Sundar Pichai put a number on the table at Google Cloud Next 2026 that stopped the room: $185 billion in capital expenditure for the year. Not over a multi-year plan. For 2026 alone.
Four years earlier, Google’s annual infrastructure investment was approximately $31 billion. The six-fold increase in that span is one of the largest capital allocation decisions in the history of enterprise technology. It is also, if you understand what the money is buying, a precise statement about what Google believes the next decade of computing will run on.
The answer is not just AI. Every major technology company is spending heavily on AI. The specific answer is agentic AI infrastructure: the hardware, networking, software, and governance layers that allow organizations to deploy not one AI assistant but thousands of coordinated AI agents, running continuously, orchestrating decisions across business processes that were never previously automable at anything close to this scale.
Pichai captured the moment with a sentence that deserves to be read carefully: “The conversation has gone from can we build an agent to how do we manage thousands of them.”
That sentence is a roadmap for where enterprise computing is going. It is also a precise description of the infrastructure problem that Google is spending $185 billion to solve.
A Physics Problem in Disguise
Amin Vahdat, Google’s VP of Systems Infrastructure, introduced the challenge as a physics problem, and the framing is more illuminating than it might initially seem.
The speed of light is a hard constraint. When AI inference is happening in a data center thousands of miles from the user or process it is serving, the round-trip latency imposed by the physical distance between compute and application is not a software problem. It is a physics problem. You cannot software-engineer your way past the fact that light takes finite time to travel finite distances.
This constraint becomes critical in the context of agentic AI because agents are interactive systems. An agent that is waiting for instructions, processing them, and returning results to another agent or to a human is engaged in a latency-sensitive dialogue. High latency in that dialogue does not just feel slow. It fundamentally limits what agents can do and how many can be coordinated in real time.
Google’s response to this physics problem is the AI Hypercomputer: a hardware and networking architecture designed to minimize the distance and time between compute and application, and to maximize the density of AI processing that can happen in a given physical and energy footprint.
The two newest TPU systems are the clearest expression of this architecture. TPU v8-T, the training variant, packs 9,600 TPU chips into a single pod connected in a three-dimensional torus topology, delivering 121 exaflops of FP4 computing throughput per pod. The interconnect bandwidth is twice that of the previous generation, which matters because the training of large models is as much a networking problem as a compute problem: chips that cannot communicate results to each other fast enough cannot train at maximum efficiency regardless of how powerful they individually are.
TPU v8-I, the inference variant, is optimized for different tradeoffs. At 1,152 TPUs per cluster in a dragonfly topology, it delivers 9.8x improvement in inference throughput relative to the previous generation. The dragonfly topology reduces the average number of network hops between any two chips in the cluster, which reduces latency for inference workloads that are fundamentally latency-sensitive in a way that training workloads, which are primarily throughput-sensitive, are not.
Alongside its own TPUs, Google also announced availability of Nvidia’s Vera Rubin NVL72 GPUs through Google Cloud, which gives customers who have built expertise on Nvidia hardware the option to access that hardware within Google’s infrastructure management and governance stack. This is not a strategic concession. It is a recognition that the enterprise market is not going to discard Nvidia expertise overnight, and that a platform strategy serves Google better than forcing customers to choose between GPU expertise and Google Cloud.
The Virgo network, the interconnect fabric that ties all of this compute together, operates at 134,000 chips, 47 petabits per second of aggregate bandwidth, and 1.7 million exaflops of total compute. Managed Lustre, the high-performance storage layer, delivers 10 terabytes per second of throughput for the data-intensive workloads that training and large-scale inference require.
These are not incremental improvements to existing infrastructure. They are step-change advances in the raw capability of what Google can offer as a compute substrate.
The Knowledge Catalog: Context as Infrastructure
The hardware story is easier to tell than the software story, but the software may be more consequential.
Agentic AI systems fail in one of two broad ways. They fail because they lack the compute to run the models they need. Or they fail because they lack the context to make their outputs useful. The first failure mode is what the hardware investments address. The second failure mode is what the Knowledge Catalog addresses, and it is arguably the harder problem.
Enterprise data is fragmented across dozens or hundreds of systems in most large organizations: ERP systems, CRM systems, document management systems, email archives, Slack conversations, spreadsheets that live on individual laptops, databases that were designed for a different era and never properly integrated with the systems that came after them. The data exists. But it exists in forms that AI agents cannot easily access, and with contextual relationships that the agents cannot easily understand without significant human guidance.
The Knowledge Catalog is Google’s proposed solution. It sits atop Google Cloud Storage and uses Gemini to automatically analyze files as they are added to the system: understanding content, recognizing entities, inferring relationships between documents, learning the specific vocabulary and taxonomy of the organization over time. The goal is to create what Google calls a “universal context engine” for the enterprise: a layer that knows what data exists, where it is, how it relates to other data, and what it means in the context of the specific organization’s operations.
Virgin Media O2’s deployment of the Knowledge Catalog produced 20,000 tagged data assets within weeks. This is not a search index, though search is one of the capabilities it enables. It is a semantic layer that allows AI agents to understand enterprise data well enough to use it as context for decisions, rather than simply retrieving documents in response to queries.
The demonstration case Google chose to illustrate the Knowledge Catalog’s capabilities was a scenario called Midnight Swirl: a fictional beauty brand facing a viral social media trend and needing to evaluate the business opportunity. The five-minute demonstration moved from trend identification to a revenue projection of $15 million, with a path through a previously unknown soy allergen that the Knowledge Catalog surfaced by finding relevant information across two separate PDFs, a cross-cloud lakehouse query that joined data from AWS S3 and BigQuery in the same analysis, and a projected revenue model built from real inventory and demand data.
The scenario was designed to show maximum impact in minimum time, which is what demonstrations are for. But the specific challenge it illustrated, finding contextually relevant information across fragmented data sources and surfacing it at the moment when an agent needs it to make a decision, is a real and common enterprise problem. Most organizations have experienced the frustration of making a significant business decision with incomplete information, discovering after the fact that the relevant information existed but was not found in time. The Knowledge Catalog is the infrastructure bet that this failure mode can be systematically addressed rather than merely hoped-away.
Security in the Agentic Era: The Exploit That Arrives Before the Patch
Francis Dsouza, presenting on security at Google Cloud Next, offered a statistic that crystallized the threat landscape for enterprise AI: mean time to exploit critical vulnerabilities has dropped to negative seven days. The negative sign is not a typographical error. Attackers are now exploiting vulnerabilities an average of seven days before the security patches for those vulnerabilities are published.
This is not a new security problem. But it is a security problem that is dramatically worse in the context of agentic AI. An agent that has access to enterprise data and can take actions on behalf of users is a much more valuable target than a static application. Compromising an agent that manages customer communications or financial approvals is qualitatively different from compromising a web form.
The Wiz acquisition, Google’s largest acquisition in history, is the strategic response to this threat. Wiz brought a comprehensive cloud security platform whose AI Application Protection Platform operates through three parallel agent types: red agents that continuously test for vulnerabilities, blue agents that defend against discovered attacks, and green agents that help developers build more secure code from the beginning. The three-agent architecture is an acknowledgment that security in the agentic era cannot be a periodic audit. It must be continuous and automated, running at the speed of the threat environment rather than at the speed of human security review cycles.
The operational impact Dsouza cited is the evidence that the approach works: handoff time for security issues dropped from eight hours to 22 seconds. That is not a marginal improvement. It is the difference between a threat that escalates for hours before anyone with authority to address it is notified, and a threat that is in front of the right person or process within seconds of detection.
Model Armor, Google’s layer for filtering malicious inputs before they reach AI models, and Agent Identity, which assigns cryptographic identifiers to each agent in a multi-agent system, address the attack surface that is specific to agentic AI architectures. When agents can communicate with each other and take actions based on each other’s outputs, a malicious actor who can inject false outputs into that communication chain can cause significant damage without directly compromising the underlying models. Agent Identity creates an audit trail and authentication mechanism that makes that injection attack significantly harder to execute invisibly.
The Enterprise Agent Platform: Mission Control for the Agentic Enterprise
The consolidation of Vertex AI and Agentspace into the Gemini Enterprise Agent Platform is a product announcement that is more strategically significant than it might appear on a product roadmap.
Vertex AI was Google’s machine learning platform for developers and data scientists. Agentspace was the enterprise search and knowledge management product for business users. They served adjacent needs but were sold to different buyers and organized under different product teams. The consolidation reflects a strategic judgment that the enterprise AI market is converging: the developer building an agent and the business buyer deploying that agent to their workforce are increasingly using the same infrastructure, and separating them into different products creates unnecessary friction.
The Gemini Enterprise Agent Platform includes an Agent Registry that catalogs all agents running in an enterprise environment, a Skills and Tools Registry that inventories what each agent can do and what data sources it can access, and an Agent Marketplace that allows sharing and reuse of agent capability across an organization or, eventually, across organizations.
The governance implications of this architecture are significant. An enterprise that has deployed hundreds of agents without a centralized registry has no practical way to audit what those agents are doing, what data they are accessing, or whether their behavior is consistent with policy. The Agent Registry is the prerequisite for answering those questions. The Agent Observability layer, built on OpenTelemetry standards, provides the logging and tracing infrastructure that allows issues to be diagnosed when agents behave unexpectedly.
Gemini 3.1 Pro, in preview at the time of the conference, represents the model capability layer that the platform infrastructure is designed to host. The capability of the model matters, but the argument Google is making at Cloud Next 2026 is that model capability alone is insufficient. An enterprise that has a powerful model and no infrastructure for deploying, governing, and monitoring agents built on that model has a capability that it cannot reliably operationalize at scale.
What the Customer Numbers Actually Say
The customer deployments Google cited at Cloud Next 2026 range from impressive to extraordinary, and the spread across industries and use cases is worth examining for what it reveals about where agentic AI is delivering value.
Signal Iduna, the German insurance group, achieved 80 percent adoption across 11,000 employees within weeks of deployment. The specifics are important: a 400 percent weekly user growth surge, and answers to employee questions arriving 37 percent faster. Insurance is a knowledge-intensive industry where employees spend significant time finding information to answer client questions or make underwriting decisions. The 37 percent improvement in answer speed, multiplied across 11,000 employees, represents a substantial productivity gain.
KPMG reached 90 percent adoption with more than 100 agents deployed in the first month. The adoption rate is particularly notable because KPMG is a professional services firm where the primary product is human judgment and expertise. High AI adoption in that context is not employees substituting AI for their own judgment. It is employees finding that AI assistance makes their existing judgment more efficient and better-informed.
Merck’s deployment to 75,000 employees represents scale of a different kind. A pharmaceutical company has extraordinarily complex regulatory and technical information domains, and making that information accessible to 75,000 employees through AI agents is an information architecture challenge that conventional enterprise search cannot adequately address.
Walmart’s deployment across 4,700 stores, accessing operational and supply chain data through conversational interfaces, addresses a management information challenge that is specific to large-scale distributed retail: the store leaders who make daily operational decisions are often poorly equipped with the data they need to make those decisions well. An agent that can answer supply chain questions in plain language, rather than requiring the store manager to navigate business intelligence dashboards, puts better information in the hands of the people making operational decisions in real time.
Google’s own internal deployment numbers are the most striking: 75 percent of new code at Google is now generated by AI. Code migration processes are running six times faster. Marketing workflows are completing 70 percent faster with a 20 percent improvement in conversion rates. Security threat mitigation has been reduced by 90 percent.
Key Takeaways
The $185 billion infrastructure investment is the clearest signal in enterprise technology that Google believes the agentic enterprise is not a future scenario but an immediate reality that requires infrastructure investment at a scale commensurate with its eventual importance. Companies that are evaluating AI infrastructure choices should understand this investment not as a data center buildout but as the construction of the underlying platform on which the next generation of enterprise computing will run.
The Knowledge Catalog represents a genuine architectural advance in enterprise AI. The failure mode of AI agents that cannot access or understand enterprise data is more common and more consequential than the failure mode of AI agents that lack model capability. Enterprise organizations that invest in the data infrastructure layer before optimizing the model layer are making the sequence right.
Security in the agentic era requires a fundamental rethinking of what security means. The negative seven-day mean time to exploit is not a problem that periodic security audits can address. Continuous, automated security monitoring running at agent speed is becoming a prerequisite for responsible enterprise AI deployment, not a premium feature.
The adoption rates from Signal Iduna, KPMG, and Walmart suggest that enterprise AI adoption is no longer primarily constrained by employee willingness to use AI tools. The organizations that have deployed effectively are seeing 80 to 90 percent adoption within weeks. The constraint has shifted to infrastructure quality: organizations that deploy AI on well-designed infrastructure see rapid adoption; organizations that deploy on poorly-designed infrastructure see the friction that creates low adoption, and then incorrectly conclude that their workforce is resistant to AI.
Notes for Enterprise Technology Leaders, Infrastructure Architects, and AI Strategists
The negative seven-day mean time to exploit statistic deserves to become part of every enterprise security conversation. It changes the fundamental premise of patch-based security management: if attackers are exploiting vulnerabilities before patches exist, the operational response cannot be patch deployment. The response must be continuous detection, containment, and response capabilities that operate faster than the exploit window.
The Knowledge Catalog’s approach of learning organizational semantics over time rather than requiring manual data cataloging is worth studying carefully. Most enterprise data governance programs fail because they require significant ongoing human investment in tagging and metadata management that organizations cannot sustain. A system that learns semantics automatically and improves with use has a fundamentally better adoption trajectory.
For organizations currently running Vertex AI: the consolidation into Gemini Enterprise Agent Platform is a branding change that does not require immediate operational response. Existing deployments continue to function. The planning implication is that future capability additions will be governed by the Gemini Enterprise roadmap, and organizations should evaluate their vendor relationships and procurement frameworks accordingly.
The NASA Artemis 2 flight readiness use case, cited at the conference, is worth noting for what it says about where agentic AI is being deployed in high-stakes decision environments. The Artemis program’s flight readiness process involves hundreds of subsystems and thousands of parameters that must be evaluated before launch. An agent system that assists with that evaluation is not replacing human judgment. It is making human judgment more comprehensive by ensuring that no relevant parameter goes unreviewed.
Google Cloud Next 2026 ran April 22-24 at the Mandalay Bay Convention Center in Las Vegas. For product documentation, research, and the Gemini Enterprise resource library, visit cloud.google.com.
