane handles routing access requests to the right data source based on what the agent needs, in real time, across heterogeneous systems. This is architecturally significant because it decouples agent deployment from the data transformation projects that would otherwise be prerequisite.
The Agent Development Kit, announced at Cloud Next, provides structured tooling for building agents that can call external APIs, interact with other agents in a multi-agent system, and be composed into complex workflows. The Agentive API ecosystem maps the external service landscape, enabling agents to take action in the world beyond the organization’s own systems.
For enterprise deployments, the model health dashboard and agent reliability infrastructure represent the governance layer that makes running thousands of agents operationally feasible. Organizations deploying a single experimental agent can manage it manually. Organizations running thousands of agents in production need systematic visibility into what agents are doing, what they are getting right, what they are getting wrong, and where human review needs to happen.
What Actual Deployment Looks Like
The most persuasive material at Cloud Next was not the architecture slides. It was the case studies of organizations that have moved past experimentation into genuine production deployment.
Walmart deployed AI agents to 4,700 stores. The scale is worth sitting with for a moment. 4,700 locations, each with its own operational context, staffing situation, inventory state, and customer base, now have AI agents operating as part of their regular workflows. Home Depot’s Magic Apron assistant drives measurably higher conversion rates, which in a business operating at Walmart and Home Depot’s scale means the agent deployment is contributing meaningfully to top-line revenue.
Axia Energia, Brazil’s largest power company, runs advanced weather modeling on Google’s TPU clusters to forecast severe weather conditions up to ten days in advance. The operational application is direct: proactive planning and preventative measures before storms hit, reducing power outages for millions of customers. The value is not measured in AI performance metrics but in avoided outages and the customer experience and economic activity they represent.
Woven by Toyota has achieved 42 percent faster training for models predicting complex traffic events. Boston Dynamics trains and deploys vision language models for robotics across industrial applications. The U.S. Department of Energy powers AI research across all 17 national laboratories.
These are not technology companies running AI experiments. They are large, complex organizations in traditional industries using AI infrastructure to change how they operate at a level that compounds over time.
The common thread across the case studies is operational integration. The agents are not sitting alongside existing workflows generating reports that humans then act on. They are embedded in the workflows themselves, taking actions, making decisions, triggering downstream processes. The human oversight remains, but it operates at a higher level of abstraction, reviewing outputs and exceptions rather than executing every step.
The Governance Question That Cannot Be Avoided
Pichai’s framing, from can we build an agent to how do we manage thousands of them, captures the shift in the hard problem.
Building agents is now, with Google’s tooling and infrastructure, a tractable engineering challenge for organizations with competent technical teams. The much harder challenge is governing what those agents do when they are operating across an organization at scale.
An agent that processes invoices and occasionally makes a wrong classification creates a different risk profile than an agent that can commit financial resources, send communications on behalf of the organization, or modify customer records. The governance architecture has to distinguish between these risk profiles and apply human oversight appropriately: consistently and efficiently enough that it does not negate the efficiency gains from automation, but reliably enough that high-stakes errors get caught before they cascade.
Google’s model health dashboard and the broader reliability infrastructure Google announced addresses this at the infrastructure level. But infrastructure governance is the floor, not the ceiling. Organizations deploying agents at scale will need their own governance policies, review processes, and escalation structures tailored to their specific risk environments.
The enterprise deployment patterns that emerge over the next two to three years will do more to define how organizations actually govern agentic AI than any framework published today. The organizations building those patterns now, in production, across thousands of deployed agents, are accumulating operational knowledge that will be difficult to replicate quickly. Walmart’s institutional knowledge about how to deploy agents across 4,700 locations is not in Google’s documentation. It is embedded in the decisions Walmart made, the problems they encountered, and the solutions they developed through iteration.
This is what the gap between early adoption and late adoption in agentic AI actually looks like. Not a technology gap, which can be closed by buying access to the same infrastructure. An operational learning gap, which takes time and production experience to develop regardless of when an organization starts writing checks to Google Cloud.
The Strategic Question for Every Executive in That Room
The honest version of what Google Cloud Next communicated to the enterprise executives in the audience was something like this: the organizations that build operational competency with agentic AI over the next two to three years will not merely have a competitive advantage. They will have a different capacity for executing strategy than organizations that wait.
This is a stronger claim than most technology announcements justify. Most technology upgrades make existing operations somewhat faster or cheaper. The compound effect of deploying AI agents across the full range of an enterprise’s operations, over years of iteration, is something more fundamental than incremental improvement. It is a change in the rate at which the organization can translate decisions into executed action.
Citadel’s researchers going from weeks to hours is not a 10 percent productivity gain. It is a change in the temporal rhythm of how the firm does research. The decisions available to a team that can test 100 ideas in the time it previously took to test one are qualitatively different from the decisions available to a team that cannot.
Most enterprises will not achieve anything close to Citadel’s deployment sophistication quickly. They will start with more bounded applications, build operational confidence, and expand over time. But the compounding starts with the first deployment, not with full-scale rollout. And it stops accruing to organizations that have not started yet.
That was the message behind the $185 billion. Not that Google is spending an extraordinary amount of money on infrastructure. It is that the infrastructure exists, the tools to build on it exist, and the question has moved from whether agentic AI is real to whether your organization has the foundation to participate in what is already being built on top of it.
Key Takeaways
- Google plans $185 billion in capital expenditure in 2026, nearly six times its 2022 spending, with more than half of the machine learning compute going to Google Cloud customers.
- New eighth-generation TPUs split into two specialized platforms: TPU v8-T for training (121 exaflops per pod, 9,600 processors in 3D torus topology) and TPU v8-I for inference (1,152 processors optimized for millions of concurrent agents with near-zero latency, 9.8x performance improvement).
- The Virgo network architecture links 134,000 chips at 47 petabits per second of bandwidth, supporting 960,000 GPUs and 1.7 million exaflops of compute.
- Citadel Securities runs financial modeling 2x to 4x faster at 30 percent lower cost; Walmart has deployed agents to 4,700 stores; Home Depot’s Magic Apron drives measurably higher conversion rates.
- Agentspace, the Agent Development Kit, and the Unified Data Plane address the practical challenge of deploying agents in enterprises with distributed, heterogeneous data environments.
- Sundar Pichai framed the shift precisely: the question has moved from “can we build an agent” to “how do we manage thousands of them.”
Important Insights
The split between training-optimized and inference-optimized hardware reflects a maturation in how the industry understands AI workloads. The performance characteristics required to train large models efficiently are fundamentally different from those required to serve millions of concurrent agent sessions with low latency. Acknowledging that difference and building specialized hardware for each represents an architectural commitment that has compounding returns as deployment scales.
The enterprise case studies at Cloud Next carried more strategic information than the product announcements. Walmart at 4,700 stores is not an experiment. Home Depot driving conversion with an AI assistant is not a pilot. These are production deployments at enterprise scale, and the organizational learning embedded in them is not transferable through documentation. It develops through iteration in production, which means the time spent building that learning matters in ways that the technology cost does not.
The governance challenge is the most underaddressed element of the agentic enterprise transition. Infrastructure is available. Models are capable. The gap between where most enterprises are and where the leaders are is not primarily technical. It is operational: the policies, review processes, escalation structures, and institutional knowledge about what agents can be trusted to do and where human judgment needs to remain in the loop. Building that operational knowledge is the slow, irreplaceable work that Google’s platform enables but cannot do for you.
Google Cloud Next is Google’s annual cloud enterprise conference. More information about Google Cloud’s AI infrastructure and agent development tools is available at cloud.google.com.
