Google Cloud Unveils Agent Platform as CapEx Hits $185 Billion

Google Cloud Next '26 Opening Keynote — Google Cloud

Google Cloud

Alphabet argues enterprise returns require abandoning fragmented models and glue code for a vertically integrated stack capable of governing thousands of autonomous agents.

AI-assisted summary of Google Cloud Next '26 Opening Keynote — original source. Check the original for full context.

Source published:

Key ideas

  • Google is escalating its capital expenditures from $31 billion in 2022 to an anticipated $175 billion to $185 billion in 2026, with over 50% of 2026 machine learning capacity allocated directly to Google Cloud customers. (4:32)
  • Google Cloud introduced the Blueprint for the Agentic Enterprise across five foundational pillars: AI Hypercomputer, Agentic Data Cloud, Agentic Defense, Gemini Enterprise Agent Platform, and Agentic Taskforce. (10:20)
  • Eighth-generation TPU architecture has split into two dedicated silicon platforms: TPU 8t for training (1.21 FP4 exaflops per pod) and TPU 8i for inference and long-context autoregressive decoding (11.6 FP8 exaflops per pod). (42:16)
  • To address compressed cyberattack timelines—where lateral handoff has dropped from 8 hours to 22 seconds—Google integrated Wiz with its security stack to deploy autonomous Red, Blue, and Green defense agents. (64:40)
  • Workspace Intelligence natively embeds Gemini models across Google Workspace applications, enabling ambient queries, semantic file retrieval, and automated presentation creation grounded in live CRM and email data. (86:30)

Detailed notes

## Background Published on YouTube by Google Cloud on April 22, 2026, this opening keynote presents the blueprint for the agentic enterprise. Led by Google Cloud CEO Thomas Kurian alongside senior executives and global enterprise partners, the session outlines Google's full-stack strategy to guide businesses through the operational shift from early AI experimentation to large-scale, autonomous enterprise production.

## TL;DR Google Cloud introduces its comprehensive architectural framework for the agentic enterprise, organized across five core layers: the AI Hypercomputer, Agentic Data Cloud, Agentic Defense, Gemini Enterprise Agent Platform, and Agentic Taskforce. Key infrastructure and model milestones include eighth-generation TPUs (TPU 8t and 8i), NVIDIA Vera Rubin NVL72 integration, Gemini 3.1 Pro reasoning models, and native Model Context Protocol (MCP) endpoints across all Google Cloud services. The keynote highlights data innovations such as the Knowledge Catalog and Cross-Cloud Lakehouse, machine-speed autonomous security powered by Mandiant and Wiz, and generative Workspace Intelligence. Throughout the presentation, live demonstrations and enterprise case studies from organizations like Citadel Securities, Virgin Voyages, YouTube TV, and The Home Depot showcase how multi-agent systems are replacing isolated pilots with end-to-end autonomous business workflows.

## Atmosphere & style The presentation carries the polished, high-velocity energy of a premier technology keynote, balancing executive grand strategy with rapid-fire live demonstrations and deeply technical systems engineering. Speakers transition seamlessly from the physical constraints of data center power and silicon networking to practical multi-agent commercial workflows, establishing a confident, assertive tempo. The tone is pragmatic and enterprise-focused, purposefully discarding speculative AI hype to project an atmosphere of mature operational readiness, architectural scale, and relentless competitive momentum.

## Bulletin - [04:32](https://www.youtube.com/watch?v=11PBno-cJ1g&t=272s) Google is escalating its capital expenditures from $31 billion in 2022 to an anticipated $175 billion to $185 billion in 2026, with over 50% of 2026 machine learning capacity allocated directly to Google Cloud customers. - [10:20](https://www.youtube.com/watch?v=11PBno-cJ1g&t=620s) Google Cloud introduced the Blueprint for the Agentic Enterprise across five foundational pillars: AI Hypercomputer, Agentic Data Cloud, Agentic Defense, Gemini Enterprise Agent Platform, and Agentic Taskforce. - [42:16](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2536s) Eighth-generation TPU architecture has split into two dedicated silicon platforms: TPU 8t for training (1.21 FP4 exaflops per pod) and TPU 8i for inference and long-context autoregressive decoding (11.6 FP8 exaflops per pod). - [01:04:40](https://www.youtube.com/watch?v=11PBno-cJ1g&t=3880s) To address compressed cyberattack timelines—where lateral handoff has dropped from 8 hours to 22 seconds—Google integrated Wiz with its security stack to deploy autonomous Red, Blue, and Green defense agents. - [01:26:30](https://www.youtube.com/watch?v=11PBno-cJ1g&t=5190s) Workspace Intelligence natively embeds Gemini models across Google Workspace applications, enabling ambient queries, semantic file retrieval, and automated presentation creation grounded in live CRM and email data.

## Key takeaways - [02:09](https://www.youtube.com/watch?v=11PBno-cJ1g&t=129s) Enterprises have decisively shifted from AI pilots and glue-code experimentation to cross-organizational enterprise production, with nearly 75% of Google Cloud customers running daily AI workloads. - [04:32](https://www.youtube.com/watch?v=11PBno-cJ1g&t=272s) Google acts as 'Customer Zero' for its agentic stack, generating over 75% of new internal code with AI review assistance and accelerating complex code migrations by 6x. - [08:48](https://www.youtube.com/watch?v=11PBno-cJ1g&t=528s) The Gemini Enterprise Agent Platform serves as a centralized mission control center to build, scale, monitor, and govern thousands of autonomous agents with native Model Context Protocol (MCP) support. - [12:32](https://www.youtube.com/watch?v=11PBno-cJ1g&t=752s) Previewed foundation reasoning models include Gemini 3.1 Pro, Gemini 3.1 Flash Image, Veo 3.1 Lite, and Lyria 3 Pro, alongside expanded partnership access to Anthropic's Claude Opus 4.7. - [12:32](https://www.youtube.com/watch?v=11PBno-cJ1g&t=752s) Apple selected Google Cloud as its preferred cloud partner to collaborate on foundation models for Apple Intelligence and Siri. - [22:02](https://www.youtube.com/watch?v=11PBno-cJ1g&t=1322s) Google Distributed Cloud Edge enables high-performance AI deployment in disconnected or offline environments, demonstrated by Virgin Voyages' virtual concierge running at sea. - [33:21](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2001s) Google Cloud and DeepMind demonstrated spatial AI analysis with Team USA, utilizing pose-estimation models and Cloud TPUs to convert 2D broadcast footage into interactive 3D physics rigs. - [39:28](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2368s) The AI Hypercomputer addresses scaling physics through liquid cooling, TPU v8, Google Axion Arm CPUs, early deployment of NVIDIA Vera Rubin NVL72, and the 47 Pbps Virgo Network fabric. - [47:24](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2844s) Citadel Securities built an elastic quantitative research platform across TPU Ironwood chips, cutting runtimes from days to hours while reducing compute costs by 30%. - [50:01](https://www.youtube.com/watch?v=11PBno-cJ1g&t=3001s) The Agentic Data Cloud integrates the Knowledge Catalog and Lightning Engine for Spark to query across multi-cloud Apache Iceberg lakehouses on AWS and Azure without data egress fees. - [01:04:40](https://www.youtube.com/watch?v=11PBno-cJ1g&t=3880s) Autonomous triage agents and Mandiant global telemetry address the reality of negative seven-day exploit publication and 22-second lateral attacker handoffs. - [01:08:07](https://www.youtube.com/watch?v=11PBno-cJ1g&t=4087s) Wiz's AI Application Protection Platform deploys autonomous Red, Blue, and Green agents across multi-cloud environments to continuously identify, validate, and patch vulnerabilities in code repositories. - [01:21:40](https://www.youtube.com/watch?v=11PBno-cJ1g&t=4900s) YouTube TV successfully transitioned 100% of its telephone support calls to a multi-agent conversational voice system built and deployed in six weeks using CX Agent Studio. - [01:18:37](https://www.youtube.com/watch?v=11PBno-cJ1g&t=4717s) Enterprise retail adopters including The Home Depot and Reliance utilize multimodal Gemini models for conversational commerce, wayfinding, and real-time catalog enrichment. - [01:31:26](https://www.youtube.com/watch?v=11PBno-cJ1g&t=5486s) Google Workspace migration tooling enables enterprises to transition from Microsoft 365 up to five times faster, eliminating the productivity 'context tax' through native cross-app grounding. - [01:35:53](https://www.youtube.com/watch?v=11PBno-cJ1g&t=5753s) Google Cloud differentiates its enterprise strategy around an open architecture that avoids proprietary walled gardens, offering customers full flexibility over models, silicon, and data locations.

## Detailed walkthrough ### [02:09](https://www.youtube.com/watch?v=11PBno-cJ1g&t=129s) Moving from AI Experimentation to Enterprise Production

Google Cloud CEO Thomas Kurian takes the stage at Google Cloud Next '26, looking back at the promises made twelve months prior and declaring that the future envisioned then is now operating live in production at an unprecedented global scale. Over the preceding year, the enterprise narrative moved rapidly beyond initial curiosity and tentative adoption into genuine structural transformation. Today, nearly 75% of Google Cloud customers actively leverage Google AI products to run their daily operations, deploying thousands of specialized agents and automated services across sectors to reach billions of end consumers via Google’s partner network. Kurian stresses that the enterprise community has decisively moved beyond pilots, and the era of isolated experimentation is now over.

The core challenge confronting businesses now is scaling AI out of silos and across the entire enterprise workflow. Kurian cautions that this cannot be solved by assembling a patchwork of disconnected pieces, noting that organizations cannot achieve dependable enterprise value by stitching together fragmented silicon, isolated models, and brittle glue code. Instead, he argues that sustained business value requires a tightly integrated architectural stack where custom silicon is co-designed around frontier models, models are directly grounded in proprietary enterprise data, and agents are secured natively by the underlying infrastructure. Google relies on this identical unified OpenStack internally to reinvent how it powers global services across Search, YouTube, Chrome, and Android.

- Nearly 75% of Google Cloud customers currently leverage Google Cloud AI products. - Enterprises have completed experimental pilots and are shifting toward cross-organizational enterprise production. - Kurian emphasizes: “You cannot deliver AI by piecing together a puzzle piece of fragmented silicon and disconnected models.”

---

[04:32](https://www.youtube.com/watch?v=11PBno-cJ1g&t=272s) Sundar Pichai on Google’s CapEx Surge and "Customer Zero" Workflows

Alphabet and Google CEO Sundar Pichai reflects on his 22-year tenure at the company, tracing how Google has navigated successive paradigm shifts from the open web to mobile, and now to artificial intelligence. To power the emerging agentic era, Google is making massive, multi-year infrastructure investments. Pichai details that while Google’s capital expenditures totaled $31 billion in 2022, the company plans to invest between $175 billion and $185 billion in total CapEx in 2026—a nearly sixfold increase in just four years. Furthermore, for 2026, over half of Google’s total machine learning compute capacity is designated directly for Google Cloud customers, ensuring enterprises can operate at the frontier.

Pichai emphasizes that Google’s product design relies heavily on being "Customer Zero" for its own technologies before bringing them to market. Internally, Google has radically overhauled its software development life cycle: nearly 75% of all new code checked in at Google is now generated by AI and reviewed by human engineers, up from 50% the previous fall. Software teams are transitioning to autonomous digital task forces composed of planner, orchestrator, and coder agents. In a recent complex enterprise code migration, this multi-agent triad accelerated completion by six times compared to baseline metrics from a year earlier, prompting Google to implement the same agentic approach across its entire pipeline of experimentation, testing, and evaluation.

Beyond engineering, AI is reshaping core marketing and cybersecurity workflows across Alphabet. In marketing, creative teams used generative models for the launch of Gemini in Chrome to programmatically create thousands of asset variations tailored by channel, audience, and geography, reducing turnaround times by 70% while boosting conversions by 20%. In security, the Security Operations Center (SOC) previously faced an overwhelming volume of unstructured monthly threat reports. Autonomous SOC agents now ingest, parse, and triage tens of thousands of these reports each month, cutting threat mitigation times by more than 90%.

- Google CapEx grew from $31 billion (2022) to an anticipated $175B–$185 billion (2026). - Over 50% of Google’s 2026 ML compute allocation is earmarked specifically for Cloud. - Over 75% of newly authored code at Google is AI-generated and approved by engineers. - Multi-agent development task forces completed a major code migration 6x faster than previous standards. - Marketing asset personalization drove a 70% turnaround reduction and a 20% conversion lift. - Automated security triage reduced incident threat mitigation time by over 90%. - Pichai asserts: “The conversation has gone from can we build an agent to how do we manage thousands of them?”

---

[08:48](https://www.youtube.com/watch?v=11PBno-cJ1g&t=528s) Introducing the Gemini Enterprise Agent Platform

Sundar Pichai announces the launch of the Gemini Enterprise Agent Platform. While Gemini Enterprise was introduced the previous fall as an accessible entry point to bring agentic workplace AI to employees, the speed of adoption has surfaced new organizational challenges. Individual employees across departments have rapidly transitioned into builders, quickly multiplying the number of autonomous workers inside corporations. Consequently, corporate IT and engineering discussions have moved from whether an agent can be constructed to the governance, security, and orchestration of thousands of interconnected agents.

Pichai describes the Gemini Enterprise Agent Platform as the foundational connective tissue that provides full-stack security and operational control across the agent life cycle. It is designed to act as an enterprise mission control center, allowing businesses to build, scale, monitor, govern, and optimize thousands of autonomous agents with confidence.

```mermaid flowchart TD accTitle: Gemini Enterprise Agent Platform Architecture accDescr: Visual representation of how enterprise context flows through the Gemini Enterprise Agent Platform into the agentic layers.

subgraph Inputs ["Enterprise Inputs"] D["Enterprise Data & Context"] L["Business Logic & Rules"] end

subgraph Platform ["Gemini Enterprise Agent Platform"] C["Mission Control & Governance"] O["Multi-Agent Orchestration"] C --- O end

subgraph Execution ["Agentic Execution"] T["Agentic Task Force"] A["Custom Enterprise Agents"] end

Inputs --> Platform Platform --> Execution ```

---

[10:20](https://www.youtube.com/watch?v=11PBno-cJ1g&t=620s) The Blueprint for the Agentic Enterprise

Returning to the stage, Thomas Kurian frames Gemini Enterprise as the holistic end-to-end operating environment for the agentic era. He defines the fundamental equation of modern business transformation as: intelligence plus automation must equal tangible enterprise value. Kurian breaks this down into two necessary elements: context and action. Intelligence is extracted directly from enterprise data, while automated action is driven by agents operating within governed boundaries. To deliver this at scale, Google Cloud is structuring its agentic blueprint across five integrated layers:

1. **AI Hypercomputer:** The purpose-built compute infrastructure optimized for the scale, latency, and networking demands of agentic workloads. 2. **Agentic Data Cloud:** The foundational data engine that grounds AI models and agents in real-time, trusted business context. 3. **Agentic Defense:** Autonomous, end-to-end security architecture protecting data, models, and agents across their life cycle. 4. **Gemini Enterprise Agent Platform:** The governance and development platform where models, data, and business rules converge to build and orchestrate agents. 5. **Agentic Task Force:** A suite of specialized, pre-built domain agents ready for immediate enterprise deployment.

Kurian explains that this architecture elevates Vertex AI by pairing enterprise business logic directly with underlying models and datasets, giving IT organizations the same architectural rigor, observability, and compliance standards they apply to their most mission-critical systems.

- Kurian summarizes the operational imperative: “Intelligence plus automation must deliver value.” - System pillars span physical silicon (AI Hypercomputer), data grounding (Agentic Data Cloud), safety (Agentic Defense), orchestration (Agentic Platform), and turnkey workers (Agentic Task Force).

---

[12:32](https://www.youtube.com/watch?v=11PBno-cJ1g&t=752s) Frontier Reasoning Models and Global Enterprise Deployments

At the model layer of the platform, Kurian highlights Google’s latest state-of-the-art foundation models and partner offerings. Available in preview, Gemini 3.1 Pro serves as Google’s flagship reasoning engine, purpose-built for orchestrating multi-step autonomous workflows and executing API calls with minimal prompt engineering. Software leaders like Databricks, JetBrains, and Replit have integrated Gemini 3.1 Pro into their developer platforms. Alongside Gemini 3.1 Pro, Kurian unveils several additional multimodal models available in preview: Gemini 3.1 Flash Image (internally dubbed Nano Banana 2) for high-fidelity asset rendering, Veo 3.1 Lite for high-volume and cost-efficient enterprise video generation, and Lyria 3 Pro for studio-grade audio and music production. In addition to first-party models, Google Cloud expanded its partnership with Anthropic by adding support for Claude Opus 4.7 alongside existing Claude Sonnet and Haiku models.

Kurian then reveals major commercial milestones powered by Google’s stack. Apple has selected Google Cloud as its preferred cloud provider to collaborate on developing next-generation Apple foundation models based on Gemini technology, which will power upcoming Apple Intelligence features, including an updated, more personalized Siri arriving later in the year. In banking, Citi Wealth—partnering with Google Cloud and Google DeepMind—unveiled Citi Sky, an always-on multilingual AI wealth management assistant designed to equip advisors and clients with instant access to Citi’s global institutional intelligence. Industrial leader Honeywell is using Gemini models to train digital twins across more than one million product specifications to yield billions of building management insights, while Mexican department store chain Liverpool is scaling its personalized in-store advisory model to its digital storefront via a virtual shopping assistant projected to generate a tenfold return on investment.

| Model / Partner | Modality & Specialty | Availability / Adoption Status | |:--- |:--- |:--- | | **Gemini 3.1 Pro** | Advanced reasoning & complex workflow orchestration | Preview; adopted by Databricks, JetBrains, and Replit | | **Gemini 3.1 Flash Image (Nano Banana 2)** | High-fidelity visual asset generation | Preview | | **Veo 3.1 Lite** | Cost-effective, high-volume video synthesis | Preview | | **Lyria 3 Pro** | Professional-grade enterprise audio and music | Preview | | **Anthropic Claude Opus 4.7** | Advanced frontier model support | Available on Vertex AI alongside Sonnet and Haiku | | **Apple** | Foundation model collaboration using Gemini | Preferred cloud partner; powering Apple Intelligence and personalized Siri | | **Citi Wealth** | Citi Sky client/colleague intelligence tool | Built in collaboration with Google Cloud and DeepMind | | **Honeywell** | Building management digital twins | Trained on >1M engineering and product specifications | | **Liverpool** | Retail conversational shopping assistant | Projecting 10x ROI on e-commerce implementation |

[15:42](https://www.youtube.com/watch?v=11PBno-cJ1g&t=942s) Building, Governing, and Connecting Agents with Gemini Enterprise

Google Cloud has partnered with NASA to deploy Gemini Enterprise Agents in high-stakes environments, powering flight readiness and ensuring astronaut safety for Artemis II, the historic mission that set the human spaceflight record for traveling the farthest distance from Earth. Building on that foundation, Google Cloud designed its agent platform to govern the complete lifecycle of enterprise agents. The low-code Agent Studio allows any employee across an organization to build and deploy agents using plain natural language, grounding large language model reasoning directly in enterprise-specific business rules to guarantee predictable, autonomous workflow actions at scale.

```mermaid flowchart TD accTitle: Gemini Enterprise Agent Lifecycle and Ecosystem accDescr: Architecture of the Gemini Enterprise Agent Platform showing creation, registration, ecosystem integrations, and protocol connectivity. subgraph Creation["Authoring & Definition"] AS["Agent Studio (Low-Code)"] STR["Skills & Tools Registry"] end

subgraph Governance["Central Control"] AR["Agent Registry"] end

subgraph Integrations["Ecosystem & Connectivity"] AM["Agent Marketplace<br/>(Atlassian, Box, Lovable, Oracle, ServiceNow, Workday)"] MCP["Model Context Protocol (MCP)"] GCP["GCP & Workspace Services"] end

AS --> AR STR --> AR AR --> AM AR --> MCP GCP --> MCP ```

To manage these digital assets, Google Cloud introduced the Agent Registry as a unified point of control, cataloging every internal agent and tool across the company so they remain discoverable and compliant. Alongside it, the Skills and Tools Registry lets developers package reusable, modular bundles of instructions, scripts, and resources to train agents on specialized tasks. Google Cloud is exposing dedicated skills for every Google Cloud Platform (GCP) service and Workspace application within this registry. Organizations can also tap into the Agent Marketplace to discover and deploy pre-built specialized agents from partner companies, including Atlassian, Box, Lovable, Oracle, ServiceNow, and Workday. Furthermore, native integration with the Model Context Protocol (MCP) enables the platform to connect with any external MCP server, while all GCP services are exposed as MCP endpoints for frictionless programmatic control.

- NASA utilized Gemini Enterprise Agents for Artemis II astronaut safety and flight readiness. - Low-code Agent Studio translates natural language instructions into grounded autonomous actions. - Agent Registry and Skills and Tools Registry provide modular reuse and centralized discovery. - Fully supports open interoperability via native Model Context Protocol (MCP) across all GCP services.

[18:01](https://www.youtube.com/watch?v=11PBno-cJ1g&t=1081s) Agent-to-Agent Orchestration, Zero-Trust Security, and Projects

Scaling an agentic workforce requires robust coordination between autonomous systems. The platform supports agent-to-agent orchestration, letting independent agents delegate tasks back and forth using both complex generative and deterministic orchestration patterns. For strictly regulated or audit-heavy workflows, deterministic orchestration ensures agents adhere to rigid, pre-specified execution paths every single time. These workflows can be triggered across a variety of runtime patterns: real-time events, fixed schedules, ad-hoc system triggers, or batch inference jobs.

```mermaid flowchart LR accTitle: Agentic Security and Execution Architecture accDescr: Diagram illustrating the security, gateway, and observability layers wrapping agent orchestration. subgraph Ingestion["Triggers & Events"] TR["Real-time / Schedule / Batch"] end

subgraph Security["Zero-Trust Security Layer"] AG["Agent Gateway"] AID["Agent Identity (Crypto ID)"] MA["Model Armor"] end

subgraph Execution["Agent Orchestration"] AO["Deterministic & Generative Flows"] end

subgraph Monitoring["Observability"] OTel["OTel Instrumentation & Tracing"] end

TR --> AG AG --> AID AID --> MA MA --> AO AO --> OTel ```

Enterprise security is enforced through an end-to-end zero-trust architecture. Through Agent Identity, every single agent receives a unique cryptographic ID accompanied by auditable, traceable authorization policies to track every autonomous action across corporate networks. The Agent Gateway acts as a central command console for organization-wide policy enforcement, working in tandem with Model Armor to defend models and sensitive enterprise data against leakage. For complete visibility, the platform integrates OpenTelemetry-compliant (OTel) observability, allowing engineering teams to inspect execution traces, track tool invocation, and diagnose reasoning loops through granular telemetry.

While engineers operate within the underlying agent platform, end users interact through the Gemini Enterprise application—the organizational front door to AI. To solve the problem of memory retention across ongoing business tasks, Google introduced Gemini Enterprise Projects. Projects provide a dedicated, high-fidelity workspace with long-term memory where agents apply deep think capabilities without context pollution. The application also supports Microsoft 365 interoperability, enabling users to export Google Docs and Slides created inside Canvas straight into native Microsoft Office formats.

[22:02](https://www.youtube.com/watch?v=11PBno-cJ1g&t=1322s) Customer Spotlight: Virgin Voyages on Distributed Edge and Rovey

Virgin Voyages demonstrated how deploying Gemini Enterprise drives real business value across its cruise operations. Speaking on their people-first operational philosophy, Virgin Voyages CEO Nirmal Saverimuttu remarked:

> “Our people are a secret sauce. So anytime we can use data to make our people more empowered, we're all in.”

To deliver continuous AI capabilities on the high seas where standard connectivity fails, Virgin Voyages turned to Google Distributed Cloud Edge. Billy Bohan Chinique highlighted the challenge of maritime environments, noting: “There's not exactly a fiber cable running out the back of the ship. With Google Distributed Cloud Edge, we had data resiliency at the core of the project. And that means, even when we're offline, our intelligence is still online.”

Through "Project 3D" and their virtual concierge named Rovey, Virgin Voyages assists sailors from pre-trip vacation planning through disembarkation, while giving onboard crew members instant access to operational answers. Saverimuttu added that “Gemini Enterprise is a front door for AI in our business, and we're using it to bring natural language intelligence to our crew.” The results have shown measurable business impact: - Production timelines dropped by up to 60%. - Month-over-month sales grew by 28%, culminating in a record sales quarter. - Virgin envisions a unified customer journey where travelers transition seamlessly from a Virgin Atlantic flight to a Virgin Hotel and onto a Virgin Voyage, all coordinated by Gemini Enterprise and Rovey.

[23:58](https://www.youtube.com/watch?v=11PBno-cJ1g&t=1438s) Cross-Agent Workflow in Action: From Strategy to Code and Marketing

Thomas Kurian returned to the stage to underline the industry inflection point:

> “The era of the pilot is over. The era of the agent is here.”

Kurian then introduced Erika Chuong to demonstrate an end-to-end multi-agent business workflow. Playing the role of an employee at a global furniture retailer, Chuong showcased her personalized Gemini Enterprise homepage, which aggregates internal enterprise systems, external live data, and an approved agent gallery into a single pane of glass.

```mermaid flowchart TD accTitle: Multi-Agent Retail Campaign Execution accDescr: Step-by-step workflow showing how multiple agents collaborate to revive a retail furniture line. subgraph Strategy["Research & Strategy"] P1["Prompt: Analyze Trends & Dead Stock"] --> MR["Market Research Agent (Deep Research)"] P1 --> DI["Data Insights Agent (Agentic Data Cloud)"] MR & DI --> PS["Product Strategy Agent"] end

subgraph Production["Asset Generation & Development"] PS -->|"Rebrand & Reprice Recommendation"| VEO["Veo 3.0 Video Generation"] PS -->|"Generate Jira Ticket"| DEV["Dev Agent (CLI & Chat)"] DEV -->|"Deploy Web Code"| SITE["Updated Landing Page"] end

subgraph Field["Enablement"] SITE --> SO["Store Operations Session"] SO --> GWA["Google Workspace Agent"] GWA --> DECK["Branded Distributor Deck"] end ```

To revive slow-moving products, Chuong issued a single high-level natural language prompt asking the system to research interior design trends, identify warehouse dead stock, and launch a marketing campaign:

1. **Market Research Agent**: Using Deep Research, the agent analyzed current Google Search trends alongside internal sales figures and CSAT feedback, identifying "Organic Modern" as a surging high-spend trend. 2. **Data Insights Agent**: Querying the global catalog across the Agentic Data Cloud, it mapped "dead stock" to stale inventory in the retailer's "Tuscany" collection. 3. **Product Strategy Agent**: Synthesized the findings and recommended rebranding and repricing the Tuscany line above clearance discounts but below competitor prices, while proposing a complete website refresh. 4. **Creative Media Generation**: The strategy agent called Veo 3.0, generating high-definition video assets placing the existing furniture pieces into styled, modern living spaces. 5. **Developer Handoff**: The dev agent created a detailed Jira ticket, dispatched an alert into Google Chat, and picked up the task in the command-line interface (CLI) to build and deploy the landing page according to brand and coding standards. 6. **Store Operations Enablement**: In a separate Gemini Enterprise session, a store operations prompt utilized organizational context—such as past launch workflows and team sales targets—collaborating with the Google Workspace agent to generate a custom, branded presentation deck for regional distributors.

[28:42](https://www.youtube.com/watch?v=11PBno-cJ1g&t=1722s) Gemini Enterprise Canvas Mode and Enterprise Governance

Wrapping up the live demonstration, Erica shows how content generation integrates smoothly into workplace collaboration. Because generating slide decks can take a minute or two, she pulls up a deck generated ahead of time to spotlight Canvas mode. This allows users to make granular edits and collaborate directly within Google Slides without ever having to leave Gemini Enterprise. Reviewing the presentation, she notes that it looks great overall, swaps in the word "reimagined," and shares it with her team, watching colleagues jump into the deck in real time.

Erica delivers the core takeaway of the demo: a single prompt can orchestrate multiple agents, pass shared context between them, and produce collaborative content in Google Workspace. Backing every step is enterprise-grade governance designed to let organizations manage and protect agents at scale while maintaining total custody of their business data. “That's the magic of the Gemini Enterprise Agent platform,” Erica concludes before handing the stage back to Thomas Kurian.

```mermaid flowchart LR accTitle: Gemini Enterprise Agent Orchestration accDescr: Diagram illustrating multi-agent orchestration, workspace integration, and enterprise governance in Gemini Enterprise. subgraph Input ["User Interaction"] A["Single Prompt"] end subgraph Platform ["Gemini Enterprise"] B["Multi-Agent Orchestration"] C["Shared Context & Governance"] end subgraph Workspace ["Collaboration"] D["Canvas Mode"] E["Google Slides / Workspace"] end A --> B B --> C C --> D D --> E ```

[30:03](https://www.youtube.com/watch?v=11PBno-cJ1g&t=1803s) Global Enterprise Adoption: From Signal Iduna to Walmart

Thomas Kurian returns to detail how thousands of organizations across industries are adopting Gemini Enterprise, driving a fundamental organizational shift. “Companies aren't just redesigning workflows,” Kurian explains. “They're turning their everyday employees into AI builders, empowering them to solve their own hardest problems.” He runs through early enterprise proof points across insurance, manufacturing, professional services, and healthcare:

- **Signal Iduna:** Achieved 80% adoption within weeks in Germany, making Gemini Enterprise their AI front door. Today, 11,000 employees build specialized agents, including a health agent that evaluates customer coverage across a century of historical policy data, producing answers 37% faster amid a 400% surge in weekly active users. - **Bosch:** Deploying the platform globally across business lines from finance to engineering, freeing up team bandwidth for intensive research. - **KPMG:** Reached 90% adoption within the very first month, spinning up over 100 specialized agents. - **American Society of Clinical Oncology (ASCO):** Utilizing the platform to deliver clinical cancer expertise more rapidly. - **Merck:** Equipping 75,000 employees with Gemini Enterprise to accelerate their core mission to save and improve lives.

Kurian highlights Walmart's extensive rollout to empower retail managers on the front lines. Kieran Shanahan, Walmart's EVP and Chief Growth Officer, explains in a video segment that retail execution depends on interpersonal trust: field leaders need to be active out on the sales floor with associates and shoppers rather than tethered to back-office screens. To make this possible, Walmart equips store and supply chain leaders with Google Pixel Fold devices connected straight to enterprise systems, allowing them to pull critical operational answers in seconds instead of hours.

[33:21](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2001s) Team USA Collaboration and Shaun White on Stage

Kurian pivots to Google Cloud's sports engineering partnerships, explaining how Google engineers worked directly with Team USA winter athletes ahead of major international competitions like the Milan Winter Olympics. As team members Michael Santoro, Katherine Litinsky, and Sarah Kennedy explain, action sports often suffer from sparse training data, making them an ideal testbed for advanced vision and 3D modeling developed alongside Google DeepMind. “We've moved so far past talking to an LLM with text,” Litinsky emphasizes, noting that models must process movements in real time at the speed of competitive sports.

Snowboarding legend and three-time Olympic gold medalist Shaun White joins the stage to dynamic rock music and falling digital snow, joking about seeing winter weather in the Las Vegas desert. Reflecting on his career, White recalls how training historically relied on hand-held camcorders and pure intuition: athletes would perform high-risk aerial maneuvers and squint at tiny camera screens guessing how to gain an edge. With trick technicality escalating and half-pipe walls growing higher, White brought his domain expertise to Google Cloud engineers to build tools that eliminate the guesswork.

[35:34](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2134s) Spatial AI Analysis of the Switch Cab Double Flip 1440

Google Cloud engineer Jason Davenport joins Shaun White to break down a legendary clip from White’s winning run at the 2017 Burton US Open. The maneuver—a switch cab double flip 1440—takes place in under three seconds. White describes it as riding in his unnatural stance and executing four complete 360-degree rotations alongside two off-axis flips mid-air. Davenport demonstrates how Google Cloud AI processes the rapid split-second sequence frame-by-frame:

- **3D Spatial Pose Estimation:** Developed with Google DeepMind, the model converts flat, 2D broadcast footage into an interactive, 3D spatial skeletal rig of the athlete, accurately tracking White despite the low-contrast black leather jacket and tight pants he wore in 2017. - **Biomechanical Flight Dynamics:** Gemini-powered metrics extract critical physical parameters, such as tuck compression, rotational speed, and overall flight duration. White highlights that being able to compare airtime between landed and missed tricks provides measurable data that coaches and future generations never had access to before. - **Ribbon Motion Overlay:** A continuous colored path maps out the rotational trajectory of the board and rider, shifting color from blue to green. White explains that this visualization captures the precise, make-or-break apex of the trick—the exact pause between the opening front flip and the transition into the backflip.

Davenport ties the demonstration back to Google Cloud’s underlying infrastructure: specialized spatial analysis models, high-scale training using Cloud TPUs, and a reliable foundation built on the Gemini Enterprise Agent Platform. Reflecting on the impact for both athletes and spectators, Shaun White observes: “Learning a trick on the mountain is one thing, but actually understanding the physics of a trick is a whole other thing.”

[39:28](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2368s) The AI Hypercomputer and the Physics of Enterprise Compute

Amin Vahdat, SVP and Chief Technologist for AI and Infrastructure at Google Cloud, takes the stage following Shaun White. He bridges the gap between athletic performance and enterprise data: while athletes analyze data to master their sport, modern enterprises rely on it to fundamentally transform their industries. Handling continuous massive video streams alongside decades of historical records while defending data in near real time represents the ultimate stress test. Vahdat emphasizes that expanding these capabilities globally transcends standard software engineering—it is fundamentally a physics problem. Engineers are constantly pushing against the boundaries of software architectures, silicon designs, electrical power, thermal cooling, and the speed of light.

```mermaid flowchart LR accTitle: AI Hypercomputer Architecture accDescr: Interconnected layers of Google Cloud AI Hypercomputer from clean energy to data center scale. subgraph Physics ["Physical Foundations"] A["Clean Energy & Power"] B["Liquid Cooling Infrastructure"] end subgraph Compute ["Silicon & Systems"] C["TPU v8 Generations\n(8t Training / 8i Serving)"] D["Google Axion Arm CPUs"] E["NVIDIA Vera Rubin NVL72"] end subgraph Fabric ["Storage & Interconnect"] F["Virgo Network Fabric\n(47 Pbps Bisection)"] G["Managed Lustre\n(10 TBps Throughput)"] end Physics --> Compute Compute --> Fabric ```

To solve this, Google Cloud built the AI Hypercomputer, bringing together clean energy, massive physical scale, and custom silicon into a unified engine of efficiency. Vahdat highlights a core shift in perspective for modern system design: “Because in the agentic era, compute is no longer defined by chip. Compute is the entire data center.” He points to real-world production deployments demonstrating this architectural flexibility:

- **Axia Energia**: Brazil’s largest power utility runs complex weather simulations on TPU clusters, predicting severe storm fronts up to 10 days in advance to execute proactive preventative measures and reduce outages year over year. - **Thinking Machine Labs**: Powers Tinker, an open-source platform for reinforcement learning and fine-tuning frontier models for specialized enterprise tasks, on NVIDIA-based Google Cloud infrastructure. - **Woven by Toyota**: Accelerated training times by 42% on models designed to predict complex, split-second traffic incidents. - **US Department of Energy**: Deploys autonomous AI co-scientists across all 17 national laboratories to accelerate the pace of scientific discovery. - **Boston Dynamics**: Trains and safely deploys flexible vision-language models (VLMs) to scale industrial robotics across varied factory settings.

[42:16](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2536s) Eighth-Generation TPUs: Splitting Training and Serving (TPU 8t and TPU 8i)

Vahdat announces that because the computing demands of model training and model serving have diverged dramatically in the agentic era, Google Cloud is for the first time launching two distinct, specialized hardware platforms: the eighth-generation Tensor Processing Units, designated TPU 8t and TPU 8i. Calling them “things of beauty,” he explains that building each platform from scratch allows hardware architects to eliminate bottlenecks specific to training runs versus high-concurrency inference and reinforcement learning.

```mermaid flowchart TD accTitle: TPU Eighth Generation Architectural Divergence accDescr: Workload divergence between TPU 8t training and TPU 8i inference platforms. A["Eighth-Generation TPU Platform"] --> B["TPU 8t: Training Powerhouse"] A --> C["TPU 8i: Inference & RL Engine"]

B --> B1["Native MXU Block-Scale Quantization"] B --> B2["9,600 TPUs in 3D Torus Topology"] B --> B3["1.21 Exaflops FP4 Compute per Pod"] B --> B4["2 PB Shared Memory with TPUDirect Storage"]

C --> C1["Collectives Acceleration Engine (5x lower latency)"] C --> C2["On-Silicon Memory Cache (Breaks Memory Wall)"] C --> C3["Board Flight Topology (1,152 TPUs per Pod)"] C --> C4["11.6 Exaflops FP8 Compute (9.8x vs. Ironwood Pod)"] ```

The TPU 8t serves as the dedicated training powerhouse. By incorporating block-scale matrix multiplication directly inside the Matrix Multiply Units (MXUs), Google eliminated Vector Processing Unit (VPU) quantization overhead, delivering nearly triple the compute performance per pod over prior hardware generations. TPU 8t features an upgraded inter-chip interconnect with double the bandwidth of TPU Ironwood, scaling up to 9,600 TPUs in a unified 3D torus topology to output 1.21 FP4 exaflops per pod. Each superpod provides two petabytes of shared memory bandwidth—enough to hold the entire digital collection of the US Library of Congress one hundred times over—and pairs with TPUDirect storage for high-speed streaming from managed data stores.

Conversely, TPU 8i is tailored for real-time model thinking, reasoning, and reinforcement learning. It integrates a dedicated on-chip collectives acceleration engine that slashes inter-chip latency by an additional 5x. By hosting the memory cache directly on silicon, TPU 8i overcomes the traditional memory wall that cripples long-context auto-regressive decoding. Deployed through a novel board flight network topology containing 1,152 TPUs per pod, it delivers 11.6 FP8 exaflops—a 9.8x performance surge over a standard 256-chip Ironwood pod—designed to host millions of concurrent agent sessions with near-zero latency.

[45:05](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2705s) Google Axion, NVIDIA Vera Rubin NVL72, Managed Lustre, and the Virgo Network

Vahdat turns to general-purpose computing and accelerated GPU systems. Google Cloud Axion, powered by custom-designed Arm-based CPUs, powers the new N48 compute instances. These deliver up to 2x better price-performance and 80% higher performance per watt than comparable x86 instances, running sustained, continuous execution designed to eliminate cold starts in logic apps and keep agents persistently active. Vahdat also announces that Google Cloud will be among the first cloud providers to deploy the NVIDIA Vera Rubin NVL72 system, achieving a 10x leap in performance efficiency specifically tuned for long-context windows and interactive multi-agent reasoning.

Recognizing that advanced processors starve without equivalent data throughput, Google re-engineered its storage and networking stack. Managed Lustre now achieves an industry-leading throughput rate of up to 10 terabytes per second. To connect large-scale clusters, Google unveiled the Virgo Network fabric:

- **Virgo Network Scale**: Extends inter-superpod connectivity across 134,000 chips using up to 47 petabits per second of non-blocking bisection bandwidth, driving aggregate compute capacity past 1.7 million exaflops. - **Accelerated Turnaround**: Compresses model training runs that previously required several months down into a matter of weeks by orchestrating over one million TPU accelerators in a single cluster via Google Pathways and JAX. - **Bandwidth Efficiency**: Supplies up to 4x higher bandwidth per accelerator compared to prior generations, yielding near-linear scaling curves. - **GPU Compatibility**: The Virgo network architecture is also being deployed for NVIDIA Vera Rubin NVL72 fabrics, capable of scaling to support clusters of up to 960,000 GPUs.

[47:24](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2844s) Citadel Securities on Nanosecond Research Scale

Vahdat introduces Citadel Securities to demonstrate how high-frequency trading and market-making operations leverage Google Cloud’s silicon and network architecture to secure a competitive edge. Josh Woods, representing Citadel Securities, explains the operational scale: the firm executes trades exceeding half a trillion dollars daily, where success hinges on executing calculations within nanosecond margins across global pension funds and retail retirement accounts.

```mermaid flowchart LR accTitle: Citadel Securities Cloud Research Acceleration accDescr: The progression of quantitative research timelines at Citadel Securities using Google Cloud TPUs. A["Quantitative Researcher\n(100 Novel Ideas)"] --> B["Traditional Cluster Execution\n(Weeks or Days)"] A --> C["Google Cloud TPU Ironwood\n(Hours or Minutes)"] C --> D["2x to 4x Faster Runtime\n30% Cost Reduction"] ```

Citadel Securities partnered with Google Cloud to construct an elastic, cloud-based quantitative research platform. Working closely with Google engineers on TPU architectures, the firm now runs parallel workloads across thousands of TPU Ironwood chips simultaneously. Haris Nair of Citadel Securities notes that research workloads that previously dragged on across multiple days or weeks are now completed in hours or minutes. This computational shift allows quantitative researchers with 100 novel market hypotheses to test their models without hitting hardware or financial limits, constrained only by their creative insight. Woods summarizes the partnership’s intensity: “As we like to say, steel sharpens steel. Google Cloud and Citadel Securities' collaboration has been exactly that.”

[49:00](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2940s) Autonomous Cloud Telemetry and the Agentic Data Cloud

Vahdat returns to emphasize that managing data centers operating at hypercomputer scale is beyond manual human oversight: “You need a cloud that drives itself.” Google implemented the Model Context Protocol (MCP) to expose every Google Cloud service as an orchestratable tool for autonomous agents. By injecting Gemini reasoning directly into internal cloud telemetry pipelines, the infrastructure autonomously conducts root-cause analysis, discovering and rectifying configuration drifts before human administrators detect an anomaly.

Amin Vahdat then introduces Karthik Narain to explain how enterprise data must feed this compute engine. Narain outlines the central challenge of the next phase of enterprise AI adoption:

> “Amin just showed you how we built the world's most powerful reasoning engine. But here's the reality. Reasoning without context is just a guess. And when you expect your AI to make decisions and your agents to take actions, you cannot afford to guess. Trusted context turns an intelligent guess to a decisive action.”

Narain introduces the Agentic Data Cloud, grounded in four core architectural innovations aimed at unlocking data trapped in corporate silos, PDFs, video recordings, and SaaS systems. The foundational layer is the Knowledge Catalog, designed as a universal context engine for enterprise environments:

- **Structured Data Integration**: Natively maps BigQuery relational tables, views, and operational metadata directly into unified enterprise business logic. - **Smart Storage for Unstructured Data**: Eliminates manual pipeline construction. When raw image files or PDF documents arrive in Google Cloud Storage buckets, they are automatically parsed, tagged, enriched with metadata, and made agent-ready upon ingestion. - **Semantic Understanding**: Using native Gemini multi-modal reasoning, the Knowledge Catalog continuously reads files, extracts named entities, constructs relationship graphs, and learns domain-specific enterprise semantics—ensuring that terms like "net revenue" or "risk" map to identical corporate calculations across all deployed autonomous agents. - **Enterprise Adoption**: Virgin Media O2 is already utilizing the Knowledge Catalog to unlock and contextualize value across more than 20,000 previously dormant data assets.

[52:06](https://www.youtube.com/watch?v=11PBno-cJ1g&t=3126s) The Agentic Data Cloud: Knowledge Catalog, Data Agent Kit, and Lightning Engine

Direct zero-copy access to enterprise systems, applications, and operating platforms—including Palantir, Salesforce, SAP, ServiceNow, and Workday—supplies unified context across the entire enterprise so agents can act with certainty. To extend that power across every employee's daily workflow, Google Cloud integrated its Knowledge Catalog with the Gemini Enterprise Deep Research Agent. The integration performs multi-step reasoning across internal enterprise documents and web data, delivering precise citations and compressing weeks of manual research into seconds.

Building on that contextual foundation, Google Cloud introduced the Data Agent Kit to transform data science workflows directly inside the developer's existing tools rather than forcing them into a new platform. The kit brings AI skills and plugins straight into IDEs, notebooks, and terminals—spanning VS Code, Cloud Code, and the Gemini CLI:

```mermaid flowchart LR accTitle: Data Agent Kit Workflow accDescr: Shows the journey from stating business intent to running autonomous execution on the Agentic Data Cloud. subgraph Input ["User Input"] A["State Business Intent<br/>(e.g., 'Predict customer churn')"] end subgraph DevEnv ["Workspace Environments"] B["VS Code / Cloud Code / Gemini CLI"] end subgraph Execution ["Agentic Data Cloud"] C["Autonomously Build Pipelines"] D["Deploy Models & Orchestrate"] end A --> B B --> C C --> D ```

Data scientists simply state a business intent—such as "predict customer churn"—and the environment autonomously generates data pipelines and deploys machine learning models onto the Agentic Data Cloud. At Bayer Crop Science, researchers use these agentic capabilities to eliminate routine data engineering and analytics overhead, freeing them to focus entirely on agricultural innovations for sustainable food production.

Addressing the severe compute pressures created by agentic workloads, Google Cloud unveiled the Lightning Engine for Apache Spark. Rebuilt from the ground up to supply autonomous agents with real-time context at massive scale, the engine delivers up to 2x the price-performance of the previous proprietary market leader. Early adopters operating large-scale Spark clusters include Flipkart, Lowe's, and Meesho.

- Knowledge Catalog integration: Powers the Gemini Enterprise Deep Research Agent with multi-step reasoning and full citations across internal and public web data. - Data Agent Kit: Embeds autonomous pipeline creation, model deployment, and orchestration natively into VS Code, Cloud Code, and the Gemini CLI. - Lightning Engine for Apache Spark: Delivers 2x price-performance improvement over the prior market-leading Spark engine for high-scale agentic data processing.

---

[55:23](https://www.youtube.com/watch?v=11PBno-cJ1g&t=3323s) The Cross-Cloud Lakehouse and Enterprise Implementations

Addressing the persistent debate over physical data location, the speaker noted that enterprise data inevitably spans Google Cloud, AWS, Microsoft Azure, and diverse SaaS applications. Calling the traditional lakehouse model—which required storage and analytical compute to sit inside the exact same cloud—fundamentally broken, Google Cloud introduced the Cross-Cloud Lakehouse.

```mermaid flowchart TD accTitle: Cross-Cloud Lakehouse Architecture accDescr: Cross-Cloud Lakehouse built on Apache Iceberg connecting engines to distributed clouds without data replication. subgraph MultiCloud ["Distributed Data Sources"] AWS["AWS S3<br/>(Apache Iceberg)"] AZ["Microsoft Azure"] SaaS["SaaS Applications<br/>(SAP, Salesforce, Workday)"] end subgraph EngineLayer ["Google Cloud Analytical Layer"] CCL["Cross-Cloud Lakehouse Engine"] LE["Lightning Engine for Spark"] BQ["BigQuery & Spanner Context"] end AWS <-->|"Low-Latency Direct Access (Zero Egress/No Copy)"| CCL AZ <-->|"Low-Latency Direct Access (Zero Egress/No Copy)"| CCL SaaS <-->|"Direct Zero-Copy Access"| CCL CCL --- LE CCL --- BQ ```

Built upon the open Apache Iceberg format, the Cross-Cloud Lakehouse delivers borderless analytics. Instead of requiring complex networking configurations or incurring massive data egress fees to move datasets, it provides low-latency direct connectivity across AWS and Azure as if the data were stored natively in Google Cloud. Eliminating forced data duplication removes vendor lock-in and gives autonomous agents access to global business context.

Several global enterprises rely on Google's Agentic Data Cloud to anchor large-scale agentic transformations: - **Vodafone:** Unified network operations data in BigQuery and rolled out hundreds of Gemini Enterprise agents to proactively detect and remediate outages, optimizing infrastructure and saving millions of dollars annually. - **Macquarie Bank:** Deployed Gemini Enterprise to its entire workforce and launched a 24/7 autonomous customer-facing banking assistant named "queue" for 2 million customers. Unified data across BigQuery and Spanner has helped Macquarie cut client fraud and scam losses by 50%. - **American Express:** Centralized core transaction and customer data on Google Cloud to accelerate global fraud and risk analysis. - **Costco:** Deployed BigQuery to drive operational member insights, giving store associates instant visibility to optimize member experiences across millions of shoppers.

---

[58:44](https://www.youtube.com/watch?v=11PBno-cJ1g&t=3524s) Live Demo: From Trend to $15 Million Launch in Five Minutes

Yasmeen Ahmad took the stage to demonstrate how the Agentic Data Cloud solves real-world data bottlenecks, promising to compress a decision cycle that usually takes weeks into a five-minute sprint:

> “Most companies take five weeks to turn a trend into a decision. We're going to do it in five minutes.” — Yasmeen Ahmad

To do this, the demonstration tackled three core barriers: dark data using Knowledge Catalog, data silos using the Cross-Cloud Lakehouse, and manual coding via Google's new agentic skills and tools.

The scenario began when automated agentic workflow triggers identified a viral consumer food trend: "Midnight Swirl" frozen yogurt. Before commercializing the flavor, Ahmad needed to address three critical business hurdles: safety (hidden allergens), market sizing (soy-sensitive customer segmentation), and return on investment.

```mermaid flowchart TD accTitle: Demo Workflow: Viral Trend to Investment Decision accDescr: Step-by-step resolution of safety, cross-cloud audience filtering, and revenue forecasting within five minutes. subgraph Step1 ["1. Safety & Dark Data Extraction"] P1["Recipe PDF: 'Base 204'"] P2["Supplier Manual: 'Base 204 = Soy'"] KC["Knowledge Catalog + Gemini Entity Extraction"] P1 & P2 --> KC KC --> RES1["Identified Allergen: Soy<br/>(Product Specs Citation)"] end subgraph Step2 ["2. Cross-Cloud Sizing"] RES1 --> CCLH["Cross-Cloud Lakehouse"] S3["AWS S3 Iceberg<br/>(Millions of Loyalty Profiles)"] --> CCLH CCLH -->|"Lightning Engine Execution"| RES2["Filtered Soy-Safe Customer Cohort"] end subgraph Step3 ["3. Forecasting & ROI"] RES2 --> DS["Data Science Plan & Monte Carlo Notebook"] DS -->|"2,000 Simulations"| FIN["$15M Validated Demand Forecast"] end ```

First, verifying ingredient safety proved challenging across thousands of recipe and ingredient PDFs. A standard keyword search for "soy" across the Midnight Swirl recipe returned zero direct results because the recipe merely referenced an ingredient designated as "base 204." However, the supplier manual revealed that base 204 contained soy:

> “A simple GenAI search would miss this connection because the information is trapped across two separate PDFs.” — Yasmeen Ahmad

By combining agent skills with Knowledge Catalog, Gemini extracted structured entities across both the supplier manual and recipe PDFs—linking recipe, ingredient, and allergen into a unified `product specs` schema with full data citations.

Second, sizing the market required joining the newly generated allergen schema in Google Cloud with a loyalty database containing millions of customer dietary preferences stored externally in an AWS S3 Apache Iceberg table. Rather than waiting days for a data engineering migration pipeline, Ahmad leveraged the Cross-Cloud Lakehouse. Gemini drafted an execution plan respecting security perimeters, which executed across AWS and Google Cloud using the Lightning Engine to isolate soy-sensitive customers in seconds without copying data.

Finally, to calculate projected annual ROI, Gemini acted as an orchestrator, generating a multi-step data science notebook. Ahmad edited the execution plan to increase Monte Carlo simulations from 1,000 to 2,000 runs ("Yes, Chef," she quipped). The Lightning Engine executed the models to project an annual revenue outcome of $15 million, completing the discovery, multi-cloud query, and predictive forecast in under five minutes.

---

[01:04:40](https://www.youtube.com/watch?v=11PBno-cJ1g&t=3880s) Agentic Defense: Security at Machine Speed

Francis deSouza stepped on stage to address the security realities facing the agentic enterprise. He argued that as autonomous agents take actions across enterprise systems, cybersecurity defense must operate completely autonomously to counter weaponized AI:

> “In an agentic enterprise, giving agents the power to act means that security must become an autonomous force, moving faster than the threats themselves.” — Francis deSouza

DeSouza highlighted alarming telemetry on the acceleration of modern cyberattacks: - **Mean Time to Exploit:** Has fallen to negative seven days (-7 days), meaning threat actors routinely deploy functional exploits a full week before software vendors publish a security patch. - **Threat Actor Lateral Handoff:** The elapsed time between an attacker gaining initial system access and transferring control to a specialized secondary threat group has plummeted from eight hours down to just 22 seconds.

Because human analysts can no longer react within that window, Google is shifting the defensive burden to a Gemini-native, agentic Security Operations Center (SOC). Autonomous triage agents take over investigation workflows, collapsing detailed manual 30-minute triage and analysis routines into rapid, machine-speed resolutions.

[01:05:36](https://www.youtube.com/watch?v=11PBno-cJ1g&t=3936s) Global Telemetry and Securing the Agentic Enterprise with Wiz

Francis deSouza continues highlighting Google Cloud’s defensive speed, noting that threat hunting and detection agents proactively scour enterprise environments for vulnerabilities at speeds human analysts could never match. Backed by Google’s vast telemetry—including Mandiant, VirusTotal, and Chrome—the platform delivers a defense infrastructure grounded in global intelligence, featuring integrated dark web intelligence that spots external threats with 98% accuracy. DeSouza cites CME Group, which manages the world's largest derivatives marketplace; they rely on a unified stack consisting of Mandiant, Security Command Center, and Google SecOps to protect against novel attacks while hitting nanosecond precision and sustaining billions of transactions every single day.

Turning to modern operational hazards, deSouza warns that cybersecurity risk is no longer confined to outside adversaries; internal shadow AI—unauthorized models and agents running untracked across corporate networks—presents an equally dangerous vulnerability. To address this exposure, Google Cloud has woven deep security context directly into its AI architecture. Marked by enthusiastic applause from the audience, deSouza officially welcomes the Wiz team to Google Cloud, explaining that while Gemini Enterprise manages and secures native agents, bringing Wiz into the fold extends that governance to on-premises systems and every major public cloud.

- Threat hunting agents proactively sweep environments at machine scale. - Google telemetry stack integrates Mandiant, VirusTotal, Chrome, and dark web intelligence with 98% threat-identification accuracy. - CME Group uses Google SecOps, Mandiant, and Security Command Center for nanosecond-level processing across billions of daily derivatives transactions. - Google Cloud formally integrates Wiz to safeguard against shadow AI across hybrid and multi-cloud footprints.

[01:08:07](https://www.youtube.com/watch?v=11PBno-cJ1g&t=4087s) Wiz AI Application Protection Platform and Autonomous Agents

Taking the stage to upbeat music, Wiz co-founder Yinon Costica thanks deSouza and reflects on Wiz’s foundational mission: helping enterprises safeguard everything they develop and operate across code, cloud, and runtime. The rise of autonomous tools, hyperscale automated code generation, and weaponized AI attacks has reshaped the threat landscape, demanding security operations that move at machine speed rather than developer speed. To solve these realities, Wiz created the AI Application Protection Platform (AI APP), which tackles four foundational priorities: establishing complete visibility into enterprise AI stacks, preemptively uncovering and remediating critical risks, embedding secure-by-design standards inside AI IDEs for developers, and providing SecOps teams with tools to neutralize active cloud and AI attacks.

Costica explains that Wiz counters automated threat actors by fielding its own lineup of specialized AI agents named after the functional teams they assist: red, blue, and green. Drawing upon Wiz's underlying security graph, these agents autonomously discover, analyze, and remediate high-priority risks. Because a company's modern AI footprint spans diverse foundational models (Gemini, Claude, OpenAI) and disparate development environments (Lovable, Copilot Studio, Salesforce Agentforce, and the Gemini Enterprise Agent Platform), Wiz monitors the full architecture—from cloud infrastructure and databases to individual models and agents—correlating every interdependency to illuminate the attack vectors that present genuine danger.

> “In order to outpace attackers, Wiz delivers a set of expert AI agents.” — Yinon Costica

```mermaid flowchart TD accTitle: Wiz AI APP Tri-Agent Architecture accDescr: Diagram showing how Wiz Red and Green Agents interact with the Security Graph and enterprise tools to validate and fix risks.

subgraph INGESTION["Multi-Cloud & AI Assets"] A["Cloud Infrastructure & DBs"] B["Models: Gemini, Claude, OpenAI"] C["Agent Studios: Lovable, Agentforce"] end

subgraph ENGINE["Wiz Security Graph"] D["Agentless Inventory & Architecture Map"] end

subgraph AGENTS["Autonomous AI Layer"] E["Wiz Red Agent<br/>(Validates Attack Paths)"] F["Wiz Green Agent<br/>(Automates Triage & PRs)"] end

subgraph REMEDIATION["Resolution Pipeline"] G["Developer Pull Request (PR)"] H["CodeMender Coding Agent"] end

INGESTION --> D D --> E E -->|"Validated Risk"| F F --> G F --> H ```

[01:10:40](https://www.youtube.com/watch?v=11PBno-cJ1g&t=4240s) Live Demonstration: From Validated Risk to Verified Fix

Walking through a live platform demonstration, Costica illustrates how Wiz generates an agentless, dynamic inventory of all code and cloud assets, granting security teams complete visibility into corporate AI builds without causing operational friction. The Wiz security graph serves as a living architectural schematic of every live AI application. On screen, Costica highlights a cloud-based agent provisioned with tools to query databases and execute code; Wiz instantly discovers that the agent is directly exposed to the internet and holds active pathways into sensitive customer data, diagnosing the issue without requiring meetings or developer interviews.

Costica then triggers the Wiz Red Agent, which functions as an internal penetration tester to validate exposures before external adversaries find them. Simulating real-world tactics, the Red Agent confirms a critical authentication bypass vulnerability that would permit malicious actors to breach the agent and siphon data from the backend database. Instead of forcing teams into slow triage cycles marked by endless spreadsheets, the Wiz Green Agent takes over. It autonomously pinpoints the responsible asset owner, locates the precise faulty line of code, and generates a remediation patch. The green agent can dispatch a pull request straight to the engineering team or hand off the patch to automated coding agents like CodeMender, resolving vulnerabilities in source code in minutes.

> “And in the AI era, the best defense is to continuously use AI against ourselves in order to give the defenders the first-mover advantage over the attackers.” — Yinon Costica

- Agentless discovery maps all deployed models, IDEs, and tools into a dynamic inventory. - Wiz Red Agent simulates elite penetration testers to convert theoretical exposures into validated vulnerabilities. - Wiz Green Agent automates triage by mapping code ownership, drafting code-level repairs, and routing fixes to developers or automated remediation tools like CodeMender.

[01:14:23](https://www.youtube.com/watch?v=11PBno-cJ1g&t=4463s) Enterprise Validation and Unified Agentic Defense

Francis deSouza returns to the stage, emphasizing how uniting Google’s agentic security operations center (SOC) with Wiz’s deep cloud context empowers global enterprises to build and deploy generative AI with confidence. He lists concrete implementations worldwide: the Los Angeles Department of Water and Power relies on the integrated platform to protect vital municipal systems ahead of the 2028 Olympic Games, Singapore’s CSIT employs it for proactive national threat defense, and DBS Bank embeds Google Cloud security directly into its core financial services to enable real-time detection and response.

DeSouza points out that financial heavyweight Morgan Stanley has integrated Wiz as a cornerstone of its cloud security program, deploying Google Cloud Security capabilities to oversee expanding multi-cloud architectures. International enterprises including Nestlé, LVMH, BMW, and Shell similarly depend on this unified framework. DeSouza frames this partnership as closing the historical divide separating security analysts from application developers, establishing a comprehensive defense that is natively autonomous and secure by design.

> “Google and Wiz provide a unified agentic defense that is secure by design and autonomous by nature.” — Francis deSouza

- Los Angeles Department of Water and Power (LADWP) secures municipal infrastructure for the LA 2028 Olympic Games using the unified stack. - Singapore's CSIT relies on Google and Wiz for proactive digital threat defense. - DBS Bank incorporates Google Cloud protections directly into financial system architectures for immediate threat remediation. - Morgan Stanley, Nestlé, LVMH, BMW, and Shell trust Wiz and Google Cloud Security to ensure multi-cloud and AI-driven systems remain secure by design.

[01:16:20](https://www.youtube.com/watch?v=11PBno-cJ1g&t=4580s) Agentic Taskforce: Reimagining Customer Experience and Commerce

Welcomed to the stage by deSouza, Carrie Tharp explains that establishing an unshakeable security baseline unlocks the freedom to innovate rapidly. Google Cloud directs this foundation toward two enduring operational hurdles: improving customer engagement and augmenting employee productivity. To achieve this, organizations can deploy an "agentic taskforce"—fleets of specialized agents designed to work directly alongside human staff as strategic operational partners rather than static software utilities.

Tharp highlights Gemini Enterprise for Customer Experience, a dedicated solution designed to streamline the consumer lifecycle from initial search to long-term post-purchase care. The suite introduces agentic commerce, featuring pre-built shopping and food-ordering agents capable of handling exploratory queries, recommendations, and checkouts entirely via natural dialogue. At Papa John's, for example, a custom food-ordering agent retains unique customer dietary preferences to speed up delivery times. For complex retail support, Best Buy uses Agent Assist to guide customer service employees and autonomously walk shoppers through technical product specifications, issue resolution, and appointment booking. To keep these workflows cohesive, an omnichannel gateway preserves complete customer context across web, mobile, and phone channels without conversational drop-offs.

> “In the agentic era, an agent is no longer just a tool. It's a strategic extension of your business, built to expand your reach, deepen engagement, and personalized service at scale.” — Carrie Tharp

- Agentic commerce suite provides conversational agents that navigate discovery, product recommendations, and final checkout. - Papa John's deploys conversational food-ordering agents that track personal preferences to speed up orders. - Best Buy leverages Agent Assist to autonomously handle appointment schedules, complex technical queries, and customer issue resolution. - Omnichannel gateway maintains unbroken conversational context as customers switch across voice, mobile apps, and web chat.

[01:18:37](https://www.youtube.com/watch?v=11PBno-cJ1g&t=4717s) Gemini Enterprise for Customer Experience: The Home Depot and Reliance

Gemini Enterprise for Customer Experience delivers conversational AI agents capable of maintaining context across channels, seamlessly remembering where a user left off, and scaling globally with human-like voice synthesis across multiple languages. At The Home Depot, this technology serves as the foundation for the "Magic Apron" digital assistant, designed to translate the retailer's renowned in-store orange apron customer service into an end-to-end digital experience across inspiration, product research, purchasing, and ongoing project execution.

```mermaid flowchart LR accTitle: The Home Depot Customer Journey via Magic Apron accDescr: A linear flow of the customer journey supported across all touchpoints by the Magic Apron digital agent. A["Inspiration"] --> B["Research"] B --> C["Shopping & Buying"] C --> D["Doing & Support"] ```

Jordan Broggi and Angie Brown of The Home Depot emphasize that Gemini Enterprise provides the raw speed, reasoning power, and agentic back end required to unify their web and in-store touchpoints. In physical stores, shoppers rely on Magic Apron for indoor wayfinding, product specifications, and hardware compatibility checks; at home, they use it for project planning and post-purchase support. Crucially, Jordan Broggi notes that this unified experience delivers concrete business results: “Maybe most importantly, we see a higher conversion rate when they leverage the tool versus when they don't.”

Beyond home improvement, the agentic paradigm is transforming digital-first scale retail internationally. In India, Reliance has deployed a shopping agent powered by Gemini to help millions of consumers assemble complex orders effortlessly. A shopper can type a conceptual prompt such as "plan a birthday party," prompting the agent to curate items across disparate product categories, suggest complementary items, and build a unified shopping cart. This capability is anchored by Gemini-driven catalog enrichment, which ingests and classifies millions of product images in mere minutes rather than requiring months of manual indexing. Internally, Google is drinking its own champagne by deploying Gemini Enterprise for Customer Experience across Google Store, Google Photos, Google One, and YouTube.

- **The Home Depot Magic Apron:** Digital equivalent of the in-store orange apron staff, providing wayfinding, inventory lookups, and compatibility recommendations across physical stores and web platforms. - **Reliance Shopping Agent:** Drives increased average order value and basket sizes via cross-category bundling prompted by natural language life events (e.g., party planning). - **Automated Catalog Enrichment:** Employs multimodal Gemini models to index, tag, and enrich millions of retail product images in minutes instead of months. - **Internal Google Adoption:** Customer experience scaling currently active across Google Store, Google Photos, Google One, and YouTube TV.

---

[01:21:40](https://www.youtube.com/watch?v=11PBno-cJ1g&t=4900s) Live Demo: YouTube TV Support Voice Agent and CX Agent Studio

Senior Product Manager Patrick Marlow demonstrates a live production deployment of the YouTube TV voice support agent, currently handling 100% of incoming telephone traffic for NFL Sunday Ticket and YouTube TV plan subscriptions. Calling the live support number from the keynote stage, Marlow acts as a prospective subscriber who only watches sports and needs a solution for an upcoming NFL draft party with friends. The agent immediately identifies the specialized YouTube TV Sports Plan, notes that it includes over 30 sports channels, calculates that it costs $18 less per month than the standard base plan, and dispatches a direct sign-up link via SMS text message to his phone.

```mermaid flowchart TD accTitle: YouTube TV Voice Agent Architecture in CX Agent Studio accDescr: Multi-agent system orchestration showing supervisor and subagents handling caller requests. User["Inbound Caller"] --> Orch["Primary Orchestrator Agent"] subgraph MultiAgentSystem ["CX Agent Studio Orchestration"] Orch --> Sub1["Plan & Subscription Agent"] Orch --> Sub2["Price Finder Tool (Grounded KB)"] Orch --> Sub3["Multilingual Agent (Spanish)"] Orch --> Sub4["Promotions Agent (Subagent)"] end MultiAgentSystem --> Output["Telephony & SMS Delivery"] ```

To showcase real-time multimodal adaptability, Marlow asks the agent to explain the plan in Spanish to his visiting father-in-law, specifically confirming access to both American football and soccer (*fútbol*). The voice agent pivots instantly into fluent Spanish without losing session state or call context. When asked whether the household can stream across multiple television sets simultaneously (such as in the living room and out on the patio), the agent confirms concurrent streaming support on up to three screens at the same time. Marlow concludes the call with an aside that he is standing in a stadium surrounded by thousands of people with millions more watching the live stream.

Moving behind the scenes into CX Agent Studio, Marlow displays the Visual Builder interface that powers the YouTube TV production environment. Instead of requiring software engineering teams to modify endless lines of code, the system allows non-technical product owners to manage orchestration across specialized subagents using declarative natural language business logic and grounded knowledge bases. Marlow adds a brand-new "Promotions" subagent and configures its business logic in real time in seconds, showing how conversational insights monitor call performance directly in the same pane:

> “Now, let's recap. We just turned a routine support call into a multilingual upgrade and a globally deployed promotion, all in under four minutes.”

- **Full Production Deployment:** The YouTube TV voice agent handles 100% of customer support call volume for NFL Sunday Ticket and base tier subscriptions. - **Rapid Development Lifecycle:** YouTube TV customer service teams built, integrated, grounded, and launched the full multi-agent support agent in just six weeks. - **Dynamic Policy and Knowledge Integration:** Utilizes an integrated Price Finder grounding tool connected directly to authoritative internal rate sheets to guarantee factual, non-hallucinatory quoting. - **Low-Code Visual Orchestration:** CX Agent Studio enables instant addition of specialized subagents (demonstrated via a live promotional campaign agent) configured strictly through plain text instructions.

---

[01:26:30](https://www.youtube.com/watch?v=11PBno-cJ1g&t=5190s) Introducing Workspace Intelligence: Ending Context Fragmentation

Yulie Kwon Kim, Vice President of Product for Google Workspace, steps onto the stage to introduce Workspace Intelligence, an enterprise capability designed to eradicate the severe productivity loss caused by context switching. Modern office work is plagued by fragmented applications: an employee searches for a single data point and quickly finds themselves with 15 disconnected browser tabs open, jumping between outdated email threads, collaborative slide decks being edited live, and dense spreadsheets. As Kim observes, professional workers routinely spend half their working hours searching for disparate context and the remaining half deciding what action to take.

Workspace Intelligence integrates Gemini's frontier reasoning capabilities and advanced multimodal embedding models directly into the foundational layer of Google Docs, Sheets, Slides, Drive, Gmail, and Google Chat. Rather than treating productivity tools as isolated files, it functions as a unified ambient cognitive layer that automatically establishes relational links between dispersed enterprise artifacts.

```mermaid flowchart TD accTitle: Workspace Intelligence Unified Layer accDescr: Workspace Intelligence connecting background enterprise data into Google Chat as a command center. subgraph DataSources ["Workspace Context & Agents"] D1["Drive Pitch Decks"] D2["Gmail Threads"] D3["Sheets Merchandising Playbooks"] D4["Gemini Enterprise Ops Agents"] end DataSources --> WI["Workspace Intelligence Layer (Gemini + Embeddings)"] WI --> Chat["Google Chat Command Center (Ask Gemini)"] Chat --> Action["Surfaced Priority Actions & Semantically Grounded Assets"] ```

In a live walkthrough set inside the ongoing furniture rebrand narrative, Kim assumes the role of a regional distributor working inside Google Chat. Inside the Regional Managers chat space, an automated Regional Operations Agent (constructed in Gemini Enterprise) issues an proactive alert that new retail display kits are arriving ahead of schedule, sparking rapid cross-functional chatter across multiple rooms.

Instead of reading through backlogged chat histories, Kim activates "Ask Gemini" directly inside Google Chat, converting the messaging client into an executive command center. In a single pane, Gemini aggregates the critical items: it identifies an impending 4:00 PM regional plan deadline, extracts the specific presentation deck requiring geographical localization, and prepares the user to take immediate action without opening a single external tab. When she needs to locate historical quarterly sales figures buried in past documentation, she enters a semantic request: *"Find the merchandising playbook from last quarter, the one with the chart showing regional sales."* Rather than relying on simple keyword matching, the underlying vector embeddings interpret the conceptual intent of finding a specific chart embedded inside historical presentation assets.

- **Unified Intelligence Layer:** Runs natively across Gmail, Drive, Docs, Sheets, Slides, and Google Chat, powered by advanced Gemini reasoning and state-of-the-art embedding models. - **Context Unification:** Eliminates cross-app tab sprawl by automatically surfacing relevant source documents, presentation slide decks, and deadlines within the active communication surface. - **Embedded Agent Integration:** Enables autonomous business process agents created in Gemini Enterprise to interact directly with human teams inside Google Chat spaces. - **Semantic File Retrieval:** Replaces traditional lexical keyword file queries with conceptual understanding capable of retrieving specific sub-elements (like specific charts or tables) inside complex enterprise documents.

[01:29:45](https://www.youtube.com/watch?v=11PBno-cJ1g&t=5385s) Orchestrating Regional Plans with Workspace Intelligence

Continuing the live demonstration of Google Workspace, the presenter locates a critical graph buried deep within meeting contexts and file contents, describing the instant retrieval as “a needle in a haystack, solved in seconds.” Moving from data gathering to execution, the presenter sets out to convert this fragmented information into an executive regional plan for store managers. Rather than manually building slides, the presenter invokes Ask Gemini with a conversational command: tagging the specific “regional campaign skill” and prompting it to generate a deck centered on organic living.

```mermaid flowchart TD accTitle: Workspace Intelligence Regional Deck Generation accDescr: Diagram illustrating how Workspace Intelligence synthesizes multiple internal and external data sources into a Google Slides presentation.

subgraph Inputs["Context & External Sources"] A["Emails & Chat Threads"] B["Internal Drive Docs & Meeting Context"] C["Live HubSpot Win-Loss Data"] end

subgraph Engine["Workspace Intelligence"] D["Regional Campaign Skill"] E["Corporate Brand & Template Engine"] end

subgraph Output["Artifact Generation"] F["Formatted Google Slides Deck with Citations"] end

A --> D B --> D C --> D D --> E E --> F ```

Workspace Intelligence orchestrates the generation by reading across emails, chat messages, and background documents while pulling live win-loss metrics directly from HubSpot. The system formats the content using standard corporate branding templates to assemble a complete Google Slides presentation. The generated deck includes transparent citations listing every source file, chat, and email used to compile the briefing. Remarking on the quality and structure of the final deck, the presenter observes:

> “And it looks like it was designed by a creative pro, not a terminal.”

- **Context retrieval**: Resolves cross-platform searches across meeting transcripts and file contents instantly. - **Skill-based actions**: Users tag predefined organizational skills (e.g., regional campaign skill) inside Ask Gemini to trigger cross-application workflows. - **Integrations and citations**: Pulls third-party CRM records (HubSpot) alongside Google Workspace interactions with source transparency.

---

[01:31:26](https://www.youtube.com/watch?v=11PBno-cJ1g&t=5486s) Eliminating the Context Tax and Accelerating Workspace Migration

The presenter frames Workspace Intelligence as a structural solution to workplace fragmentation, declaring: “Workspace Intelligence is the end of the context tax.” Instead of forcing workers to juggle disparate tools and manually transfer information, the platform orchestrates enterprise workflows natively with built-in security and integration. Global enterprise customers are already leveraging these capabilities to scale productivity:

- **Colgate-Palmolive**: Deployed Google Workspace to 34,000 employees; teams use Gemini agents to transform raw data into new consumer product concepts within minutes rather than months, mining decades of historical sales data for actionable growth insights. - **Natura**: Custom agents accelerate data-driven reporting cycles by 10x. - **Korean Air**: More than 22,000 employees deploy AI agents to streamline operational workflows and customer care tasks. - **Compass Real Estate**: Gemini automates routine administrative workloads so agents can prioritize client relationships.

Addressing organizations hesitant to switch cloud office suites due to migration friction, Google announced major updates to its migration tooling and interoperability framework. Organizations, including specialized finance and compliance teams, can now migrate from Microsoft 365 to Google Workspace up to five times faster. With automated context retrieval handling background coordination, employees are urged to “close those 15 tabs” and let Workspace Intelligence take over the administrative burden.

---

[01:33:43](https://www.youtube.com/watch?v=11PBno-cJ1g&t=5623s) Customer Spotlight: Unilever's Multi-Agent Sourcing Engine

Thomas Kurian returns to the stage to highlight global consumer goods giant Unilever, an organization serving 3.7 billion people daily across the world. Leandro Barreto, Chief Digital Officer at Unilever, explains that partnering with Google Cloud enables the company to deliver personalized experiences at scale, describing personalization as the core of Unilever's digital strategy. Sam Kini, Chief Information Officer, emphasizes the construction of an enterprise-wide digital foundation powered by Google Cloud to accelerate AI deployment across demand generation and retail collaborations.

Daphne Coates, Global Procurement Digital & Tech Director, details how Unilever operationalizes multi-agent systems via Horizon3 Labs. Built on Gemini Enterprise, the Gemini model family, and Google Cloud's Agent Development Kit (ADK), Unilever developed a specialized competitive buying multi-agent solution. The system orchestrates diverse sourcing and analysis agents beneath a single unified user interface, giving procurement professionals real-time market intelligence. Where complex procurement analyses previously required days of manual collation, the multi-agent system completes assessments in minutes, enabling faster, data-driven supplier decisions. Reflecting on the impact across their supply chain, Kini notes:

> “With Google Cloud, we are turning intelligence into a growth engine.”

---

[01:35:53](https://www.youtube.com/watch?v=11PBno-cJ1g&t=5753s) Open AI Ecosystem, Partner Expansions, and Closing Remarks

Thomas Kurian returns to outline Google Cloud's philosophical approach to enterprise AI, contrasting Google's strategy with proprietary platforms. Kurian insists that the future of enterprise AI must be open rather than closed inside proprietary silos:

> “While others want to lock you into a wall garden that owns your models, your data, and your agents, we offer you an integrated stack, but the freedom to choose the world's best chips and models, the freedom to run AI wherever your data may live, the freedom to control your own destiny with deep governance features.”

This open mission is expanded through major systems integrators and specialized AI practices. Global consultancies—including Accenture, BCG, Deloitte, and McKinsey—have significantly expanded their dedicated Google Gemini practices, fielding forward-deployed engineering teams to implement agentic systems. Additionally, AI-native service partners like Quantium, Distyl, and Tribe.AI work alongside independent software vendors (ISVs) and SaaS platforms to rebuild enterprise solutions on top of the Gemini Enterprise Agent Platform.

```mermaid flowchart LR accTitle: Google Cloud Next '26 Agentic Blueprint accDescr: High-level overview of the five architectural pillars comprising Google Cloud's Blueprint for the Agentic Enterprise.

subgraph Infrastructure & Foundation A["AI Hypercomputer<br/>(Purpose-Built Compute)"] B["Agentic Data Cloud<br/>(Trusted Context Engine)"] end

subgraph Security & Management C["Agentic Defense<br/>(Autonomous Protection)"] D["Gemini Enterprise Agent Platform<br/>(Build, Deploy & Manage)"] end

subgraph Application Layer E["Agentic Taskforce<br/>(Gemini Enterprise App)"] end

A --> D B --> D C --> D D --> E ```

Kurian closes the keynote by summarizing the five architectural pillars of the Agentic Enterprise Blueprint introduced throughout the morning:

1. **AI Hypercomputer**: An infrastructure foundation engineered specifically for the computational physics and memory demands of agentic workloads. 2. **Agentic Data Cloud**: The grounding engine supplying autonomous agents with governed, trusted enterprise context across distributed sources. 3. **Agentic Defense**: Autonomous security mechanisms protecting the complete lifecycle of models, data, and active agent workflows. 4. **Gemini Enterprise Agent Platform**: A comprehensive platform featuring state-of-the-art Gemini models to build, deploy, evaluate, and orchestrate agent networks. 5. **Agentic Taskforce**: Specialized operational agents accessible to end users via the Gemini Enterprise application, transforming everyday customer experiences and organizational productivity.

Kurian thanks Google's engineers, partners, and customers for preparing the platform for production scale, closing with a call to action for attendees as they begin three days of technical sessions: “This platform is ready. So what will each of you build?”

## Notable quotes > [23:58](https://www.youtube.com/watch?v=11PBno-cJ1g&t=1438s) The era of the pilot is over. The era of the agent is here. — Thomas Kurian

> [02:09](https://www.youtube.com/watch?v=11PBno-cJ1g&t=129s) You cannot deliver AI by piecing together a puzzle piece of fragmented silicon and disconnected models. — Thomas Kurian

> [04:32](https://www.youtube.com/watch?v=11PBno-cJ1g&t=272s) The conversation has gone from can we build an agent to how do we manage thousands of them? — Sundar Pichai

> [49:00](https://www.youtube.com/watch?v=11PBno-cJ1g&t=2940s) Trusted context turns an intelligent guess to a decisive action. — Karthik Narain

> [01:04:40](https://www.youtube.com/watch?v=11PBno-cJ1g&t=3880s) In an agentic enterprise, giving agents the power to act means that security must become an autonomous force, moving faster than the threats themselves. — Francis deSouza

## Mentioned Accenture, Agent Assist, Agent Development Kit (ADK), Agentic Data Cloud, Agentic Defense, Agentic Taskforce, AI Application Protection Platform (AI APP), AI Hypercomputer, American Express, American Society of Clinical Oncology (ASCO), Amin Vahdat, Anthropic, Apache Iceberg, Apple, Apple Intelligence, Artemis II, Atlassian, AWS S3, Axia Energia, Bayer Crop Science, BCG, Best Buy, BigQuery, Billy Bohan Chinique, BMW, Bosch, Boston Dynamics, Box, Carrie Tharp, Citadel Securities, Citi Sky, Citi Wealth, Claude Haiku, Claude Opus 4.7, Claude Sonnet, Cloud Code, CME Group, CodeMender, Colgate-Palmolive, Compass Real Estate, Costco, CSIT, CX Agent Studio, Data Agent Kit, Databricks, DBS Bank, Deloitte, Distyl, Erika Chuong, Flipkart, Francis deSouza, Gemini, Gemini 3.1 Flash Image, Gemini 3.1 Pro, Gemini CLI, Gemini Enterprise, Gemini Enterprise Agent Platform, Gemini Enterprise Deep Research Agent, Gemini Enterprise for Customer Experience, Gmail, Google, Google Axion, Google Chat, Google Cloud, Google DeepMind, Google Distributed Cloud Edge, Google Docs, Google Drive, Google One, Google Pathways, Google Photos, Google Pixel Fold, Google SecOps, Google Sheets, Google Slides, Google Store, Google Workspace, Honeywell, Horizon3 Labs, HubSpot, Jason Davenport, JAX, JetBrains, Jordan Broggi, Karthik Narain, Katherine Litinsky, KPMG, Korean Air, LADWP, Lightning Engine for Apache Spark, Liverpool, Lovable, Lowe's, LVMH, Lyria 3 Pro, Macquarie Bank, Magic Apron, Managed Lustre, Mandiant, McKinsey, Meesho, Merck, Microsoft 365, Microsoft Azure, Model Context Protocol (MCP), Morgan Stanley, Nano Banana 2, NASA, Natura, Nestlé, NFL Sunday Ticket, Nirmal Saverimuttu, NVIDIA Vera Rubin NVL72, OpenTelemetry (OTel), Oracle, Palantir, Papa John's, Patrick Marlow, Quantium, Reliance, Replit, Rovey, Salesforce, Salesforce Agentforce, Sam Kini, SAP, Security Command Center, ServiceNow, Shaun White, Shell, Signal Iduna, Siri, Spanner, Sundar Pichai, Team USA, The Home Depot, Thinking Machine Labs, Thomas Kurian, Tinker, TPU 8i, TPU 8t, TPU Ironwood, TPUDirect, Tribe.AI, Unilever, US Department of Energy, Veo 3.0, Veo 3.1 Lite, Virgin Media O2, Virgin Voyages, Virgo Network Fabric, VirusTotal, Vodafone, VS Code, Walmart, Wiz, Workday, Workspace Intelligence, Woven by Toyota, Yasmeen Ahmad, Yinon Costica, YouTube, YouTube TV

## Video description The Next '26 Opening Keynote introduces the blueprint for the agentic enterprise, guiding businesses through the shift from AI adoption to large-scale transformation. Join Google Cloud CEO Thomas Kurian for major product announcements and hear from global customers who are partnering with Google Cloud to drive real-world impact.

To watch the Opening Keynote with ASL https://youtube.com/live/Pm7yGm6Fgl4

Ready to enter the agentic era? Try the Gemini Enterprise app today. https://business.gemini.google

AI agent-ready? Find out with our expert diagnostic. Book now. https://cloud.google.com/events/next26-offer-2

Watch more: 100+ sessions from Google Cloud Next 26 → https://www.googlecloudevents.com/next-vegas/ Subscribe to Google Cloud → https://goo.gle/GoogleCloud

Chapters 00:00 Countdown 00:12 Agentic Enterprise Blueprint 12:02 Gemini Enterprise Agent Platform 39:58 AI Hypercomputer 50:01 Agentic Data Cloud 1:04:40 Agentic Defense 1:16:20. Agentic Taskforce 1:33:52 Closing Remarks

[Source](https://www.youtube.com/watch?v=11PBno-cJ1g)

Reading summary

At Google Cloud Next '26, Google leadership argued that enterprise artificial intelligence has transitioned from isolated experimentation into large-scale production, driven by massive infrastructure spending, specialized silicon, and multi-agent systems. Infrastructure Surge and Split Eighth-Generation Silicon (4:32) Alphabet CEO Sundar Pichai revealed that Google plans to invest between $175 billion and $185 billion in capital expenditures in 2026—a nearly sixfold increase from $31 billion in 2022—with over half of its 2026 machine learning compute allocated directly to Google Cloud customers. Pichai noted that Google acts as 'Customer Zero' for its stack, generating nearly 75% of its internal new code using AI with human engineer reviews, up from 50% the previous fall. To address diverging computational demands, Google Cloud Chief Technologist Amin Vahdat announced the split of its eighth-generation Tensor Processing Units into two dedicated hardware lines: TPU 8t for model training and TPU 8i for inference and reinforcement learning. TPU 8t features native block-scale quantization in its Matrix Multiply Units to deliver 1.21 exaflops of FP4 compute per 9,600-chip pod, while TPU 8i delivers 11.6 exaflops of FP8 compute across 1,152 chips with an on-silicon memory cache. These platforms are complemented by Google Axion Arm CPUs, upcoming NVIDIA Vera Rubin NVL72 systems, and the Virgo Network fabric connecting up to 134,000 chips. Gemini Enterprise Platform and Multi-Agent Orchestration (8:48) Google Cloud introduced the Gemini Enterprise Agent Platform to tackle the operational bottleneck of securing and coordinating fleets of autonomous agents. The environment pairs natural-language low-code creation in Agent Studio with governance through the Agent Registry and Skills and Tools Registry, allowing models to interoperate across ecosystems using the Model Context Protocol. Workflows support both generative paths and deterministic orchestration for audit-heavy corporate environments. At the frontier model layer, Google unveiled Gemini 3.1 Pro for multi-step reasoning, alongside Gemini 3.1 Flash Image, Veo 3.1 Lite, and support for Anthropic's Claude Opus 4.7. Commercial deployments highlight widespread operational adoption: Apple selected Google Cloud to collaborate on foundation models for upcoming Apple Intelligence features, Citi Wealth launched its Citi Sky assistant, and Unilever operationalized a multi-agent procurement engine to complete complex sourcing reviews in minutes rather than days. Context Grounding and the Cross-Cloud Lakehouse (49:00) Karthik Narain outlined the Agentic Data Cloud, arguing that reasoning models require verified business context to avoid ungrounded guesses. The platform's Knowledge Catalog extracts context across relational databases and unstructured files like PDFs and images, while the Data Agent Kit enables developers to instruct agents to build data pipelines and deploy models directly from IDEs such as VS Code and the Gemini CLI. To resolve data fragmentation across cloud providers without replication or egress costs, Google Cloud introduced the Cross-Cloud Lakehouse. Built on open Apache Iceberg formats, the lakehouse allows agents and BigQuery or Spark analytics engines to run zero-copy, low-latency queries directly across AWS S3 and Microsoft Azure storage, removing traditional vendor lock-in. Autonomous Agentic Defense and Wiz Integration (64:40) Cybersecurity lead Francis deSouza warned that modern threats operate at unprecedented velocities, with mean time to exploit dropping to negative seven days and attacker lateral handoffs plunging to 22 seconds. Google Cloud deployed an agentic Security Operations Center utilizing Gemini-driven triage agents alongside Mandiant, VirusTotal, and Chrome telemetry to remediate threats autonomously at machine speed. Wiz co-founder Yinon Costica presented Wiz's integration into Google Cloud, introducing its AI Application Protection Platform and autonomous Red, Blue, and Green agents. The system builds an agentless security graph of multi-cloud architectures; during a live demonstration, the Wiz Red Agent validated an authentication bypass on an exposed agent, after which the Green Agent located the faulty code and generated an automated patch to resolve the vulnerability without manual triage. Agentic Task Forces in Operations and Workspace (76:20) Google Cloud demonstrated task-specific agents across consumer-facing operations and office productivity. In customer service, Gemini Enterprise powers conversational voice and shopping agents, including YouTube TV's live telephony support agent that handles inbound calls and dynamically switches between languages, as well as The Home Depot's 'Magic Apron' digital assistant for in-store navigation and project planning. Within productivity environments, Google Workspace Vice President Yulie Kwon Kim introduced Workspace Intelligence to eliminate the 'context tax' of toggling across fragmented applications. Operating across Docs, Gmail, Sheets, and Google Chat, the ambient layer synthesizes dispersed corporate documents, emails, and CRM metrics into formatted presentations and executive summaries with source citations, supported by migration tools that accelerate transitions from Microsoft 365 by up to 5x. Google Cloud framed its integrated stack—spanning custom silicon, open data formats, autonomous defense, and agent governance—as an open alternative to proprietary corporate walled gardens, though maintaining performance across multi-cloud environments and managing autonomous code generation remain complex long-term engineering challenges.

Source quotes

[23:58] Thomas Kurian: The era of the pilot is over. The era of the agent is here. [2:09] Thomas Kurian: You cannot deliver AI by piecing together a puzzle piece of fragmented silicon and disconnected models. [4:32] Sundar Pichai: The conversation has gone from can we build an agent to how do we manage thousands of them? [49:00] Karthik Narain: Trusted context turns an intelligent guess to a decisive action. [64:40] Francis deSouza: In an agentic enterprise, giving agents the power to act means that security must become an autonomous force, moving faster than the threats themselves.

Original transcript

[0:24] [MUSIC PLAYING] [2:04] [CHEERING] [2:09] SPEAKER 1: Please welcome to the stage CEO [2:12] of Google Cloud, Thomas Kurian. [2:16] [UPBEAT MUSIC] [2:34] THOMAS KURIAN: Hello. [2:39] [CHEERING, APPLAUSE] [2:43] THOMAS KURIAN: Welcome everyone to Google Cloud Next. [2:46] Just one year ago, we stood on this same stage [2:50] and promised a new future for AI. [2:53] Today, that future is running in production at a scale [2:58] that the world has never seen. [3:00] Over the last year, we didn't just see adoption. [3:05] We saw transformation. [3:07] Nearly 75% of Google Cloud customers [3:11] now leverage our AI products to power their businesses. [3:15] We have thousands of agents and services across every industry, [3:20] reaching billions of people through the global scale [3:23] of our partner network. [3:25] You have moved beyond the pilot. [3:28] The experimenting phase is behind us. [3:31] And now, the real challenge begins. [3:34] How do you move AI into production [3:37] across your entire enterprise? [3:40] The answer is a unified stack. [3:43] You cannot deliver AI by piecing together a puzzle piece [3:47] of fragmented silicon and disconnected models. [3:51] To drive real value, you need an architecture [3:55] where chips are designed for the models, [3:58] models are grounded in your data, agents and applications [4:03] are built with models and secured by the infrastructure. [4:09] Google uses this exact same OpenStack [4:13] to reimagine how we serve users using our AI [4:17] tools across Search, YouTube, Chrome, and Android. [4:23] To take you inside that journey, let's hear from Sundar Pichai. [4:28] [APPLAUSE] [4:32] SUNDAR PICHAI: Thanks, Thomas, and hello, everyone. [4:34] We are so glad you are here. [4:37] The pace of technological change has [4:39] been faster than I've ever seen, and I've been around technology [4:43] for a while. [4:44] In fact, this Sunday, I'm celebrating my Googleversary-- [4:49] 22 years, and I'm still feeling lucky. [4:52] And I hope that luck extends to all of you in Vegas this week. [4:56] I was drawn to Google all those years [4:57] ago because of its ambitious mission. [5:00] From the web to mobile to AI, every platform shift [5:05] has given us new opportunities to advance our mission [5:09] and to help our customers and partners achieve yours. [5:13] As we move into the agentic era, we [5:16] are taking this to the next level. [5:18] We are making big investments now and for the future. [5:22] In 2022, we were investing $31 billion in CapEx. [5:27] This year, we plan to invest between $175 to $185 billion [5:33] in total CapEx, a nearly 6x increase in just four years. [5:38] And for 2026, just over half of our machine learning compute [5:41] is expected to go towards the cloud business, [5:44] so it will greatly benefit all of you. [5:47] These investments are how we continue to stay at the frontier [5:51] and ensure you're at the cutting edge as well. [5:55] A big focus for us is to always be [5:57] customer 0 for our own technologies, [6:00] so we can be a better partner to all of you. [6:03] Let me share a few examples of how we [6:05] are rewiring our work with AI. [6:08] First, coding-- we've been using AI to generate code internally [6:12] at Google for a while. [6:14] Today, nearly 75% of all new code at Google [6:18] is AI-generated and approved by engineers, [6:21] up from 50% last fall. [6:23] We are now shifting to truly agentic workflows. [6:27] Our engineers are orchestrating fully autonomous digital task [6:31] forces, firing off agents, and accomplishing incredible things. [6:37] As one example of how we are using agents, [6:40] we recently executed a particularly complex code [6:43] migration. [6:45] Anyone familiar knows these migrations [6:47] can take a while to get done. [6:50] We created a system of agents taking on three different types [6:53] of roles-- [6:54] planners, orchestrators, and coders. [6:58] Working together with engineers, we completed the migration 6x [7:01] faster than we could have a year ago. [7:04] And now we are applying that approach [7:06] across our entire development cycle, [7:09] from experimentation to testing to evals. [7:13] It's not just developers who are infusing AI [7:15] into their workflows. [7:17] Our marketing teams are fundamentally [7:19] rethinking the creative process from concept to launch. [7:23] Historically, adapting a campaign [7:25] for every audience, channel, and country took weeks. [7:28] For the recent launch of Gemini and Chrome, [7:31] teams used our models to rapidly generate thousands of variations [7:35] of our creative assets. [7:37] This enabled personalization at a massive scale. [7:41] It also led to 70% faster turnaround and a 20% increase [7:45] in conversion. [7:47] We are not only getting to market faster. [7:49] We are also doing it more effectively with AI. [7:52] Our security teams are also seeing significant gains [7:55] with AI. [7:57] Each month, our teams receive unstructured threat reports [8:00] at a scale that would take thousands of hours to review-- [8:04] a nearly impossible task. [8:07] Today, our Security Operations Center agents automatically [8:11] triage tens of thousands of unstructured threat reports [8:14] each month. [8:15] By accelerating the extraction of critical intelligence [8:18] and filtering out the noise, it's [8:20] reduced threat mitigation time by over 90% [8:24] We are more on the front foot than ever before. [8:28] Those are just a few examples of what we are trying at Google. [8:31] We are moving in a bold and responsible way. [8:35] There's a lot of change happening. [8:37] And it can feel like we are in the messy part of the innovation [8:40] cycle. [8:40] But we are starting to see the foundational building [8:43] blocks coming together, unlocking [8:45] a new wave of innovation. [8:48] One thing that is super clear-- [8:50] we are firmly in the agentic Gemini era. [8:53] Last fall, we introduced Gemini Enterprise as a front door [8:57] to agentic AI in the workplace. [9:00] What we have seen in just a few months since launch [9:03] is how every employee in every organization [9:06] can become a builder. [9:08] This is an incredible shift. [9:10] We are accomplishing bigger things faster, [9:13] but this comes with complexity. [9:15] The conversation has gone from can we build an agent to how [9:19] do we manage thousands of them? [9:22] Today, I'm excited to announce our new Gemini Enterprise Agent [9:26] Platform. [9:28] It provides the secure, full-stack connective tissue [9:32] you need to build, scale, govern, and optimize [9:35] your agents with confidence. [9:38] Think of it as mission control for the agentic enterprise [9:42] to help move your organization into the next phase [9:46] of the agentic era. [9:48] I'm going to turn it back over to Thomas to share more. [9:51] But first, thanks again for being here. [9:54] Together, we are building a blueprint for true business [9:57] transformation, and I'm excited to see where it takes us next. [10:01] Thomas, back to you. [10:05] [UPBEAT MUSIC] [10:20] THOMAS KURIAN: Thank you, Sundar. [10:22] Today, we're taking the next step [10:24] in bringing Google AI to every employee and every workflow. [10:29] Gemini Enterprise is now the end-to-end system [10:32] for the agentic era, the connective tissue [10:35] between your data, your people, and your goals. [10:38] It transforms disconnected processes [10:40] into a single intelligent flow. [10:43] This is our blueprint for the agentic enterprise. [10:47] It's also the answer to that fundamental question. [10:51] Intelligence plus automation must deliver value. [10:56] To make this work, you need context and action. [11:01] Intelligence comes from your data. [11:04] Automation is driven by agents. [11:06] To solve this equation at scale, you [11:09] need a complete integrated system. [11:13] Today, we will show you the layers of that system and all [11:17] the innovations we're introducing at every layer, [11:20] starting with the AI Hypercomputer, [11:23] the purpose-built foundation optimized [11:26] for the physics of the agentic era; [11:29] the Agentic Data Cloud, the engine that provides agents [11:34] with trusted business context; Agentic Defense, the autonomous [11:40] protection that secures your entire AI lifecycle; [11:45] the Agentic Platform, the system to build, deploy, and manage [11:49] agents; and finally, the Agentic Task [11:53] Force, the pre-built specialized agents that we [11:57] offer that are ready to transform your business. [12:02] Let's dive in. [12:03] Now, the Gemini Enterprise Agent platform [12:06] is the environment where your business logic, your data, [12:10] and your models converge to drive autonomous action. [12:14] It expands on the previous capabilities of Vertex AI [12:17] and brings new capabilities to enable your teams [12:22] to build, scale, govern, and optimize agents [12:26] with the same architectural rigor you apply to your most [12:30] mission-critical systems. [12:32] At its core, we build on our support for state-of-the-art [12:36] models. [12:37] Our most advanced reasoning model, Gemini 3.1 Pro, [12:41] is available in preview. [12:42] It's optimized for complex workflow orchestration. [12:46] It bridges the gap between strategy and autonomous [12:48] execution, interacting with your APIs and systems [12:52] with minimal tuning. [12:54] Industry leaders, including Databricks, JetBrains, [12:57] and Replit, have chosen Gemini 3.1 Pro. [13:01] We're also announcing Gemini 3.1 Flash Image, [13:06] also known as Nano Banana 2, for high-fidelity visual assets; [13:11] Veo 3.1 Lite, our most cost-effective video model, [13:16] designed to build high-volume video applications; [13:20] and Lyria 3 Pro, a state-of-the-art model [13:24] for enterprise and professional-grade audio [13:28] and music. [13:30] All of these models are available in preview. [13:33] Finally, we support all the leading models from Anthropic, [13:37] including Claude Opus, Sonnet, and Haiku. [13:41] And today, we're adding support for Anthropic's Claude Opus 4.7. [13:47] While all of these models represent a massive leap [13:51] in intelligence, their true value [13:53] is realized when they're operationalized to solve [13:56] mission-critical problems. [13:59] Earlier this year, we announced a monumental partnership [14:03] with one of the world's most iconic brands, [14:07] that will bring the power of our technology [14:09] to users everywhere around the world. [14:12] We're collaborating with Apple as their preferred cloud [14:15] provider to develop the next generation of Apple foundation [14:19] models based on Gemini technology. [14:22] These models will help power future Apple Intelligence [14:26] features, including a more personalized Siri coming [14:31] later this year. [14:33] Leading organizations aren't just [14:35] using Gemini Enterprise to work faster. [14:38] They're using it to redefine what their businesses can do. [14:43] Citi Wealth, in partnership with Google Cloud and DeepMind, [14:47] today unveiled Citi Sky, an always-on AI-powered member [14:53] of the Citi Wealth team that brings [14:56] Citi's global intelligence to clients' fingertips. [15:03] Citi Sky will provide an exemplary experience, [15:06] allowing clients and colleagues to ask [15:09] more of Citi Wealth whenever they need it, [15:12] and in multiple languages. [15:15] Honeywell generates billions of insights for managing buildings [15:20] by training digital twins on over a million product [15:23] specifications. [15:25] Liverpool is bringing their signature in-store [15:29] service online, projecting a 10 times return on investment [15:34] on their new shopping assistant. [15:36] And for a mission that's quite literally out of this world, [15:42] we were very proud to partner with NASA-- [15:45] [CHEERING, APPLAUSE] [15:53] --to use Gemini Enterprise Agents to power flight readiness [15:58] and ensure astronaut safety for Artemis II, which [16:03] set the human spaceflight record for traveling [16:06] the furthest distance from Earth. [16:09] Now, we built the agent platform to manage the entire lifecycle [16:13] of an agent. [16:15] The low-code Agent Studio enables every employee [16:18] to build and deploy agents using natural language. [16:22] It grounds LLM reasoning in your specific business rules, [16:27] bringing predictable, autonomous action [16:30] to every workflow at scale. [16:33] To manage these assets, the agent registry [16:37] provides a single point of control, [16:39] indexing every internal agent and tool [16:42] across your organization to ensure they're [16:44] discoverable and governed. [16:47] Similarly, our Skills and Tools Registry [16:51] allows you to define modular, reusable packages [16:55] of instructions, scripts, and resources [16:58] to teach agents to perform specialized tasks or repeatable [17:02] workflows. [17:04] We also expose skills for every GCP service [17:09] and every service in Workspace in our skills registry. [17:14] All of this is also supported by our Agent Marketplace, [17:18] allowing you to search and deploy [17:21] specialized agents from our global partner ecosystem [17:25] directly in Gemini Enterprise, including Atlassian, Box, [17:31] Lovable, Oracle, ServiceNow, Workday, and many, many more. [17:38] Finally, with our native integration [17:41] of Model Context Protocol, or MCP, [17:44] you can connect our agent platform to any MCP server. [17:49] We're also exposing all of our GCP services [17:54] as MCP to allow you to interact seamlessly [17:57] with GCP from any agent. [18:01] In addition to developing this agent platform, [18:04] it also provides the framework to orchestrate and scale [18:08] your entire agentic workforce. [18:11] We enable agent-to-agent orchestration, [18:14] allowing agents to seamlessly delegate tasks from one [18:18] to another, including support for complex generative [18:24] and deterministic orchestration patterns. [18:27] This helps ensure that for your critical workflows, [18:31] such as those that need compliance, [18:34] your agents can follow well-specified paths [18:38] every single time, guaranteeing predictable outcomes. [18:44] Orchestration flows and agents can respond to events. [18:47] These events can be real-time. [18:49] They can be schedule-based. [18:51] They can be trigger-based, or they can be batch inference job. [18:56] We're bringing zero-trust verification [18:59] to every agent and every orchestration step. [19:04] And now with Agent Identity, every agent [19:09] has a unique cryptographic ID and well-defined authorization [19:15] policies that are traceable and auditable, [19:19] ensuring you can track every action in your company [19:24] and manage all your agents. [19:27] They're also centrally managed through our Agent Gateway, which [19:32] provides a single command center for policy enforcement [19:36] across the organization. [19:38] Paired with Model Armor, you protect your models. [19:42] You protect your proprietary enterprise data [19:45] from threats such as sensitive data leakage. [19:49] This integrated approach, from secure sandboxes [19:54] to a single management console, provides the visibility [19:57] and isolation you need to run your most-sensitive workloads [20:03] with a high degree of confidence. [20:06] Now, optimization and observability [20:08] are also built into the fabric of the platform. [20:12] Agent observability delivers granular instrumentation, [20:16] allowing you to visualize the full execution [20:19] path of any agent using OTel-compliant telemetry. [20:25] You can retrieve traces, monitor tool use, [20:28] quickly diagnose reasoning loops with fine-grained logging. [20:33] All of these capabilities that we've just detailed-- [20:37] the orchestration, zero-trust security, and developer tools-- [20:42] form the core engine of Gemini Enterprise. [20:46] While the agent platform is where your technical teams build [20:49] and govern agent, the Gemini Enterprise application [20:53] is the primary environment where your business actually operates. [20:58] It's that new front door to AI for all your employees, [21:02] turning complex agentic capabilities [21:04] into a simple, new way to work for every employee. [21:10] By unifying data across your entire stack, [21:13] including Google Workspace, coding tools, [21:17] and enterprise systems, we've created [21:20] a single intelligent flow for the entire organization. [21:24] To give your teams a permanent memory, [21:28] we're introducing Gemini Enterprise projects. [21:32] Projects give your agents a high-fidelity workspace [21:36] and the ability to use deep think [21:39] to solve your most complex business challenges [21:42] without context pollution. [21:45] We're also adding Microsoft 365 interoperability, [21:48] allowing you to export the Docs and Slides you've [21:51] created within Canvas into common Microsoft Office formats. [21:56] [APPLAUSE] [22:02] Leading innovators already use this platform [22:05] to redefine their industries. [22:07] Let's hear from one of them, Virgin Voyages. [22:11] [STIRRING MUSIC] [22:15] NIRMAL SAVERIMUTTU: Our people are a secret sauce. [22:17] So anytime we can use data to make our people more empowered, [22:21] we're all in. [22:23] Gemini Enterprise is a front door for AI in our business, [22:26] and we're using it to bring natural language intelligence [22:29] to our crew. [22:30] BILLY BOHAN CHINIQUE: Gemini Enterprise is helping us bring [22:32] Project 3D to life. [22:33] It's going to allow our crew to feel confident [22:36] with any question they're asked. [22:38] So for sailors, Rovey is a personal concierge [22:41] to maximize the value they get out of their vacation, [22:44] inspired at home all the way through the last day [22:46] of their voyage to make sure they have the best [22:48] time on vacation with us. [22:50] Our crew are already world-renowned and award-winning [22:53] for the exceptional service they deliver as humans. [22:55] We want Project 3D and Gemini Enterprise to be able to take [22:58] that even further. [23:00] So you can imagine when you're building [23:02] something for a cruise ship, there [23:03] are challenges that pop up. [23:05] There's not exactly a fiber cable running out [23:07] the back of the ship. [23:08] With Google Distributed Cloud Edge, [23:10] we had data resiliency at the core of the project. [23:13] And that means, even when we're offline, [23:15] our intelligence is still online. [23:17] Using our AI stack in Google Cloud has helped us [23:21] not only reduce our production timeline by up to 60%, [23:24] but we have contributed to a 28% increase month [23:27] on month for a record sales quarter. [23:31] We could truly see a day where you get on a Virgin Atlantic [23:34] flight, stay at a Virgin hotel, and end up [23:37] on a Virgin Voyage, all connected seamlessly by Project [23:40] Ruby and Gemini Enterprise. [23:41] NIRMAL SAVERIMUTTU: Google has been [23:43] a foundational partner for us. [23:44] The extension of Gemini. [23:45] Enterprise was the natural way to bring that mission to life. [23:53] [APPLAUSE] [23:58] THOMAS KURIAN: The era of the pilot is over. [24:00] The era of the agent is here. [24:03] But the true power comes from how it changes your workflow. [24:07] Let's see it in action. [24:09] Please welcome Erika Chuong. [24:11] [UPBEAT MUSIC] [24:21] ERIKA CHUONG: Thanks, Thomas. [24:23] Gemini Enterprise is a platform to build, manage, [24:27] and interact with agents. [24:28] I'll demonstrate how it can orchestrate multiple agents [24:31] across platforms, share context for seamless handoffs, [24:34] and streamline workflows. [24:36] Imagine I work for a global furniture retailer. [24:39] Here's my personalized homepage where [24:41] I can interact with internal business [24:43] context and external sources in a single pane of glass. [24:49] The agent gallery hosts my approved selection of agents, [24:52] including built-in ones by Google and my company's agents, [24:55] like this one for price and margin optimization, [24:58] which autonomously orchestrates across agents, tools, [25:02] and sources. [25:03] Now, let's see Gemini Enterprise in action. [25:07] To bring some less popular product lines back to life, [25:11] we'll ask our agent to analyze current interior design trends, [25:14] identify dead stock in our warehouse, [25:17] and orchestrate a relaunch campaign. [25:20] With that one prompt, multiple agents [25:22] complete a series of actions in minutes instead of hours. [25:26] My market research agent, powered by Deep Research, [25:29] analyzes the latest Google Search information [25:32] alongside my own sales and CSAT data. [25:36] My data insights agent is connecting to our global product [25:39] catalog, regardless of location or format. [25:43] It knows that dead stock refers to stale inventory, [25:46] and it uses our Agentic Data Cloud [25:48] to identify the right data sets. [25:51] And my product strategy agent pulls everything together. [25:55] I'll go ahead and approve this plan. [25:57] This can take a few minutes, so we'll fast-forward. [26:00] Here are the highlights. [26:03] Organic Modern is a huge trend, and customers [26:06] are paying top dollar. [26:08] Our Tuscany collection has leftover inventory [26:11] that actually fits right in, but sales haven't been great. [26:17] The recommendation-- rebrand and reprice it [26:20] above our current discount but below competitor pricing. [26:25] All of that with one simple prompt. [26:28] I can even see the sources that it analyzed [26:30] to make these recommendations. [26:33] Gemini Enterprise is also suggesting a new landing [26:36] page and new media to put front and center. [26:40] I'll ask my product strategy agent to generate some videos. [26:47] This part can also take a little bit of time, [26:49] so we'll look at some pre-generated options. [26:56] This looks amazing. [26:59] Veo 3.0 placed the exact pieces in a whole new organic living [27:03] style space. [27:05] Now, let's get that website updated. [27:08] I'll ask my dev agent to coordinate with our engineering [27:11] team. [27:13] It connects directly with Jira to give a developer full context [27:17] for a seamless handoff. [27:20] And on the developer's end, I'll automatically [27:23] get a notification right here in my Google Chat with that Jira [27:27] ticket. [27:28] I'll get this going with CLI. [27:33] Here, I'll ask my dev agent to start [27:35] working on that Jira ticket. [27:37] It's going to give me an overview of the strategy, [27:40] the assets, and how it plans to execute on this web page [27:44] using all of the context from the previous Gemini Enterprise [27:47] session, the ticket, and our brand and coding guidelines. [27:52] This is really easy. [27:53] Let's go ahead and get this built and deployed. [27:56] From here, this can take a little bit of time-- [27:59] I'm sure the devs in the audience [28:00] know that-- so I gave it a little bit of a head start. [28:03] Let's check out that final product. [28:07] This looks perfect. [28:10] Now let's head back over to Gemini Enterprise [28:13] to prepare our stores for the launch. [28:16] Our store operations team would usually handle this piece, [28:20] so I'm in a brand-new session. [28:22] I'll ask for a deck to get our regional distributors up [28:25] to speed. [28:27] Gemini Enterprise uses my company's context [28:30] to find exactly what I need to know about that launch, how [28:33] I've managed these in the past, and my team's sales goals. [28:37] Then it works with the Google Workspace agent [28:39] to create a branded deck in my personal style. [28:42] Again, this can take a minute or two, [28:44] so we'll look at one I generated ahead of time. [28:47] Get this. [28:48] With our brand-new Canvas mode, I [28:50] can make edits and collaborate in Google Slides [28:53] without ever having to leave Gemini Enterprise. [28:57] Overall, I think this looks great, [28:59] but I actually prefer the word "reimagined" here. [29:05] Perfect, and I'll go ahead and share this with my team. [29:14] And it looks like some of them are already jumping in. [29:17] Let's recap. [29:18] We saw Gemini Enterprise orchestrate multiple agents [29:21] with a single prompt, share context, [29:24] and generate and collaborate content [29:26] with Google Workspace, all supported [29:29] by our enterprise-grade governance, which [29:31] gives organizations the power to manage and secure agents [29:35] at scale, all while maintaining control [29:37] over their business data. [29:39] That's the magic of the Gemini Enterprise Agent platform. [29:44] [UPBEAT MUSIC, APPLAUSE] [30:03] THOMAS KURIAN: Thank you, Erica. [30:05] Thousands of companies around the world and across industries [30:08] are choosing Gemini Enterprise to transform their business. [30:12] We're seeing a profound shift. [30:13] Companies aren't just redesigning workflows. [30:16] They're turning their everyday employees into AI builders, [30:20] empowering them to solve their own hardest problems. [30:25] Signal Iduna is redefining insurance in Germany [30:29] with Gemini Enterprise. [30:31] They hit 80% adoption within weeks, [30:34] making it their front door for AI. [30:37] 11,000 employees are now building specialized agents. [30:41] Their health agent automatically verifies coverage [30:45] against a century of complex policy data, [30:49] driving a 400% surge in weekly users [30:53] and providing answers 37% faster. [30:57] Bosch is adopting Gemini Enterprise globally. [31:00] From finance to engineering, employees [31:04] are deploying custom agents so that they can reclaim time [31:08] for complex research. [31:09] KPMG reached 90% adoption with over 100 agents [31:14] in just the first month. [31:17] The American Society for Clinical Oncology [31:19] is now delivering cancer expertise faster. [31:23] Merck is bringing Gemini Enterprise to 75,000 employees [31:28] to support their purpose to save and improve lives. [31:34] And Walmart is rolling out Gemini Enterprise-- [31:38] [APPLAUSE] [31:41] Walmart is rolling out Gemini Enterprise [31:43] to help their store leaders spend more time with customers. [31:47] Let's take a look. [31:52] KIERAN SHANAHAN: At Walmart, we're [31:53] known for being people-led and tech-powered. [31:56] It's how we deliver on our promise [31:57] to help our customers save money and live better. [32:02] But in a retail landscape that's moving faster than ever, [32:05] there's one thing that's critical to our success-- [32:09] trust. [32:10] It's the backbone of our relationship with our associates [32:12] and customers. [32:14] And our field leaders build trust [32:15] by being present on the sales floor, [32:18] not in their offices at a screen. [32:20] So we're giving our associates the latest tech [32:23] and tools at their fingertips so that they [32:25] can spend more time on the floor with their teams, [32:29] serving our customers. [32:31] By combining our data with advanced AI technology, [32:35] Google Cloud is helping our leaders make better decisions [32:38] with clearer insights. [32:40] Gemini Enterprise is helping us take an important step forward, [32:44] helping our leaders lead and solving problems [32:47] so we can serve our customers better. [32:49] Our store and supply chain leaders [32:51] now have a Pixel Fold connected to Walmart's enterprise data [32:56] and giving them the answers they need in seconds, not hours. [32:59] By putting people first and taking their hands, [33:03] we're building a stronger future. [33:05] Thanks to our partnership with Google Cloud, [33:07] we continue to transform retail every day. [33:11] I'm excited for what's next. [33:15] [APPLAUSE] [33:21] THOMAS KURIAN: This winter, ahead [33:23] of one of the biggest moments in global sports, [33:25] our world-class engineers collaborated [33:28] with world-class athletes. [33:30] We're incredibly proud of how Google Cloud is [33:33] helping athletes push the boundaries of their sport [33:37] with AI. [33:41] SARAH KENNEDY: There are sports that [33:43] have a limited amount of data that's captured. [33:45] And I think that's a beautifully perfect challenge [33:48] for a company like Google. [33:49] SPEAKER 2: Let's roll it. [33:50] SARAH KENNEDY: One of the really cool things that we're [33:53] partnering with Team USA to do is [33:54] to help them understand what's that edge that we can possibly [33:57] give them? [33:58] MICHAEL SANTORO: They understand what's happening in the moment [34:00] better than anybody. [34:01] KATHERINE LITINSKY: So to have their input and their say [34:03] can really help us create something that's of benefit. [34:06] SARAH KENNEDY: The side-by-side comparison thing [34:08] I'm excited for. [34:08] KATHERINE LITINSKY: Everything is running on Google Cloud. [34:11] The 3D model that we're using is a collaboration with Google [34:14] DeepMind. [34:15] We've moved so far past talking to an LLM with text. [34:18] SPEAKER 3: It's so clear as day right here. [34:20] It's amazing. [34:20] SARAH KENNEDY: Whoa. [34:21] Is that me? [34:22] That's crazy. [34:23] Operating at the speed of sports is [34:25] one of the greatest challenges that we could be helping solve. [34:28] If we can do that for the Winter Olympics in Milan, [34:31] we can do that for any industry all over the world. [34:36] SPEAKER 1: Please welcome three-time Olympic gold [34:39] medalist, entrepreneur, and snowboarding legend, [34:42] Shaun White. [34:44] [CHEERING, APPLAUSE] [34:46] [ROCK MUSIC] [34:49] SHAUN WHITE: What's up, Google Cloud Next? [34:55] Oh, wow. [34:59] Man, I love seeing it snow in the desert. [35:02] This is wild. [35:03] All right. [35:04] Look, back when I was training, our tools [35:07] were camcorders and basically guesswork. [35:10] You'd land a trick and watch it back, [35:12] and you'd be looking at it thinking, how can I [35:14] make that trick better? [35:16] Over time, I've seen the tricks get bigger [35:18] and the walls get higher. [35:20] And now, Google Cloud is bringing AI to the mountain. [35:23] This winter, I worked with some crazy-smart Google Cloud [35:26] engineers on some awesome tech. [35:29] And to show you how rad this is, give it up for Jason Davenport. [35:34] [APPLAUSE] [35:36] [ROCK MUSIC] [35:42] JASON DAVENPORT: What's up, Shaun? [35:43] SHAUN WHITE: What have we got for everyone? [35:45] JASON DAVENPORT: All right, Shaun. [35:46] This is going to be super fun. [35:48] Let's unleash the power of Google Cloud AI [35:50] into the half-pipe. [35:52] And to do that, we're going to pull up a trick from your 2017 [35:56] Burton US Open. [35:57] SHAUN WHITE: Oh, wow. [35:58] JASON DAVENPORT: I'm sure you remember this one. [35:59] SHAUN WHITE: It's a throwback. [36:00] It's a throwback. [36:00] JASON DAVENPORT: It is. [36:01] So let's analyze this clip for the audience. [36:03] And so they can see all the cool things that AI sees. [36:06] SHAUN WHITE: All right. [36:06] JASON DAVENPORT: All right, let's start it off. [36:08] First, this trick is over in under three seconds. [36:12] You are literally-- it's a split-second blur. [36:14] And for everyone watching, what trick is this? [36:17] SHAUN WHITE: This is a switch cab double flip 1440. [36:20] JASON DAVENPORT: That is insane. [36:21] But for everyone else, what does that actually mean? [36:25] SHAUN WHITE: Basically, it's my unnatural way of riding. [36:28] I do four full rotations with two flips in the middle. [36:31] JASON DAVENPORT: I tried stopping using the camcorder, [36:34] and I couldn't do it. [36:35] So let's slow this down. [36:37] And to do this, we need to analyze this frame by frame [36:40] and see what Google Cloud sees. [36:42] And we're going to do this essentially [36:44] with every single stop in this. [36:46] So first, let's start with your pose. [36:49] So this is what's super cool. [36:51] We built a model in collaboration [36:53] with Google DeepMind that can track you spatially, [36:56] and it creates a three-dimensional, [36:57] essentially pose of you from a flat, two-dimensional video. [37:01] SHAUN WHITE: So awesome. [37:02] JASON DAVENPORT: I mean, I love it. [37:03] I can't even see you here. [37:05] How does it do it? [37:06] SHAUN WHITE: This was my tight pant, leather jacket phase. [37:09] So you got to really punch in on it. [37:11] JASON DAVENPORT: I know. [37:13] I got some out back, too. [37:14] SHAUN WHITE: It was a good look. [37:15] JASON DAVENPORT: I know. [37:16] It's still a good look. [37:17] All right. [37:17] Let's talk about the next thing. [37:19] These are stats powered by Gemini. [37:21] Here, we're tracking your flight dynamics, rotational velocity, [37:25] and even your tuck compression. [37:27] SHAUN WHITE: Yeah, this is amazing. [37:28] We didn't have this before. [37:29] I can see how much time I've spent in the air. [37:31] And now take old footage that I've done [37:34] and compare it with new footage of how much time was I [37:36] in the air when I landed the trick, [37:38] and when I didn't land the trick. [37:39] And I can compare the two. [37:40] And that data is really going to help me progress. [37:42] And obviously, the next generation [37:44] can use these tools too. [37:45] JASON DAVENPORT: It's true. [37:46] And what I love is I love seeing how [37:47] quickly your rotational velocity gets up here. [37:50] It's all about how you exit the pipe. [37:52] SHAUN WHITE: Torque-- exactly, and coming out of it. [37:54] JASON DAVENPORT: All right. [37:54] So let's look at one more cool thing that we've done here, [37:56] which is the ribbon overlay. [37:58] SHAUN WHITE: Oh, I love this. [37:58] JASON DAVENPORT: So here's what I love about this. [38:01] We're actually showing your cork here [38:02] with the whole visualization. [38:04] So what happens here? [38:05] We're moving from blue to green on the thing. [38:08] What is that? [38:08] SHAUN WHITE: When I was working with the engineers, [38:10] I was trying to describe this trick. [38:12] And there's a turning point where normally I [38:14] do one front flip and then come in for the second, [38:17] but then you have to pause and then go backflip. [38:20] If you miss that turning point in the trick, well, [38:23] it's not good. [38:25] It's not good. [38:26] But this really pinpoints it exactly. [38:28] JASON DAVENPORT: It is so cool to watch this. [38:29] And what I love is that this is making sports [38:31] more accessible for fans. [38:33] We can see all of the craziness that you're [38:35] putting into this sport. [38:36] SHAUN WHITE: Exactly. [38:37] JASON DAVENPORT: All right. [38:38] So let's bring this back and talk a little bit more tech. [38:41] Google Cloud is helping everyone innovate here [38:44] by creating new use-case specific models, even [38:46] for things like spatial analysis, [38:49] training these large amounts of data and even models [38:53] with Google Cloud TPUs, and helping everyone [38:56] to build a secure data and agent foundation with Gemini [38:59] Enterprise Agent Platform. [39:02] Compute, models, and platform-- this [39:05] is the power of Google Cloud AI. [39:07] SHAUN WHITE: And look, from my personal experience, [39:09] I can tell you, learning a trick on the mountain is one thing, [39:12] but actually understanding the physics of a trick [39:15] is a whole other thing. [39:16] And this is really going to help not only [39:18] the next generation of athletes learn new skills [39:21] but also help the fans at home understand the sport better. [39:24] And not just for snowboarding, but I think sports globally. [39:28] Look, I'm so stoked to see what's ahead. [39:30] Thank you, Jason. [39:31] And thank you, Google Cloud Next. [39:34] [APPLAUSE, UPBEAT MUSIC] [39:39] SPEAKER 1: Please welcome SVP and chief technologist, [39:42] AI and infrastructure, Amin Vahdat. [39:47] [UPBEAT MUSIC] [39:59] AMIN VAHDAT: Thank you, Shaun. [40:00] We're taking the exact same technology [40:03] that empowers athletes and applying it to the enterprise. [40:06] Because while athletes use data to understand the game, [40:08] the enterprise needs it to change the game. [40:11] That kind of scale is the ultimate stress test [40:14] for a unified system. [40:15] When you process massive video feeds, reaching over [40:18] decades of historical data, and defend [40:20] every byte in near real time, you [40:22] have a foundation built for anything. [40:26] But to solve this at a global scale, [40:28] we first have to solve a physics problem. [40:31] It is a constant struggle against the limits [40:33] of software architecture, hardware architecture, power, [40:37] cooling, and even the speed of light. [40:40] This is the foundation of our AI Hypercomputer. [40:43] We integrate clean energy, massive physical scale, [40:47] and our purpose-built infrastructure [40:48] into a single unified engine of efficiency and innovation. [40:54] Because in the agentic era, compute [40:55] is no longer defined by chip. [40:57] Compute is the entire data center. [41:02] We give you the ultimate flexibility [41:04] to pick the right architecture without compromise. [41:07] You can see this in how different [41:08] industry leaders are innovating on Google Cloud. [41:11] Axia Energia, Brazil's largest power company, [41:14] helps prevent power outages for millions of customers [41:17] by running advanced weather modeling on TPU clusters. [41:21] By forecasting severe weather conditions up [41:23] to 10 days in advance, the company [41:25] can proactively plan and execute preventative measures [41:28] before storms hit, significantly reducing service interruptions [41:33] year over year. [41:35] Thinking Machine Labs shows our NVIDIA-based infrastructure [41:38] to power Tinker, an open platform [41:41] for reinforcement learning and fine-tuning [41:43] of frontier models for specialized use cases. [41:47] Woven by Toyota achieved 42% faster training [41:50] for models that predict complex traffic events. [41:54] The US Department of Energy is powering AI coscientists [41:58] across all 17 national labs to accelerate the pace [42:02] of scientific discovery. [42:05] And Boston Dynamics is training and safely [42:08] deploying flexible vision-language models [42:10] to scale robotics across diverse industrial applications. [42:16] As we enter the agentic era, we see [42:18] that the demands of training and serving [42:20] have completely diverged. [42:22] To meet the performance needs of these explosive workloads, [42:26] I am proud to announce our eighth-generation TPUs. [42:31] [UPBEAT MUSIC] [42:43] [APPLAUSE] [42:51] These are things of beauty. [42:53] For the very first time, we are launching [42:55] two specialized platforms, each built [42:57] from the ground up for the distinct demands of training [43:00] and serving. [43:02] TPU 8t is a powerhouse optimized for training. [43:06] We have redefined performance capability [43:08] by moving block scale multiplication directly [43:11] inside the MXUs. [43:13] This native MXU quantization eliminates VPU overhead, [43:17] delivering nearly three times the compute performance per pod [43:19] over previous generations. [43:22] This allows us to push the absolute limits of model flops [43:25] utilization at a massive scale and reduces the training time [43:28] of frontier models. [43:31] It leverages our breakthrough inter-chip interconnect [43:33] technology, which now delivers twice the bandwidth compared [43:36] to Ironwood, scaling up to 9,600 TPUs connected via our 3D torus [43:42] topology. [43:45] This is a 2.8x improvement over Ironwood to deliver 1.21 [43:49] exaflops of FP4 compute per pod. [43:55] 8t provides two petabytes of shared bandwidth memory [43:58] in a single superpod and utilizes the new TPUDirect [44:01] storage to enable high-speed data transfers from managed [44:05] storage. [44:07] To put that scale in perspective, [44:08] two petabytes is enough to hold the entire digital connection [44:12] of the Library of Congress 100 times over. [44:18] TPU 8i is optimized for the thinking part of the equation-- [44:21] inference and reinforcement learning. [44:23] At the chip level, we have integrated a specialized [44:26] collectives acceleration engine, which reduces latency [44:29] by an additional 5x. [44:31] By hosting the memory cache entirely on silicon, [44:34] we've finally broken the memory wall that slows down [44:36] long-context decoding. [44:38] We then scaled this on-chip performance [44:40] through our new board flight topology, [44:43] which deploys 1152 TPUs into a single pod [44:47] to run millions of concurrent agents with near-zero latency. [44:52] This delivers 11.6 FP8 exaflops for a substantial 9.8x increase [44:58] in performance over the 256-chip Ironwood pod. [45:05] We are extending our infrastructure leadership [45:08] to general purpose workloads as well with Google Cloud Axion. [45:12] Our Google Axion N48 compute instances, [45:15] powered by a custom-designed ARM CPU, [45:17] deliver up to twice better price performance [45:20] and 80% better performance per watt [45:23] than comparable x86 instances. [45:25] This provides a sustained, continuous operation [45:28] designed to eliminate cold starts in logic apps, [45:31] ensuring your agents are always on and ready to respond. [45:36] We are the preferred cloud for NVIDIA GPUs [45:39] by some of the largest scale and most innovative customers [45:42] in the world. [45:43] Today, I am pleased to announce that Google Cloud will [45:46] be among the first to offer the NVIDIA Vera Rubin NVL72. [45:51] [APPLAUSE] [45:57] Vera Rubin NVL72 on Google Cloud achieves 10 times [46:01] performance efficiency, optimized for high interactivity [46:04] and long-context workloads. [46:06] Of course, silicon is only half the battle. [46:09] To feed these chips, we've redesigned our entire storage [46:12] and networking pipeline. [46:14] I am pleased to announce that Managed Lustre now supports [46:16] throughput up to an industry-leading 10 [46:19] terabytes per second. [46:21] We tied it all together with our new Virgo Network. [46:24] Virgo doubles connectivity to scale training [46:27] beyond the superpod. [46:28] It links 134,000 chips with up to 47pb per second [46:33] of non-blocking bandwidth to deliver over 1.7 million [46:37] exaflops of compute. [46:40] We can now turn months of training [46:41] into weeks with the power of a million-plus TPU [46:44] chips in a single cluster, all orchestrated [46:48] by Pathways and JAX. [46:52] [APPLAUSE] [46:58] With up to four times the bandwidth [47:00] per accelerator over the previous generation, [47:02] this combination delivers near-linear scaling, [47:04] ensuring you get the full power of every chip you pay for. [47:08] We are also making Virgo available on NVIDIA, Vera Rubin, [47:11] and NVL72, supporting up to 960,000 GPUs. [47:18] Leading financial institutions are [47:19] using this exact combination of silicon networking [47:22] to gain a massive edge. [47:24] Let's hear from Citadel Securities on how [47:26] we are innovating together. [47:29] JOSH WOODS: We trade over half a trillion dollars per day. [47:32] But it's not just about volumes. [47:34] It's also about speed. [47:35] In financial markets, time is measured in nanoseconds. [47:40] We connect markets all around the world, [47:42] whether pension funds or somebody doing their first trade [47:45] as part of their retirement, meet their investment [47:48] objectives. [47:49] We've partnered with Google Cloud [47:50] to build what we think is the preeminent scalable, cloud-based [47:54] research environment. [47:56] For instance, we've been working in depth on TPU [47:59] to ensure that we get the efficiency we [48:02] need to run our business. [48:06] Today, we're able to run workloads two to four times [48:09] as fast with a 30% lower cost. [48:12] HARIS NAIR: With TPU Ironwood, we [48:14] can now run thousands of parallel chips [48:16] for a single workload. [48:17] And what that means for our business is workloads [48:20] that used to take weeks or days, we [48:23] can now execute those in hours or even minutes. [48:27] Imagine you're a quantitative researcher. [48:29] You have 100 novel ideas that you want to execute. [48:32] Can we empower our researchers to be [48:34] able to test all of their ideas and really [48:36] be limited by not the scale of the platform [48:39] or the economics of the problem, but by their own creativity? [48:44] JOSH WOODS: As we like to say, steel sharpens steel. [48:46] Google Cloud and Citadel Securities' collaboration [48:49] has been exactly that. [48:54] [APPLAUSE] [49:00] AMIN VAHDAT: Citadel Securities is just one example [49:02] of many innovators who are gaining a competitive edge [49:04] on our foundational AI Hypercomputer. [49:07] At this scale, you cannot have humans manually troubleshooting [49:10] configurations. [49:12] You need a cloud that drives itself. [49:15] We have used the Model Context Protocol to turn every Google [49:18] Cloud service into a tool that agents can orchestrate directly. [49:22] By integrating Gemini reasoning into our own telemetry, [49:26] the system now performs autonomous root cause analysis, [49:29] identifying and fixing misconfigurations [49:32] before you even realize there's a problem. [49:35] This is the AI Hypercomputer-- purpose-built infrastructure [49:39] to accelerate the entire AI lifecycle. [49:42] To show you how we feed this engine with your data, [49:44] please welcome Karthik Narain. [49:47] [UPBEAT MUSIC] [50:01] KARTHIK NARAIN: Thank you, Amin. [50:03] Amin just showed you how we built the world's most--powerful [50:06] reasoning engine. [50:08] But here's the reality. [50:09] Reasoning without context is just a guess. [50:13] And when you expect your AI to make decisions and your agents [50:16] to take actions, you cannot afford to guess. [50:20] Trusted context turns an intelligent guess [50:23] to a decisive action. [50:26] Today, we are completely rethinking the data platform. [50:29] We call it the Agenetic Data Cloud. [50:34] To make this real, I'm thrilled to share [50:37] four foundational innovations. [50:40] It all starts with the foundation of truth. [50:43] Right now, your data is trapped in PDFs, video calls, [50:47] multiple data stores, and SaaS applications. [50:51] How do we help AI make sense of it all? [50:54] Today, we are introducing the Knowledge Catalog, [50:57] a universal context engine for your enterprise. [51:01] Starting with your structured world, [51:03] the Knowledge Catalog natively integrates [51:06] with BigQuery, mapping tables and metadata [51:09] into unified business logic. [51:12] And we are going even further, extending the same capability [51:15] to the unstructured world as well through Smart Storage. [51:20] In the past, a file landing in your storage [51:22] sat passive, waiting for a pipeline. [51:25] Now, the second that image or PDF hits Google Cloud Storage, [51:30] they are instantly tagged, enriched, and made agent-ready-- [51:34] zero manual data engineering. [51:37] Powered natively by Gemini, the Knowledge Catalog [51:40] goes even deeper, reading files, autonomously extracting [51:44] entities, mapping relationships, and learning [51:48] about your unique business semantics. [51:51] When an agent hears net revenue or risk, [51:54] it understands the exact meaning. [51:57] Take Virgin Media O2. [51:59] They had over 20,000 data assets, [52:01] many of which were untapped. [52:04] Knowledge Catalog is helping them [52:06] activate this data to empower their global product teams. [52:11] Now, combined with direct zero-copy access [52:14] to applications, operating systems, and AI platforms [52:17] like Palantir, Salesforce, SAP, ServiceNow, and Workday, [52:22] you now have instant context to your entire business. [52:28] [APPLAUSE] [52:30] That's right. [52:33] This trusted context enables agents to act with certainty. [52:39] Now imagine this power in the hands [52:41] of every employee, every day. [52:44] We've integrated the Knowledge Catalog with Gemini Enterprise's [52:48] Deep Research Agent. [52:50] You can now enable multi-step reasoning [52:53] across internal and web data for precise cited business answers. [52:58] What took weeks of manual effort now happens seamlessly. [53:05] This is the Knowledge Catalog. [53:08] Second, we are unveiling a Gemini-powered data science [53:12] authoring experience that fundamentally transforms [53:16] how data practitioners work with our new Data Agent Kit. [53:22] We didn't just want to build another platform [53:24] you have to learn. [53:25] Instead, the Data Agent Kit integrates comprehensively [53:31] with libraries of AI skills and plugins directly [53:34] into your workflows that you already use. [53:37] Your IDE, your notebook, your terminal, [53:40] whether you are in VS Code, Cloud Code, or Gemini CLI, [53:44] we instantly turn your everyday workspace [53:47] into a native data environment. [53:49] Here's how it works. [53:51] You just state your intent, predict customer churn, [53:56] and your environment simply takes over. [53:59] It autonomously builds the pipelines [54:01] and deploys the models right on Google's Agentic Data Cloud. [54:06] It handles all the complex orchestration for you, [54:09] creating a direct path from your business objective [54:12] straight to the business outcome. [54:15] For researchers at Bayer Crop Science, [54:17] this eliminates manual data management and analytics, [54:21] freeing them to focus on pioneering agricultural [54:24] innovations for a sustainable and productive future. [54:30] Third, the demand on data is extreme in this agentic era. [54:35] To address the scale challenge, we're [54:37] unleashing the Lightning Engine for Apache Spark. [54:41] It not only delivers industry-leading performance [54:44] for Spark. [54:45] Lightning Engine delivers up to 2x the price performance over [54:50] the previous market leader. [54:52] [APPLAUSE] [54:57] We fundamentally rebuilt this engine for the agentic era, [55:01] reasoning with trusted context enables autonomous agents, [55:05] and with the Lightning Engine, that autonomy [55:08] operates at massive scale. [55:12] Industry-leading organizations like Flipkart, [55:14] Lowe's, and Meesho are accelerating their Apache Spark [55:18] workloads with Lightning Engine. [55:23] Finally, let's address the biggest [55:25] and the longest-running debate of all. [55:28] Where should your data reside? [55:30] The reality is it lives everywhere-- [55:33] at Google, at AWS, Azure, and across your SaaS applications. [55:38] Your old lakehouse expected the analytical engine and the data [55:42] storage to reside in the same cloud. [55:45] This approach is broken. [55:48] Today, we are introducing the Cross-Cloud Lakehouse, [55:53] where your analytical engine reasons over data in any cloud. [55:59] [APPLAUSE] [56:04] Built on the open Apache Iceberg standard, [56:07] it is completely borderless. [56:09] Instead of forcing you to accept complex networking hurdles [56:13] or massive egress fees, we deliver low-latency direct [56:16] connectivity to AWS and Azure as if the data sat natively [56:21] in Google Cloud-- [56:23] no more moving data, no more vendor lock in-- just freedom. [56:29] [APPLAUSE] [56:33] No matter where your data lives, from traditional systems [56:36] to SaaS applications, we provide the trusted context [56:40] and connectivity your business needs to act upon. [56:44] This is why customers are choosing our Agentic Data [56:48] Cloud to drive their agentic transformation. [56:53] One example of this transformation is Vodafone. [56:57] Vodafone unified their data on BigQuery [57:00] to build a more resilient global network for their customers. [57:04] And with Gemini Enterprise, they are launching hundreds of agents [57:08] to proactively resolve outages and optimize infrastructure, [57:12] scaling network reliability and saving millions every year. [57:18] Macquarie Bank's goal was to securely scale [57:21] AI across their entire operation to deliver [57:24] more personalized customer experiences. [57:27] They deployed Gemini Enterprise to all staff [57:30] and supercharged productivity. [57:32] And for their two million customers, [57:34] Macquarie's 24 by 7 AI assistant queue autonomously [57:39] answers banking questions. [57:42] All of this was made possible by unifying their data [57:45] on BigQuery and Spanner. [57:48] With this foundation, Macquarie has already cut client losses [57:52] from scams by half. [57:55] That's secure, frictionless banking at scale. [58:00] American Express is redefining trust [58:03] by centralizing its core data on Google Cloud [58:06] to enable faster fraud and risk analysis at global scale. [58:11] Costco leverages BigQuery to accelerate member insights, [58:15] empowering associates to optimize experience and maximize [58:19] value for millions of members. [58:22] This is the Agentic Data Cloud. [58:25] You and your AI agents now have a platform [58:27] that operates with universal context, [58:30] is defined by your defined intent, [58:33] and scales autonomously across your borderless data [58:37] and the applications you already use today. [58:40] [APPLAUSE] [58:44] To show you how to bring trusted context to your business, [58:48] please welcome Yasmeen Ahmad. [58:51] [UPBEAT MUSIC] [59:02] YASMEEN AHMAD: Karthik gave us the recipe, [59:04] so now let's see how it all plates up. [59:07] Most companies take five weeks to turn a trend into a decision. [59:11] We're going to do it in five minutes. [59:14] To get there, we're breaking three barriers-- [59:17] dark data with our Knowledge Catalog, data [59:20] silos with our Cross-Cloud Lakehouse, and manual code [59:23] with Google's new agentic tools and skills. [59:27] So let's get cooking. [59:29] Our new agentic workflow triggers [59:31] have detected a tasty trend-- [59:33] Midnight Swirl froyo. [59:36] Looks delicious. [59:37] But with any new flavor, I need to know is it safe? [59:41] Are there any hidden allergens? [59:43] Is there a market? [59:44] Where are my hungry customers? [59:46] And is it worth it? [59:47] What's the real ROI? [59:49] But first, safety. [59:52] We have thousands of PDF recipes. [59:55] I'm going to search for soy. [59:57] It's a top food allergen, and safety is non-negotiable. [60:01] Zero results-- looks good so far. [60:05] But this recipe does contain an ingredient, base 204. [60:10] Let's check the supplier manual for base 204. [60:14] And here it is. [60:16] Base 204 actually contains soy. [60:19] A simple GenAI search would miss this connection [60:22] because the information is trapped [60:25] across two separate PDFs. [60:27] So how do we digest all of this data [60:30] to find those hidden connections? [60:32] Well, I need an agent that has the skill [60:35] to work with my PDF data combined with our Knowledge [60:38] Catalog. [60:40] Watch. [60:41] Our agent helps us find hidden allergens [60:43] in a recipe for Midnight Swirl. [60:46] And here it is. [60:48] Midnight Swirl contains soy, and it's giving me a data citation, [60:53] a schema called product specs. [60:55] Now, this schema was generated by our Knowledge Catalog, [60:59] working with Gemini to reason over our supplier and recipe [61:03] PDFs and extract entities like recipe, ingredient, and allergen [61:09] and mapping those previously invisible connections [61:12] between them. [61:13] This is trusted context built over dark data. [61:18] Next, market size-- to launch this globally, [61:22] I need a precision list of customers [61:24] with zero soy allergies. [61:27] However, I have two data sets that have never met. [61:30] Our allergen schema is here in Google, [61:33] while my loyalty list with millions [61:36] of customers and dietary preferences [61:38] is in AWS S3 Iceberg. [61:42] Now, usually you'd call your data engineer [61:44] and wait three days for a migration. [61:47] Not today. [61:48] No more slow churn. [61:51] With our cross-cloud Lakehouse, the data [61:53] stays exactly where it is while we dynamically build the list. [61:58] Before writing a single line of code, [62:00] Gemini builds a plan using our data engineering [62:03] tools while respecting my built-in security [62:06] and permissions. [62:07] As the data scientist, you're the chef. [62:11] You can review. [62:13] Let's accept and see what our agents serve up. [62:17] They execute on our Lightning Engine, [62:20] which is two times more price-performant than the market [62:23] proprietary alternative, filtering [62:26] our soy-sensitive customers in mere seconds. [62:32] AWS and BigQuery sources connected-- [62:35] zero complexity. [62:36] It's as smooth as a batch of Midnight Swirl. [62:40] Finally, ROI-- is this worth a global launch? [62:44] An annual forecast would help. [62:48] Again, our agents don't guess. [62:50] Gemini builds a multi-step execution plan. [62:54] Can you make changes? [62:55] Yes, Chef. [62:57] So let's take a look and change the number of simulations from [63:02] 1,000 to 2,000. [63:06] And we will accept. [63:10] Acting as the orchestrator, Gemini delegates [63:13] to our specialized data science tools, [63:15] choosing the right models, and again, [63:17] executing on our Lightning Engine for Spark. [63:21] It's built an entire notebook for me. [63:24] Now, training usually takes a bit of time, [63:26] so here's one I ran earlier. [63:29] $15 million. [63:32] Even after protecting our soy-sensitive market, [63:35] the demand is massive. [63:37] Can we bring back the timer? [63:39] Look what we accomplished. [63:41] In less than five minutes, we turned a viral trend [63:44] into a $15 million decision. [63:47] First, Knowledge Catalog found the hidden soy connection. [63:51] Second, Cross-Cloud Lakehouse, powered by the Lightning Engine, [63:55] connected to AWS with zero complexity. [63:59] And third, our agentic skills and tools built the forecast [64:03] with zero manual code. [64:05] This is the Agentic Data Cloud. [64:08] [TIMER BEEPING] [64:10] [CHEERING, APPLAUSE] [64:17] We did it. [64:18] And now, here to share how we're securing [64:21] your agentic enterprise, please welcome Francis deSouza [64:26] [UPBEAT MUSIC] [64:40] FRANCIS DE SOUZA: In an agentic enterprise, [64:42] giving agents the power to act means [64:45] that security must become an autonomous force, moving faster [64:49] than the threats themselves. [64:51] Human analysts simply can't keep up with AI-driven attacks. [64:56] The mean time to exploit the vulnerability [64:59] has dropped to minus seven days, meaning [65:01] that today exploits routinely occur before a patch is [65:06] released, and the handoff time between initial access [65:10] and handover to a secondary threat group [65:13] has dropped from eight hours to 22 seconds. [65:17] Your security must operate at machine speed. [65:21] Now, to do that, we're moving the heavy lifting [65:24] to a Gemini-native agentic Security Operations [65:28] Center, or SOC. [65:30] And the results are already here. [65:32] Our triage agents are turning 30-minute investigations [65:36] into 60-second resolutions. [65:40] And our new threat hunting and detection agents [65:43] proactively sweep your environments [65:44] for risks at a scale and speed that no human team could match. [65:51] By leveraging Google's unparalleled telemetry, [65:54] including Mandiant, VirusTotal, and Chrome, [65:57] we are delivering a defense system [65:59] based on global intelligence. [66:02] And now, with our integrated dark web intelligence, [66:06] we can identify external threats with 98% accuracy. [66:11] Customers are already seeing big benefits [66:13] from Google Cloud's security offerings. [66:16] CME Group depends on a unified stack of Mandiant Security [66:21] Command Center and Google's SecOps [66:23] to protect the world's largest derivatives [66:26] system against novel threats. [66:30] In an ultra high-stakes environment, [66:33] Google's security offering is matched by speed. [66:37] CME Group achieves nanosecond precision [66:40] with Google's ultra low-latency solution, [66:43] supporting billions of transactions [66:46] daily and lowering the barrier of entry for new traders. [66:50] Now, the biggest threats today aren't just hackers. [66:53] It's also shadow AI, unauthorized models and agents [66:58] that are operating in your enterprise [67:00] but outside your control. [67:02] And to meet this challenge, we have [67:04] integrated the industry's deepest security context [67:08] directly into our AI fabric. [67:11] That is why I am thrilled to officially welcome [67:14] the Wiz team to Google Cloud. [67:17] [APPLAUSE] [67:24] Together, we are building a new security posture [67:27] for the agentic enterprise. [67:29] While the New agent platform in Gemini Enterprise [67:32] secures and governs your agents, with Wiz, [67:35] we're now extending that protection [67:37] to every asset on your premises and across all major clouds. [67:43] To explain how we're solving this, [67:45] I'm excited to introduce to the stage Wiz co-founder Yinon [67:49] Costica. [67:51] [UPBEAT MUSIC] [68:07] YINON COSTICA: Thank you, Francis. [68:09] At Wiz, our mission from day one has [68:12] been to help customers protect everything they build and run. [68:17] Wiz began by unifying code and cloud and runtime context [68:21] to move at developer speed. [68:23] But AI has fundamentally changed the environment. [68:28] With autonomous products, hyperscale code generation, [68:31] and AI-weaponized threats, the stakes have never been higher. [68:36] To protect this new frontier, security [68:39] must now move at machine speed. [68:42] And that is exactly why we built Wiz [68:45] as the first AI Application Protection [68:48] Platform, or in short, AI APP. [68:51] It solves four key security challenges for the AI era-- [68:55] giving security unparalleled visibility into the AI stack, [69:00] finding and proactively remediating critical risks [69:05] before attackers can use them, enabling our builders [69:09] to start secure by design in their AI [69:12] enabled IDEs, and lastly, arming SecOps teams to stop threats [69:17] to cloud and AI environments. [69:19] In order to outpace attackers, Wiz [69:22] delivers a set of expert AI agents. [69:26] It's a really cool concept. [69:28] Our red, blue, and green agents are [69:32] named after the teams they help-- red, blue, [69:35] and green teams. [69:36] They form an AI layer to autonomously identify, [69:40] investigate, and then fix critical risks at machine speed, [69:44] leveraging context from our unique security graph. [69:49] Security starts with visibility, always. [69:52] But today. [69:53] Your AI footprint is an orbit of interconnected tools. [69:57] It spins models like Gemini, Claude, OpenAI. [70:01] Your teams are using agent studios [70:03] such as Gemini Enterprise Agent Platform, Lovable, Copilot [70:09] Studio, Salesforce's Agentforce. [70:12] And to secure all of these environments, [70:14] we examine every layer-- [70:17] the clouds, the data, the models, the agents. [70:21] And we continuously correlate these risks [70:24] to find and fix the attack path that matter most. [70:28] Now let's see it in action. [70:31] [UPBEAT MUSIC] [70:40] So this is Wiz. [70:42] Wiz basically starts by automatically building [70:45] a dynamic inventory from your code and cloud, [70:48] completely agentlessly. [70:50] This means that you can see every visibility [70:54] into the environment, into everything [70:56] that your teams are building using AI, without any friction. [71:02] Wiz then builds the security graph. [71:05] The security graph-- think of it as a living map that [71:08] explains the architecture and the logic of any AI application [71:12] it sees. [71:13] Here we see an example of an agent that [71:15] is actually running on cloud. [71:18] It has tools to query a database and even execute code. [71:22] And you can see that Wiz actually [71:25] flags that this agent is internet-exposed. [71:29] And also, it has access to sensitive customer data. [71:33] Now, this is security moving at the speed of AI. [71:38] You don't need any reviews with your development teams. [71:41] It just works. [71:43] And in the AI era, the best defense [71:47] is to continuously use AI against ourselves [71:50] in order to give the defenders the first-mover advantage [71:55] over the attackers. [71:56] And this is exactly what the Wiz Red Agent does. [72:00] It validates every single exposure that Wiz identified. [72:04] And you can think of it as a friendly hacker [72:07] that continuously scans your outside like an elite red team. [72:12] And look, this is a finding that the Red Agent actually [72:16] found-- an authentication bypass vulnerability. [72:20] This means that it enables an attacker [72:23] to gain access to that agent but also to that sensitive database [72:28] behind it to exfiltrate sensitive information. [72:32] Now, it's not a potential risk anymore. [72:35] This is a validated risk. [72:37] And this gives our developers the prioritization [72:41] and the confidence in fixing it immediately. [72:44] Now, in the past, triage used to mean very, very long meetings [72:49] and waiting and spreadsheets. [72:51] And you don't need that anymore because we have the Wiz Green [72:56] Agent. [72:57] The Wiz Green Agent automates the entire triage process [73:02] from identifying the owner, suggesting the fix, [73:06] identifying the exact line of code that [73:09] caused the risk in the first place-- [73:11] all automated-- no phone tag, no friction. [73:14] And from here, you have a choice. [73:15] You can send it to the developers as a PR, [73:18] or maybe better, send it to a coding agent like CodeMender [73:23] to execute the final fix at the source automatically. [73:26] And this is how we can move from a validated risk [73:30] to a verified fix, end-to-end visibility, control, [73:34] and full confidence. [73:36] So to summarize, whenever you have an AI team that [73:40] ships a new product or even someone in finance [73:44] that then codes a new agent, now, within minutes, security [73:48] is able to, one, identify the agent and its architecture, two, [73:53] automatically conduct a security review to find and validate [73:57] the risks, and three, automatically suggest the fix, [74:00] send it to the dev teams in their own native tools. [74:03] Ta-da. [74:04] This is how AI agents unlock secure innovation [74:08] with AI at scale. [74:11] Thank you, all. [74:12] Back to you, Francis. [74:14] [APPLAUSE, UPBEAT MUSIC] [74:23] FRANCIS DE SOUZA: The unique combination [74:24] of Google's agentic SOC and Wiz's deep cloud context [74:29] gives customers the confidence to build and deploy generative [74:32] AI everywhere safely. [74:35] We're helping the Los Angeles Department of Water and Power [74:38] secure critical infrastructure ahead of the LA 28 games, [74:43] while Singapore's CSIT enables proactive defense [74:47] against advanced digital threats. [74:50] DBS strengthens security by embedding Google Cloud's [74:53] protection directly into their architecture, [74:56] enabling real-time threat detection and response, [74:59] reinforcing customer trust. [75:02] Morgan Stanley chose Wiz as a key component of its cloud [75:06] security strategy and is expanding visibility and control [75:10] across its cloud environments. [75:12] Now, as part of this strategy, the firm [75:14] is deploying Google Cloud Security capabilities to support [75:18] its evolving cloud platforms. [75:21] And trusted global icons from Nestlé [75:24] to LVMH, from BMW to Shell, rely on Wiz [75:28] and Google Cloud Security to ensure their AI-driven future [75:32] is secure by design. [75:34] We have closed the loop between the defender and the builder. [75:38] Google and Wiz provide a unified agentic defense [75:43] that is secure by design and autonomous by nature. [75:47] Together, we are redefining cybersecurity for the AI era [75:52] so security teams can protect their organizations at machine [75:56] speed. [75:58] To share more about how our agents are [76:00] helping your customers, please welcome Carrie Tharp. [76:04] [UPBEAT MUSIC] [76:20] CARRIE THARP: Thank you, Francis. [76:22] A secure foundation gives you the confidence to innovate. [76:26] At Google, we are using that foundation [76:28] to solve two of your biggest challenges-- [76:31] serving your customers and helping your employees [76:34] work smarter. [76:35] We do this by deploying a coordinated agentic task [76:39] force, specialized agents that don't just [76:43] assist your team-- they operate alongside them. [76:47] In the agentic era, an agent is no longer just a tool. [76:52] It's a strategic extension of your business, built [76:55] to expand your reach, deepen engagement, and personalized [77:00] service at scale. [77:02] Earlier this year, we launched Gemini Enterprise [77:05] for Customer Experience, a suite designed [77:09] to enhance the entire customer journey [77:11] from the first moment of discovery [77:14] to ongoing service interactions. [77:17] Central to this is agentic commerce. [77:20] We have introduced pre-built shopping and food [77:24] ordering agents that handle everything [77:27] from discovery to checkout entirely [77:30] through natural language. [77:31] By pairing these agents with our customer engagement agents, [77:36] we've created a seamless journey from search [77:39] to service, turning every interaction into revenue [77:43] and retention opportunities. [77:45] For example, with the food ordering agent, [77:48] Papa John's is building a hyper-personalized system [77:53] that remembers customer preferences to get [77:56] the food to your door faster. [77:58] During live interactions, Agent Assist [78:01] coaches employees to deliver fast and more accurate [78:05] answers to customer questions. [78:08] For example, Best Buy is guiding shoppers [78:11] through those complex tech specs, [78:13] issue resolution, and appointment scheduling, [78:17] all autonomously. [78:19] And our new omnichannel gateway helps [78:21] ensure your agents maintain contacts across every surface-- [78:26] web, mobile, voice-- while preserving universal consumer [78:31] context. [78:33] If a customer moves from a text chat to a phone call, [78:37] the agent seamlessly remembers exactly where they left off, [78:41] and this scales globally, conversing [78:43] with human-like voice capabilities [78:46] in multiple languages. [78:48] Gemini Enterprise for Customer Experience [78:51] is the technology powering the Home Depot's Magic Apron [78:54] Assistant to reach customers across every channel. [78:58] Let's hear from the Home Depot. [79:00] JORDAN BROGGI: There's a customer journey [79:02] that often starts with inspiration, goes into research, [79:04] and then into shopping and buying and then into doing. [79:07] And Magic Apron is really a digital agent [79:09] to help through that whole journey, [79:11] and it is really there to provide that orange apron [79:15] experience for help and customer service, the same way [79:18] that as you come into a store, you're used [79:20] to that great customer service. [79:23] So we've had a partnership with Google for quite some time, [79:26] and the latest things that they're bringing to market [79:28] are incredibly exciting to us. [79:31] ANGIE BROWN: We're really excited about Magic Apron [79:34] Shopping Assistant. [79:36] Our relationship with Google has really made this possible [79:38] in a variety of different ways. [79:40] If you just focus for a second on Gemini Enterprise [79:43] and you think about how it's the bedrock of what we are doing [79:46] with Magic Apron, it gives us the power [79:49] and the speed in which we need in order to power [79:53] these great technologies. [79:54] JORDAN BROGGI: Now, with Gemini Enterprise, [79:56] we think that we can blend those channels because we'll [79:59] have an agentic back end. [80:01] Maybe most importantly, we see a higher conversion rate when they [80:05] leverage the tool versus when they don't. [80:08] When customers come into the store and they use Magic Apron, [80:10] typically it's about wayfinding. [80:12] It's about product knowledge. [80:13] It's about compatibility. [80:15] But they also use it at home on our website. [80:17] Those experiences should all get better-- easier [80:20] to be inspired, easier to find what I need, easier [80:23] to order it and get it, easier to get support after the fact. [80:26] ANGIE BROWN: The partnership really [80:27] keeps us at the forefront of building [80:29] great experiences for our customers and our associates. [80:35] [APPLAUSE] [80:41] CARRIE THARP: The same agentic approach [80:43] is also reshaping digital-first industries. [80:47] One example of this is Reliance. [80:49] Reliance is transforming retail for millions in India [80:53] with a shopping agent. [80:55] Customers can simply type "plan a birthday party," [80:58] and the agent guides their journey across categories [81:01] and assembles their cart, which drives more revenue per visit. [81:06] Gemini-powered catalog enrichment [81:09] supports this, analyzing millions of images [81:12] in minutes instead of months. [81:15] Gemini Enterprise for Customer Experience [81:17] is hard at work across many industries and enterprises. [81:22] Internally, teams at Google, including Google Store, [81:26] Google Photos, Google One, and YouTube, [81:30] are partnering with us to power our own scaled customer [81:34] experiences. [81:35] Let's take a look at how one of our first internal customers [81:39] is doing this. [81:40] Over to you, Patrick. [81:42] [UPBEAT MUSIC] [81:51] PATRICK MARLOW: All right. [81:52] Thanks, Carrie. [81:53] Hey, y'all. [81:54] I'm Patrick Marlow, a senior product manager [81:56] here at Google Cloud. [81:57] YouTube TV customer support recently launched a voice agent [82:00] to help users with NFL Sunday Ticket and YouTube TV plan [82:04] subscriptions. [82:05] It's currently live in production, [82:07] serving 100% of their users. [82:09] We're going to take a look at it in action today. [82:11] I'm going to call the support number now. [82:18] SPEAKER 4: Thanks for calling the YouTube TV and NFL Sunday [82:21] Ticket Support Line. [82:22] So tell me, how can I help you today? [82:24] PATRICK MARLOW: Hey yeah, I've got [82:25] a bunch of my friends coming over tomorrow night for the NFL [82:27] draft. [82:28] I don't have cable. [82:29] And to be honest, I really only watch sports. [82:31] Does YouTube TV have a plan that's just for that? [82:37] SPEAKER 4: Yes, we do. [82:38] The YouTube TV Sports Plan has exactly what [82:41] you're looking for, plus over 30 other sports channels. [82:45] It's $18 less per month than the base plan. [82:48] Would you like me to text you a link to the number [82:50] you're calling from that will take you directly-- [82:52] PATRICK MARLOW: Yeah, go ahead and send me a link. [82:53] I'll take a look at it. [83:00] SPEAKER 4: Great, I just sent the link to your mobile phone. [83:02] Is there anything else I can help with today? [83:05] PATRICK MARLOW: Yeah, I'm going to take a look at this now. [83:07] Actually, my father-in-law is hanging out with me here today. [83:10] He's been thinking about getting the sports plan for his account [83:12] as well. [83:13] I just wanted to see what he thinks about this. [83:15] [SPEAKING SPANISH] [83:19] Actually, could you just quickly summarize this plan for him [83:21] in Spanish? [83:22] I think he just wants to know that he can get both football e [83:25] football. [83:28] SPEAKER 4: [SPEAKING SPANISH] [83:42] PATRICK MARLOW: Look, one other question. [83:44] I was planning on setting up multiple TVs [83:46] throughout the house, like one in the living room, [83:48] one over on the patio. [83:49] Would that be OK with this plan? [83:53] SPEAKER 4: Definitely, with the Sports Plan, [83:55] you can stream on up to three screens at the same time. [83:58] Is there anything else? [83:59] PATRICK MARLOW: Awesome. [84:00] No, look, I'm actually in the middle [84:02] of a stadium full of people right now, [84:03] and I think there's millions watching me on the live stream. [84:06] So I kind of got to run. [84:07] All right. [84:07] You've been helpful. [84:08] Later. [84:11] [APPLAUSE] [84:16] All right. [84:17] Now, that was pretty easy. [84:19] The agent was able to handle complex product logic [84:21] and seamlessly pivoted languages. [84:23] But what happens if the team wants to make a change? [84:26] They don't need a bunch of software engineers updating [84:29] endless lines of code. [84:30] They just need CX Agent Studio. [84:33] Here, we have our Visual Builder, inside [84:35] of CX Agent Studio, which is designed [84:37] to give you complete transparency [84:39] and control over your entire agent-building experience. [84:42] This is the behind-the-scenes look [84:44] at the actual YouTube TV's agent that they put in production. [84:47] The orchestration is managed across [84:49] multiple specialized agents, which can handle even [84:52] the most complex requests. [84:54] And with our built-in test interface, [84:56] you can ensure that every answer is grounded, factual, and pulled [85:01] directly from a knowledge base, like the Agents Price Finder [85:04] tool here. [85:05] And the best part? [85:06] YouTube TV customer support built [85:09] and deployed this entire experience in just six weeks. [85:13] Now, let's say that the team wants to run a promotion. [85:16] It's really simple for me to add a new subagent here. [85:19] Click Add a New Subagent. [85:21] And I will make this our new promotions agent. [85:26] Say promotions. [85:28] I'll hit Save. [85:29] And then I'm going to quickly add [85:31] a set of business logic via instructions in natural language [85:36] to our agent. [85:37] I'll hit Create. [85:39] And just like that, our multi-agent system [85:42] has instantly adapted to this new specialized agent. [85:45] It's really that simple. [85:47] Finally, with conversational insights, [85:50] the YouTube TV team can instantly [85:51] see how these calls are performing, [85:53] all within the same platform. [85:55] Here's one of my test calls from earlier this morning. [85:58] Now, let's recap. [86:00] We just turned a routine support call [86:02] into a multilingual upgrade and a globally deployed promotion, [86:06] all in under four minutes. [86:08] This is the power of Gemini Enterprise [86:10] for Customer Experience. [86:12] It gives you enterprise-grade reasoning, deep tool [86:15] integration, a powerful Visual Builder, and insights [86:18] to iterate at the speed of your business. [86:21] We can't wait to see what you will build next. [86:25] Thank you. [86:26] [UPBEAT MUSIC, APPLAUSE] [86:30] SPEAKER 1: Please welcome VP of product, [86:32] Google Workspace, Yulie Kwon Kim. [86:36] [UPBEAT MUSIC] [86:50] YULIE KWON KIM: Google Workspace is the world's most popular [86:53] productivity tool. [86:54] But even with the best tools at our fingertips, [86:58] work can still feel incredibly fragmented. [87:01] We've all been there. [87:03] You're trying to answer one simple question, [87:05] and 10 minutes later, you have 15 tabs open. [87:09] You're jumping from a stale email [87:11] to a slide deck that's being edited as you watch [87:13] and digging through a spreadsheet [87:15] just to find a quick answer. [87:18] We all spend half our day finding information [87:21] and the other half figuring out what to do with it. [87:25] What if you could skip both? [87:29] Today, we're introducing Workspace Intelligence. [87:33] [APPLAUSE] [87:37] With the advanced reasoning capabilities of Gemini [87:40] and state-of-the-art embedding models, [87:42] we're eliminating context fragmentation across [87:45] the Workspace suite. [87:48] Think of it as a unified intelligence [87:50] layer that lives inside every Workspace app. [87:53] It connects the dots and lets AI do the heavy lifting. [87:57] Let's see it in action. [88:00] [UPBEAT MUSIC] [88:10] So remember that furniture rebrand you saw earlier? [88:14] Let's say I'm the regional distributor, [88:16] and here I am in Google Chat. [88:19] This is where I collaborate with my colleagues and now my agents, [88:22] all in one place. [88:24] I'm here in the Regional Managers chat space. [88:27] And look at this. [88:28] The Regional Operations Agent, which [88:31] we built in Gemini Enterprise, just [88:34] alerted the team that our new display kits are arriving early. [88:38] And as you can see, I have multiple chats blowing up. [88:44] What do I need to do next? [88:46] With Ask Gemini, Google Chat becomes my Command Center. [88:53] Look here. [88:54] Gemini tells me exactly what matters right now. [88:58] See? [89:01] It surfaced an urgent task. [89:02] It linked the pitch deck that I need to localize, [89:05] and it flagged my 4:00 PM deadline for the regional plan. [89:09] The gathering is done. [89:10] And I haven't opened a single extra tab. [89:14] Now, to build this plan, I remember [89:17] we had a great chart that showed regional sales last quarter-- [89:22] somewhere. [89:23] Normally, this is where the hunt begins. [89:26] I jump into a folder and stare at dozens of files. [89:30] Instead, I'll ask Gemini. [89:34] Find the merchandising playbook from last quarter, [89:38] the one with the chart showing regional sales. [89:41] This isn't just keyword matching. [89:43] Workspace Intelligence understands [89:45] the context of my meetings and the content inside my files. [89:50] Here it is. [89:51] It pointed me directly to the dock with the exact graph [89:55] I need-- [89:56] a needle in a haystack, solved in seconds. [90:00] Now I need to turn this scattered data [90:03] into a regional plan. [90:05] I have a standard format for briefing store managers, [90:08] and with Ask Gemini, I basically say just do it for me. [90:12] Watch this. [90:15] Let's see, use the regional campaign skill to-- oops, [90:25] create a deck with a plan for organic living. [90:29] By tagging that specific skill, Workspace Intelligence [90:33] goes to work. [90:34] It cross-references my emails, my chats, my other docs. [90:38] It pulls live HubSpot win-loss data. [90:40] It uses our corporate branding, all to generate a new Google [90:44] slide deck. [90:46] It's hard at work pulling that deck together. [90:49] And here it is. [90:51] And in the citations up here, you [90:53] can see all the sources Workspace Intelligence pulled in [90:57] without me having to ask-- [90:59] the emails, the chats, the other files. [91:03] Should we take a look? [91:06] Yeah, look at this. [91:09] It's beautiful. [91:11] It matches my brand, and it's consistent with how [91:14] I've structured all of my regional plan decks in the past. [91:17] And it looks like it was designed by a creative pro, not [91:22] a terminal. [91:23] I'm ready to brief my team. [91:26] Workspace Intelligence is the end of the context tax. [91:30] It transforms how you work by turning fragmented information [91:34] into a clear path forward. [91:36] No complicated setup. [91:38] It's secure. [91:39] It's integrated. [91:40] It just works. [91:43] By moving from simple automation to full-scale AI orchestration, [91:47] industry leaders are using Workspace to fundamentally [91:50] redefine global productivity. [91:53] For example, customers like Colgate-Palmolive rolled out [91:57] Google Workspace to 34,000 employees. [92:00] AI agents built on Gemini helped their teams drive innovation, [92:04] turning data into new product concepts in minutes, not months. [92:09] And they drive business growth by unlocking actionable insights [92:12] from decades of sales history, empowering their teams [92:16] to help brighten millions of smiles every morning. [92:20] At Natura, custom agents are accelerating data-driven [92:24] reporting by 10x, while at Korean Air, [92:28] over 22,000 global employees use AI agents and tools [92:32] to accelerate high-impact operational and care tasks. [92:36] And at Compass Real Estate, Gemini [92:39] is managing tasks for employees so they [92:41] can focus on valuable client relationships. [92:45] We know many companies want to make the shift to Workspace [92:49] but worry about the complexity of migration. [92:52] I'm thrilled to announce that migrating [92:55] your entire organization, including complex and finance [92:59] teams from Microsoft 365 to Google Workspace, [93:03] is now up to five times faster, thanks to our new migration [93:08] and interoperability enhancements. [93:12] Today, you've seen what it looks like to fundamentally change [93:16] how your employees operate. [93:19] Go ahead, close those 15 tabs. [93:22] Workspace Intelligence is now ready to do [93:25] the heavy lifting for you. [93:27] Thomas, over to you. [93:29] [UPBEAT MUSIC, APPLAUSE] [93:43] THOMAS KURIAN: Wasn't that amazing? [93:45] [CHEERING, APPLAUSE] [93:50] Thank you, Yulie. [93:52] Before we close, let's take a look [93:54] at one of the most respected brands in the world. [93:57] Serving 3.7 billion people around the world, [94:01] Unilever chose Google Cloud to build and deploy agents [94:05] at scale. [94:06] Let's hear their story. [94:10] LEANDRO BARRETO: At Unilever, we serve 3.7 billion people [94:13] every day with our products. [94:15] We need to understand these individuals at a unique level, [94:19] and that's what the partnership with Google Cloud [94:22] allows us to do. [94:23] It's about personalization at scale, desire at scale, [94:28] and that's very unique. [94:29] And it's at the core of our strategy. [94:33] SAM KINI: We are connecting our business end-to-end [94:35] for faster AI deployment across the enterprise. [94:38] This means using AI to supercharge demand generation, [94:42] get early access to advanced AI and large language models, [94:45] and build even stronger partnerships with retailers, [94:47] connecting data in smarter ways so we can grow together. [94:50] Underpinning all of this is a strong digital backbone [94:54] powered by Google Cloud. [94:57] DAPHNE COATES: So my role is to operationalize [94:59] this really exciting vision, co-creating that with Google [95:02] to deliver our agentic foundations so [95:04] that our agents can be architected for performance, [95:08] scale, security, and observability right [95:10] from the outset. [95:12] The competitive buying multiagentic solution, [95:14] co-created with our Horizon3 Labs, [95:16] is leveraging core Google Cloud technologies such as Gemini [95:20] Enterprise and ADK, your Agent Development Kit, [95:22] and the underlying Gemini models. [95:25] So this solution orchestrates a multitude of agents [95:28] together through one single user interface. [95:30] This helps my procurement colleagues [95:32] make analysis and decisions in minutes instead of days. [95:36] This solution has enabled us to make smarter sourcing decisions. [95:40] SAM KINI: With Google Cloud, we are turning intelligence [95:42] into a growth engine. [95:47] [APPLAUSE] [95:53] THOMAS KURIAN: There's one final difference [95:55] in how we work with you. [95:56] We believe the future of AI must be open. [96:00] While others want to lock you into a wall garden that [96:03] owns your models, your data, and your agents, [96:06] we offer you an integrated stack, [96:09] but the freedom to choose the world's best chips and models, [96:13] the freedom to run AI wherever your data may live, [96:17] the freedom to control your own destiny with deep governance [96:20] features. [96:21] We scale this mission through our partners, [96:25] with whom we're building a broad and deep network [96:28] of forward-deployed engineers, including Accenture, BCG, [96:33] Deloitte, and McKinsey, who've announced [96:35] major expansion of their Google Gemini AI practices. [96:40] Along with AI-led service partners like Quantium, Distil [96:45] and Tribe.AI, we're helping independent software vendors [96:49] and SaaS companies transform their solutions with Gemini [96:53] Enterprise Agent Platform. [96:55] And we're bringing AI to small to medium-sized businesses [97:00] by helping them adopt Gemini Enterprise and our AI [97:04] advances in Workspace that work so seamlessly together now. [97:09] We've covered a lot of ground today. [97:11] We've introduced our agentic blueprint [97:13] as a foundation for agents to transform your companies-- [97:18] first, the AI Hypercomputer, the purpose-built foundation [97:22] optimized for the physics of the agentic era; second, [97:26] the Agentic Data Cloud, the engine that provides agents with [97:30] trusted business context; Agentic Defense, [97:34] the autonomous protection that secures your entire AI [97:38] lifecycle; fourth, the Gemini Enterprise Agent Platform, [97:43] a new agentic platform with state-of-the-art models [97:47] to build, deploy, and manage agents; and finally, [97:50] the Agentic Task Force, the specialized agents that are [97:53] already transforming customer experience and employee [97:56] productivity and transforming workflows all served up to users [98:02] to the Gemini Enterprise Application. [98:05] To each of our customers and partners, [98:08] thank you, each of you, for being on this amazing journey [98:12] with us. [98:14] Our teams have worked so hard this last year [98:17] to prepare this amazing technology for you, [98:21] and we're all so proud of the work they've done. [98:24] This platform is ready. [98:27] So what will each of you build? [98:30] We have three fantastic days planned for you, [98:33] and a great set of sessions and spotlights still to come. [98:38] Thank you and have an amazing Cloud Next. [98:42] [CHEERING, APPLAUSE] [98:45] [MUSIC PLAYING]