Dell, Nvidia Target Agentic AI Workloads With Modular Hardware Architecture

Right-sized Infrastructure for Agentic AI Workloads With NVIDIA And Dell Technologies — Ai4

Ai4

Autonomous AI reasoning loops create unpredictable token consumption that makes cloud hosting cost-prohibitive, driving enterprise workloads onto on-premises infrastructure to cut total spend by up to 87 percent.

AI-assisted summary of Right-sized Infrastructure for Agentic AI Workloads With NVIDIA And Dell Technologies — original source. Check the original for full context.

Source published: 2026-08-17

Detailed notes

- [00:04](https://www.youtube.com/watch?v=7gjrjgsAt94&t=4s) Enterprise AI infrastructure requires holistic optimization across all layers simultaneously—including compute hardware, software frameworks, storage, and specialized services—rather than focusing on isolated hardware specs. ## Key takeaways - [01:23](https://www.youtube.com/watch?v=7gjrjgsAt94&t=83s) The transition from predictable, linear chatbot prompts to persistent, reasoning autonomous agents shifts task execution runtimes from seconds to unbounded periods ranging from minutes to weeks. - [03:27](https://www.youtube.com/watch?v=7gjrjgsAt94&t=207s) Despite falling per-token market costs, the continuous reasoning loops of autonomous agents generate exponential volume growth, transforming AI execution into an ongoing, high-cost consumption meter. - [05:17](https://www.youtube.com/watch?v=7gjrjgsAt94&t=317s) NVIDIA Agent Architecture provides a unified software stack combining NVIDIA Neotron open-source models, CUDA libraries, and the OpenShell policy framework to enforce networking, security, and data privacy guardrails. - [08:11](https://www.youtube.com/watch?v=7gjrjgsAt94&t=491s) Developers are advised to start agentic development locally on deskside workstations to control early costs and test models before scaling to production environments. - [08:50](https://www.youtube.com/watch?v=7gjrjgsAt94&t=530s) Dell Technologies offers a tiered workstation portfolio ranging from entry-level GB10 units up to GB300-powered Precision T6 systems with 748 GB of memory for running massive 1-trillion parameter models locally. - [11:10](https://www.youtube.com/watch?v=7gjrjgsAt94&t=670s) The Dell AI Factory with NVIDIA provides modular infrastructure that allows compute, storage, and networking to scale independently from deskside units to rack-scale systems and hybrid clouds. - [13:02](https://www.youtube.com/watch?v=7gjrjgsAt94&t=782s) Independent analysis by Futurum demonstrates that shifting agentic AI workloads on-premises can reduce long-term operational spend by up to 87% compared to cloud token fees, with local workstations achieving payback within 3 months. - [15:56](https://www.youtube.com/watch?v=7gjrjgsAt94&t=956s) The Dell AI Factory automation framework integrates pre-validated ISV blueprints onto PowerEdge and NVIDIA hardware to rapidly deploy diverse enterprise workloads across software development, computer vision, and supply chain management.

## Detailed walkthrough Here is a detailed, faithful walkthrough of the transcript section.

---

[00:04](https://www.youtube.com/watch?v=7gjrjgsAt94&t=4s) The Agentic Inflection Point and Shift in Inference

Sama Bali (Senior Manager, AI Solutions Product Marketing at NVIDIA) opens the presentation alongside Robert (Rob) Hartman (Product Marketing Manager, AI Factory with NVIDIA at Dell Technologies). They outline an agenda covering the full spectrum of infrastructure required to run agentic workflows—including accelerated compute, software stacks, services, and storage systems. Bali emphasizes that when architecting infrastructure for agentic AI, organizations must optimize across all of these layers holistically rather than focusing on just one layer.

The speakers trace the evolution of AI interactions to illustrate why infrastructure demands are shifting radically:

``` [Chatbot Era] [Agentic Era] User Prompt User Request │ │ ▼ ▼ Linear Model Execution ──► Persistent Loop: Think -> Plan -> Reason -> Act │ │ ▼ ▼ Predictable Output Unpredictable Multi-Step Execution (Seconds / Cents) (Seconds to Weeks / Scale Token Usage) ```

In the earlier chatbot era, interactions followed a simple, linear model: a user asked a question, the LLM returned a response, runtime lasted a few seconds, and cost was predictable (a few cents). That era is closing in favor of persistent, self-learning, and self-improving autonomous agents. These agents do not merely answer questions; they plan, reason, and act in an ongoing execution loop until a task is fully resolved or manually stopped by a human operator. Consequently, runtimes are no longer bounded by seconds—they can stretch from minutes and hours to weeks, creating entirely unpredictable inferencing behavior.

- **Speakers:** Sama Bali (NVIDIA) and Robert Hartman (Dell Technologies). - **Key Shift:** Transition from simple prompt-and-response chatbots to persistent, reasoning autonomous agents. - **Runtime Variability:** Agent task execution times vary dramatically, spanning from seconds to weeks.

---

[03:27](https://www.youtube.com/watch?v=7gjrjgsAt94&t=207s) The Uncomfortable Reality of "Tokenomics" and P&L Impact

Bali addresses a common misconception among IT managers: the belief that AI agents are inexpensive. While individual token prices may be falling across the industry, overall AI spending is escalating rapidly. Deploying autonomous agents shifts the cost paradigm from running a discrete model to running a continuous meter. Every internal thought process, reasoning iteration, planning step, and tool invocation consumes tokens.

Even as unit costs per token drop, the exponential increase in token volume driven by continuous agent loops results in significantly higher total consumption. Managing these "tokenomics" and controlling P&L impact ultimately falls on the shoulders of system architects, who must design end-to-end infrastructure with token efficiency in mind.

- **Cost Dynamics:** Decreasing price per token is offset by massive increases in total token consumption. - **Consumption Model:** Agent loops act as continuous meters, generating tokens during every phase of planning, reasoning, and acting. - **Architectural Responsibility:** Architecture teams, rather than just IT managers or finance departments, must design infrastructure that controls token production costs.

---

[05:17](https://www.youtube.com/watch?v=7gjrjgsAt94&t=317s) NVIDIA Agent Toolkit and the OpenShell Policy Framework

To address the challenge of managing and securing autonomous agents across hybrid environments, NVIDIA provides the **NVIDIA Agent Toolkit**. This open-source platform serves as a unified software architecture for building, customizing, controlling, and deploying agents. It integrates seamlessly with existing enterprise agents, data systems, and developer tools, leveraging open-source models (such as NVIDIA Neotron models) alongside CUDA acceleration libraries.

To counter risks of agents going "off the rails," the framework incorporates an **OpenShell** policy engine### Introduction and Session Agenda

Sama Bali, Senior Manager of AI Solutions Product Marketing at NVIDIA, opens the session alongside Robert (Rob) Hartman, Product Marketing Manager for the AI Factory with NVIDIA at Dell Technologies. Highlighting the deep co-innovation between Dell and NVIDIA, the presenters outline a comprehensive walkthrough covering the full operational stack required for agentic AI workloads—spanning accelerated compute hardware, software frameworks, storage systems, and specialized services. They emphasize that designing infrastructure for agentic AI requires holistic optimization across all these layers simultaneously rather than focusing on isolated hardware specs.

``` +------------------------------------------------------------------+ | AGENTIC WORKLOAD STACK | | | | [ Services & Strategy ] --> [ Software Stacks & Guardrails ] | | | | | v | | [ Storage & Data Systems ] <-- [ Accelerated Compute Hardware ] | +------------------------------------------------------------------+ ```

The agenda covers the fundamental shift toward agentic AI, the complex financial realities of "tokenomics" on organizational P&L, right-sizing compute infrastructure for specific agent behaviors, and an overview of the joint Dell AI Factory with NVIDIA ecosystem.

- **Speakers:** Sama Bali (Sr. Manager, AI Solutions Product Marketing, NVIDIA) & Robert Hartman (Product Marketing Manager, AI Factory with NVIDIA, Dell Technologies). - **Core Theme:** Right-sizing infrastructure to support agentic AI without overbuilding budget or stalling deployment roadmaps.

---

[01:23](https://www.youtube.com/watch?v=7gjrjgsAt94&t=83s) The Agentic Inflection Point: From Chatbots to Autonomous Agents

Sama Bali contrasts the fading era of basic conversational chatbots with the emerging paradigm of autonomous, persistent AI agents. Early chatbot interactions followed a simple, linear path: a user submitted a discrete prompt, the model generated a single completion, and the process concluded within seconds for a predictable, nominal cost. In contrast, modern agentic AI systems exhibit persistence, continuous reasoning, planning, and tool execution—operating in self-improving loops to fulfill complex goals without step-by-step human intervention.

``` CHATBOT ERA (Linear & Predictable): [User Prompt] ---> [LLM Generation] ---> [Response] (Runtime: Seconds)

AGENTIC ERA (Iterative & Dynamic Loop): +----------------------------------+ | | v | [Task Goal] -> [Reason/Plan] -> [Action/Tool] -> [Observe Outcome] | | +--- Continues until Complete -----+ (Runtime: Seconds to Weeks) ```

This structural shift transforms the operational and compute dynamics of AI inference. Because an agent continuously iterates, evaluates its own progress, and calls external tools until a task is successfully resolved or manually halted, runtime durations become completely variable. A single agent task may take seconds, minutes, hours, or even weeks to complete.

- **Chatbot Era:** Linear prompt-and-response model with predictable token usage and execution times measured in seconds. - **Agentic Era:** Autonomous, persistent, self-learning loops that reason, plan, and act independently. - **Execution Runtime:** Dynamic and unbounded—ranging from seconds to weeks depending on task complexity and execution criteria.

---

[03:27](https://www.youtube.com/watch?v=7gjrjgsAt94&t=207s) The Uncomfortable Reality of Tokenomics and P&L Impact

Addressing a common misconception among IT executives, Bali highlights why agentic AI costs can escalate rapidly. While IT managers often assume individual agent runs are inexpensive, agentic architectures effectively convert model execution into an ongoing "consumption meter." Every internal thought step, intermediate planning sequence, context evaluation, and tool invocation consumes tokens.

``` +------------------------------------------------------------------+ | THE TOKENOMICS TRAP | | | | Unit Token Price Drops (📉) x Token Usage Volume (📈📈📈) | | | | = D dramatic Total Spend Increase (💰💰) | +------------------------------------------------------------------+ ```

Although market prices per individual token continue to drop, total consumption volumes are growing exponentially due to continuous reasoning loops. As a result, overall AI expenditures rise significantly. IT, finance, and enterprise architecture teams must transition from viewing AI as a static software expense to managing it as a dynamic, high-volume consumption model, requiring specialized software tools to maintain architectural control.

- **The "Metered" Model:** Every internal action, reasoning phase, and intermediate plan incurs a continuous token cost. - **Price vs. Consumption:** Falling per-token market prices are outpaced by massive increases in overall token volume per task. - **Architectural Responsibility:** Enterprise architects must take ownership of token management and infrastructure design to prevent budget overruns.

---

[05:17](https://www.youtube.com/watch?v=7gjrjgsAt94&t=317s) NVIDIA Agent Architecture & OpenShell Security Guardrails

To address the management and governance challenges of agentic workloads, NVIDIA introduced the NVIDIA Agent Architecture (part of the NVIDIA Agent Toolkit). This unified framework serves as a centralized solution for constructing, customizing, and controlling AI agents. Designed to integrate smoothly with an organization's existing enterprise tools, proprietary data platforms, and third-party agentic frameworks, the open-source platform leverages NVIDIA Neotron models alongside domain-optimized CUDA libraries.

``` +------------------------------------------------------------------+ | NVIDIA AGENT ARCHITECTURE | | | | +----------------------------------------------------------+ | | | OpenShell Security & Governance Policy | | | | (Networking, Security, and Data Privacy) | | | +----------------------------------------------------------+ | | | NVIDIA Agent Toolkit & CUDA-X Libraries | | | +----------------------------------------------------------+ | | | AI Data Platform (Enterprise Context/RAG) | | | +----------------------------------------------------------+ | | | Open-Source & Neotron LLMs / Reasoning Models | | | +----------------------------------------------------------+ | | | | [ Workstation / Desk ] <--> [ Data Center ] <--> [ Cloud ] | +------------------------------------------------------------------+ ```

A critical layer within this architecture is the "OpenShell" policy framework, which establishes robust guardrails to prevent autonomous agents from going off-rail. OpenShell allows enterprise administrators to explicitly enforce strict networking, security, and data privacy policies directly around the agent's runtime environment. Furthermore, because the entire software stack is uniform across all execution targets, developers can build and test an agent on a local workstation and seamlessly deploy it to an on-premises data center or public cloud environment without refactoring code.

- **Unified Open-Source Stack:** Combines NVIDIA Neotron open-source models, CUDA acceleration libraries, and enterprise AI data platforms. - **OpenShell Policy Framework:** Enforces strict execution guardrails across networking, security, and data privacy to maintain containment. - **Cross-Platform Uniformity:** Standardized API and runtime layers allow identical code to run seamlessly across workstations, edge devices, data centers, and public clouds.

---

[08:11](https://www.youtube.com/watch?v=7gjrjgsAt94&t=491s) Deskside Compute Portfolio for Agentic Prototyping

Bali advises enterprise teams to initiate agentic development directly at the developer's desk rather than immediately provisioning expensive, unconstrained cloud or data center clusters. To facilitate this local-first approach, Dell Technologies and NVIDIA offer a dedicated portfolio of workstation compute tiers designed to match specific developer personas and model parameter requirements.

The hardware lineup begins with ultra-compact, entry-level developer systems and scales up to high-capacity research workstations:

- **Dell Pro Max with GB10:** An ultra-compact device designed to fit in the palm of a hand. Equipped with 128 GB of unified compute memory, it supports localized model execution for models up to 200 billion parameters at FP4 precision. - **Dell Precision T2 Tower:** Targeted at non-developer end users who require a local desktop engine to run daily agentic workflows using ~200B parameter models. - **Dell Precision T4 Tower:** A mainstream professional workstation providing mid-tier compute for local agent development and regional team testing. - **Dell Precision T6 & GB300:** High-end systems featuring the GB300 superchip. Equipped with 748 GB of memory, these systems can host massive models up to 1 trillion parameters locally, enabling advanced research teams to build, test, and validate agentic architectures before deploying them to larger production infrastructure.

---

[11:10](https://www.youtube.com/watch?v=7gjrjgsAt94&t=670s) Modular Infrastructure: Dell AI Factory with NVIDIA

When localized prototype agents are ready for enterprise-wide production, workloads must transition smoothly to centralized infrastructure. The Dell AI Factory with NVIDIA offers a modular, scalable pathway spanning deskside units, rack-scale appliances, and customizable enterprise data center architectures.

``` +------------------------------------------------------------------+ | DELL AI FACTORY WITH NVIDIA ARCHITECTURE | | | | [ Deskside Workstations ] (Pro Max GB10, Precision T2/T4/T6) | | | | | v | | [ Enterprise PowerEdge Servers ] (Modular Data Center Nodes) | | | | | v | | [ PowerRack Systems ] (High-Density Accelerated Compute) | | | | | v | | [ Hybrid Cloud Options ] (Public & Edge Deployment Targets) | +------------------------------------------------------------------+ ```

This flexible deployment model lets organizations choose the optimal hosting platform based on economic and operational requirements. Development teams can use cost-effective local models and open-source tooling for early-stage experimentation, move shared workloads onto enterprise on-premises server racks, or leverage cloud environments for peak frontier-level intelligence requirements.

- **PowerEdge & PowerRack Integration:** Dense, compute-optimized server rack configurations designed for continuous, high-concurrency agentic workloads. - **Modular Foundations:** Independent scaling of compute, networking, and storage components to match changing inferencing demands. - **Workload Routing:** Gives IT architects full flexibility to choose where each workload runs—balancing local deskside systems, on-premises racks, and public cloud infrastructure.

---

[13:02](https://www.youtube.com/watch?v=7gjrjgsAt94&t=782s) On-Premises ROI and the Economics of Agent Containment

Taking over the presentation, Robert Hartman details the economic advantages of bringing agentic AI workloads on-premises. As organizations expand from initial pilot projects to multi-department deployments—spanning coding assistants, computer vision pipelines, and supply chain managers—uncontained cloud token fees can severely strain enterprise budgets.

To quantify these cost dynamics, Dell contracted third-party analyst firm **Futurum** to conduct an independent financial study comparing cloud-hosted token consumption against on-premises infrastructure across three enterprise AI use cases over a multi-year timeline:

``` +------------------------------------------------------------------+ | FUTURUM ECONOMIC STUDY HIGHLIGHTS | | | | - Payback Period (Deskside Setup): Achievable within 3 months | | - Payback Period (Server Clusters): Even faster ROI timeline | | - Long-Term TCO Spend Reduction: Up to 87% vs. Cloud Tokens | +------------------------------------------------------------------+ ```

Hartman explains that unmanaged AI agents operating in the cloud can quickly consume token budgets as they run autonomous loops. By containing and operating these digital agents behind the enterprise firewall on dedicated local infrastructure, organizations maintain direct control over costs, safeguard proprietary data, and achieve a predictable total cost of ownership.

- **3-Month Payback Period:** Futurum's study shows that deskside agentic development setups can achieve full ROI in as little as 3 months. - **87% Reduction in Spend:** Deploying on-premises infrastructure reduced total operational AI costs by up to 87% compared to relying exclusively on cloud token services. - **Budget Predictability:** On-premises hardware stabilizes financial planning by eliminating volatile, usage-based cloud token charges.

---

[15:56](https://www.youtube.com/watch?v=7gjrjgsAt94&t=956s) End-to-End Enterprise Automation Framework

Hartman outlines the structural components of the Dell AI Factory automation platform, illustrating how modular hardware building blocks transition into an integrated enterprise solution. Anchored by Dell PowerEdge hardware and NVIDIA acceleration, the system incorporates a specialized software automation suite alongside an ecosystem of Independent Software Vendor (ISV) partners.

``` +------------------------------------------------------------------+ | DELL AI FACTORY AUTOMATION STACK | | | | +----------------------------------------------------------+ | | | USE CASES: Chatbots, Coding, Vision, Supply Chain Agents | | | +----------------------------------------------------------+ | | | ISV BLUEPRINTS & AUTOMATION: Orchestration & ISV Partner | | | | Ecosystem | | | +----------------------------------------------------------+ | | | HARDWARE FOUNDATION: Dell PowerEdge & NVIDIA Accelerators| | | +----------------------------------------------------------+ | +------------------------------------------------------------------+ ```

This integrated stack uses pre-validated ISV software blueprints to rapidly launch tailored AI workloads across diverse business units—including automated customer service assistants, operational agent managers, developer coding aids, and visual inspection systems. By standardizing on a single underlying architecture, enterprise IT teams can run multiple distinct workloads concurrently, applying shared operational learnings and consistent governance policies across the entire organization.

- **Dell PowerEdge Foundation:** Flexible, high-performance compute nodes serve as the core hardware building blocks. - **ISV Blueprint Ecosystem:** Pre-integrated partner software enables rapid deployment of targeted enterprise workflows. - **Multi-Workload Support:** Runs diverse agentic, generative, and computer vision workloads on a single, shared infrastructure stack.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=7gjrjgsAt94&t=0s) Overbuilding wastes budget. Underbuilding stalls your roadmap.

> [00:00](https://www.youtube.com/watch?v=7gjrjgsAt94&t=0s) This session breaks down how the modular architecture of the Dell AI Factory with NVIDIA lets developers start right-sized and expand compute, storage, or networking independently as workloads grow.

> [00:04](https://www.youtube.com/watch?v=7gjrjgsAt94&t=4s) When architecting infrastructure for agentic AI, organizations must optimize across all of these layers holistically rather than focusing on just one layer. — Sama Bali

> [01:23](https://www.youtube.com/watch?v=7gjrjgsAt94&t=83s) These agents do not merely answer questions; they plan, reason, and act in an ongoing execution loop until a task is fully resolved or manually stopped by a human operator. — Sama Bali

> [03:27](https://www.youtube.com/watch?v=7gjrjgsAt94&t=207s) Deploying autonomous agents shifts the cost paradigm from running a discrete model to running a continuous meter. — Sama Bali

> [03:27](https://www.youtube.com/watch?v=7gjrjgsAt94&t=207s) Even as unit costs per token drop, the exponential increase in token volume driven by continuous agent loops results in significantly higher total consumption. — Sama Bali

> [05:17](https://www.youtube.com/watch?v=7gjrjgsAt94&t=317s) OpenShell allows enterprise administrators to explicitly enforce strict networking, security, and data privacy policies directly around the agent's runtime environment. — Sama Bali

> [08:11](https://www.youtube.com/watch?v=7gjrjgsAt94&t=491s) Initiate agentic development directly at the developer's desk rather than immediately provisioning expensive, unconstrained cloud or data center clusters. — Sama Bali

> [13:02](https://www.youtube.com/watch?v=7gjrjgsAt94&t=782s) By containing and operating these digital agents behind the enterprise firewall on dedicated local infrastructure, organizations maintain direct control over costs, safeguard proprietary data, and achieve a predictable total cost of ownership. — Robert Hartman

## Mentioned Sama Bali, Robert Hartman, NVIDIA, Dell Technologies, Ai4, Dell AI Factory with NVIDIA, NVIDIA Agent Toolkit, OpenShell, NVIDIA Neotron, CUDA, Dell Pro Max, GB10, Dell Precision T2 Tower, Dell Precision T4 Tower, Dell Precision T6, GB300, Dell PowerEdge, PowerRack, Futurum

## Video description Sama Bali - Sr Manager, AI Solutions Product Marketing, NVIDIA Robert Hartman - Product Marketing Manager, AI Factory with NVIDIA, Dell Technologies

Overbuilding wastes budget. Underbuilding stalls your roadmap. This session breaks down how the modular architecture of the Dell AI Factory with NVIDIA lets developers start right-sized and expand compute, storage, or networking independently as workloads grow. We'll connect this flexibility to the realities of agentic and distributed inferencing, where demand can shift fast and unpredictably. You'll learn how to match infrastructure to actual workloads and keep your path to scale predictable.

[Source](https://www.youtube.com/watch?v=7gjrjgsAt94)

中文简报

戴尔与英伟达以模块化硬件架构瞄准智能体AI工作负载自主AI推理循环带来不可预测的Token消耗，导致云托管成本高企，正促使企业工作负载转向本地基础设施，以削减高达87%的总支出。从可预测的线性聊天机器人转向具备推理能力的自主智能体，使任务执行时间大幅延长，持续推理循环引发的Token用量剧增也导致AI执行成本居高不下。对此，企业AI基础设施需要在计算硬件、软件框架及存储等全层面上进行协同优化。NVIDIA Agent Architecture通过整合Neotron开源模型、CUDA库及OpenShell策略框架构建了统一软件栈以保障安全与隐私，相关建议亦指出开发者在扩展至生产环境前应先在本地桌面工作站开展开发与测试，以控制初期成本。从简单的聊天机器人转变为具备持续推理能力的自主 Agent 后，AI 任务的执行时间从几秒延长至数分钟甚至数周，使得推理需求和计算负载变得极为不可预测。 (1:23) 尽管单 Token 市场单价持续走低，但自主 Agent 的持续思考循环导致 Token 消耗量呈指数级暴增，使 AI 部署转变为持续计费的高成本模式，极大考验企业损益（P&L）。 (3:27) NVIDIA 推出 Agent Toolkit 统一软件架构，结合 Neotron 开源模型、CUDA 库以及 OpenShell 策略框架，对 Agent 的网络、安全与数据隐私设立严格的护栏隔离。 (5:17) Dell 与 NVIDIA 合作提供从掌上型 GB10（128GB 内存）到搭载 GB300 的 Precision T6（748GB 内存、可本地运行万亿参数模型）的分层工作站算力方案，建议开发者先在本地端侧进行原型开发以控制成本。 (8:11) 第三方机构 Futurum 的研究显示，相较于云端 Token 费用，将 Agent 工作负载迁移至本地基础设施最高可降低 87% 的长期运营支出，本地工作站配置最快可在 3 个月内收回投资。 (13:02)

Reading summary

Sama Bali of NVIDIA and Robert Hartman of Dell Technologies argue that autonomous AI agents fundamentally alter inference economics through continuous reasoning loops, requiring organizations to adopt right-sized, on-premises infrastructure to curb escalating token expenses. The Operational Shift to Autonomous Agent Loops (1:23) The transition from the chatbot era to agentic AI fundamentally alters computational behavior. Conventional conversational chatbots executed a predictable, linear path: a user submitted a prompt, the model returned a completion within seconds, and costs remained nominal. In contrast, autonomous agents operate in continuous, self-improving execution loops characterized by planning, reasoning, acting, and observing outcomes until a task finishes or is halted by an operator. According to Bali, this iterative operational model makes inference runtimes dynamic and unbounded. Rather than concluding in a few seconds, agentic workflows can run for minutes, hours, or even weeks. This unpredictability prevents traditional capacity planning from adequately sizing infrastructure for multi-step reasoning. The Tokenomics Trap and Metered Consumption (3:27) Bali highlights a prevalent misconception among IT leaders: assuming AI agents are inexpensive simply because unit costs per token continue to drop across the industry. In practice, agentic architectures convert model execution into an ongoing, metered consumption process where every internal reasoning cycle, intermediate planning step, and tool invocation consumes tokens. Because the volume of tokens generated within autonomous loops grows exponentially, total enterprise expenditures increase significantly despite falling unit prices. As a result, Bali stresses that enterprise and system architects—rather than just IT managers or finance departments—must assume direct responsibility for token efficiency by designing infrastructure that actively bounds consumption. Software Governance and the OpenShell Policy Engine (5:17) To address operational governance and prevent autonomous agents from running off-rail, NVIDIA introduced the NVIDIA Agent Toolkit within the NVIDIA Agent Architecture. This open-source software stack integrates NVIDIA Neotron models, CUDA acceleration libraries, and enterprise context data platforms to provide a uniform development environment. A central component of this architecture is the OpenShell policy framework, which enforces strict network, security, and data privacy guardrails directly around the agent runtime. Bali notes that this uniform software stack functions identically across local workstations, data center servers, and cloud instances, allowing developers to build and test agents without having to refactor code across deployment targets. Tiered Infrastructure from Deskside to Data Center (8:11) Bali advises organizations to begin agent development locally on specialized deskside hardware rather than provisioning expensive cloud clusters immediately. The hardware portfolio spans compact systems like the Dell Pro Max with GB10 (128 GB memory, supporting models up to 200 billion parameters at FP4 precision) and Dell Precision T2 and T4 towers for local teams, up to the Dell Precision T6 equipped with the GB300 superchip (748 GB memory, supporting models up to 1 trillion parameters). When prototype agents reach production maturity, workloads scale into the modular Dell AI Factory with NVIDIA. This architecture deploys Dell PowerEdge servers and PowerRack clusters, allowing enterprise IT teams to scale compute, networking, and storage components independently to balance localized workstations, on-premises racks, and hybrid cloud capacity. On-Premises Cost Containment and ROI Validation (13:02) Hartman states that uncontained agents running on third-party cloud infrastructure expose organizations to severe budget volatility. Placing agents behind enterprise firewalls on dedicated hardware gives organizations predictability over operational budgets while safeguarding proprietary corporate data. To substantiate the financial argument for on-premises deployment, Dell commissioned an economic study by analyst firm Futurum comparing cloud token consumption against on-premises infrastructure across three enterprise use cases. Futurum reported that deskside workstation configurations achieved a full return on investment within three months—with server clusters showing an even faster payback timeline—while reducing long-term total cost of ownership by up to 87% compared to exclusive reliance on cloud tokens. While vendor data shows that on-premises, modular infrastructure can cut long-term token costs by up to 87%, successfully capturing these economic gains depends on enterprise architects enforcing rigorous execution guardrails and choosing appropriate compute tiers before agents enter unconstrained production loops.

精炼阅读摘要

NVIDIA 的 Sama Bali 与 Dell Technologies 的 Robert Hartman 指出，AI 正在从传统的线性聊天机器人演进为具备持续循环能力的自主智能体，这种转变导致计算推理时间变得高度不可预测，并引发“代币经济学”（Tokenomics）成本激增。两位讲者主张，企业必须跳出孤立的硬件规格思维，通过覆盖桌面工作站、模块化数据中心、统一软件防护栏及本地化部署的端到端架构，实现智能体工作负载的合理规模配置与成本控制。智能体推理范式转移与执行周期变革 (0:04) Bali 阐述了人工智能交互模式的基础性转变：早期的聊天机器人遵循简单的线性运行模式，用户提交提示词后，大语言模型在数秒内即可生成一次性回复，运行时长确定且单次调用成本仅为数美分。与之不同，新兴的自主智能体呈现出持久化、自我学习与自主推理的特征，其运行机制是在“思考（Think）—规划（Plan）—推理（Reason）—行动（Act）”的闭环中持续循环，在没有人类逐项干预的情况下自主调用外部工具推进任务，直至目标彻底达成或被人工中止。这种结构性变革彻底重构了算力基础设施的推理运行动态。由于智能体需要自主评估执行进度并展开多阶段迭代，单次任务的计算周期不再受限于数秒之内，而是可能从数秒、数分钟延伸至数小时乃至数周。推理行为因此演变为高度不可预测且动态伸缩的过程，给传统基于确定性负载构建的静态算力基础设施带来了根本性挑战。代币消耗激增带来的企业损益表冲击 (3:27) Bali 提示企业 IT 管理者警惕一个普遍误区，即误以为运行 AI 智能体的费用低廉。尽管当前整个技术行业的单代币（Per-token）市场价格持续下降，但企业在 AI 领域的总体支出却在快速攀升。自主智能体的架构实质上将离散的模型调用转变成了一个持续计费的“消费计量表”（Metered Model），智能体在内部思维发散、制定中间规划、评估上下文环境以及调用外部工具的每一个操作，都会实时产生代币消耗。这种单代币价格下降与任务内总代币用量呈指数级暴增并存的现象构成了严峻的“代币经济学”陷阱，导致企业总体支出急剧膨胀。讲者强调，管控代币消耗及其对企业损益表（P&L）财务冲击的职责，已经无法仅依靠 IT 采购或财务部门，企业系统与架构设计团队必须主动介入，在设计端到端基础设施时全面纳入代币生产效率指标，通过架构层面的优化避免预算失控。统一软件栈与 OpenShell 治理框架 (5:17) 为应对自主智能体在混合IT环境下的管理与治理难题，NVIDIA 推出了 NVIDIA Agent Architecture（隶属于 NVIDIA Agent Toolkit）。该开源软件架构旨在为构建、定制、控制和部署智能体提供集中化底座，不仅整合了 NVIDIA Neotron 开源推理模型与底层 CUDA 加速库，还支持与企业现有的专有数据平台、检索增强生成（RAG）框架及第三方开发工具深度打通。由于该软件栈在开发端到生产端保持统一，开发者在本地工作站编写测试的代码，可无缝迁移至本地数据中心或公有云环境执行，无需重构底层代码。为防范自主智能体在运行中出现偏离预定行为轨道（Off-rail）的风险，该体系集成了核心的“OpenShell”策略框架。OpenShell 允许系统管理员直接在智能体的运行时环境周围部署强制性的防护栏，涵盖网络准入控制、系统安全权限以及数据隐私边界规范。通过确定性的治理规则，企业能够确保智能体在被赋予自主规划与行动能力的同时，各项操作始终受到合规约束与隔离保护。从桌面到数据中心的分级算力布局 (8:11) 针对算力选型，Bali 建议企业团队在探索智能体时应坚持“桌面先行”的原则，直接在开发者的本地工位启动研发，避免过早向高成本、无约束的云端或数据中心集群申请资源。Dell 与 NVIDIA 针对不同开发角色与模型规模推出了对应的桌面算力分级产品：巴掌大小的 Dell Pro Max 搭载 GB10 芯片，具备 128 GB 统一计算内存，支持在 FP4 精度下本地运行最高 2000 亿（200B）参数的模型；Dell Precision T2 与 T4 塔式工作站分别满足非开发业务人员的日常智能体运行与主流团队开发；高端 Precision T6 则搭载 GB300 超级芯片，配备 748 GB 内存，支持研究团队在本地构建和验证最高达 1 万亿（1T）参数的庞大模型。当原型测试成熟并进入企业级规模化落地阶段后，工作负载可通过 Dell AI Factory with NVIDIA 方案平滑过渡到集中化基础设施。该方案采用高密度的 Dell PowerEdge 服务器与 PowerRack 机架系统，支持计算、网络和存储模块的独立弹性扩展。这种模块化架构赋予了 IT 架构师灵活的工作负载路由能力，能够根据不同任务的经济性要求，在低成本本地模型、企业专有机架集群与云端前沿智能资源之间动态调度。本地化部署的投资回报与自动化生态 (13:02) Hartman 进一步剖析了将智能体工作负载收敛至企业防火墙内（On-Premises）运行的财务与运营优势。随着智能体在客服助理、代码编写、计算机视觉分析和供应链管理等多个核心业务部门的广泛铺开，持续调用云端代币服务可能给企业带来不可承受的资金负担。根据 Dell 委托独立分析机构 Futurum 开展的跨年度多场景财务评估，相比于纯云端代币消费模式，本地桌面部署的投资回收期最短仅需 3 个月，机架式服务器集群的回收周期则更为迅速；长期来看，本地化基础设施可将总体拥有成本（TCO）降低多达 87%，并有效消除用量波动带来的计费风险。在落地支撑层面，Dell AI Factory 自动化平台以 Dell PowerEdge 硬件底座和 NVIDIA 加速器为依托，结合了经过预先验证的独立软件供应商（ISV）蓝图。这一集成的全栈方案使得企业 IT 团队能够在共享的底层软硬件资源上，安全并发地支撑自动化客户服务、内部研发辅助与工业视觉巡检等异构工作负载，在实现全企业统一治理的同时，保持资产投入的高确定性与可控性。两位讲者共同论证了自主智能体在执行周期与代币消耗上的结构性剧变，提出了“桌面原型—机架生产—受控治理”的软硬件协同应对方案。需要指出的是，关于投资回报率（如 3 个月回收期与 87% 的 TCO 降幅）的核心经济论据主要源自 Dell 委托 Futurum 完成的特定商业分析报告；在企业实际落地中，对于耗时数天乃至数周的极端长周期任务，本地硬件并发承载的物理上限，以及相较于云端最新前沿大模型能力演进的长期权衡，仍是架构师在规模化部署前需要审慎评估的技术变量。

Detailed notes — zh

## Background 该内容由 Ai4 于 2026 年 8 月 17 日发布在 YouTube 平台上。视频由 NVIDIA 的 Sama Bali 与戴尔科技的 Robert Hartman 共同主讲，聚焦于如何利用 Dell AI Factory 与 NVIDIA 的模块化架构，为快速变化且不可预测的智能体（Agentic）与分布式推理工作负载精准匹配算力，实现平滑、可预测的规模化扩展。 ## TL;DR 本场演讲深入剖析了从传统线性对话机器人向自主闭环智能体演进过程中的推理拐点，并揭示了自主循环导致的“Token 经济学”对企业财务损益表（P&L）的隐性冲击。两位讲者提出了“始于桌面、渐进扩展”的算力定型策略，展示了从 Dell Pro Max、Precision 工作站到 PowerEdge 机架的完整硬件矩阵与 NVIDIA Agent Toolkit 软件栈。独立研究表明，通过 Dell AI Factory 本地化与混合架构部署，企业不仅能在 3 个月内收回硬件投资，更可最高削减 87% 的云端 Token 支出并实现数据与资产的完全自主可控。 ## Atmosphere & style 整场分享呈现出务实、严谨且极具说服力的企业级技术发布风格。两位讲者配合默契，将前沿的架构设计理念与犀利的财务成本账本紧密结合，打破了业界对“Token 降价即降本”的乐观幻觉。现场叙事节奏清晰紧凑，既有对技术失控风险的敏锐洞察，又给出了清晰落地、可落账的系统级破局方案，让听众在感受到生成式 AI 落地挑战的同时，重获对算力预算与架构扩展的掌控感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=7gjrjgsAt94&t=4s) 构建智能体工作负载的基础设施必须采取全栈协同视角，全局优化加速计算、软件栈、数据存储与专业服务，而非孤立地看待单一层级。 - [01:30](https://www.youtube.com/watch?v=7gjrjgsAt94&t=90s) 智能体通过思考、规划、工具调用和行动形成闭环，其运行周期从传统单次问答的数秒延长至数分钟甚至数周，彻底打破了传统的线性算力预测模型。 - [03:27](https://www.youtube.com/watch?v=7gjrjgsAt94&t=207s) 智能体的自主循环引发了 Token 消耗总量的爆炸式增长，即使单个 Token 价格持续下降，企业总账单仍会激增，Token 治理已成为架构设计的核心工程问题。 - [05:17](https://www.youtube.com/watch?v=7gjrjgsAt94&t=317s) NVIDIA 提供 Agent Toolkit 作为统一软件平台，并结合开源 OpenShell 安全沙箱框架，为智能体设定网络通信、数据隐私与行为安全护栏。 - [08:53](https://www.youtube.com/watch?v=7gjrjgsAt94&t=533s) 企业构建与验证智能体应优先从桌面端起步，利用统一的 NVIDIA 软件栈在本地开源模型上验证逻辑，规避公有云高昂的初期调用账单。 - [08:53](https://www.youtube.com/watch?v=7gjrjgsAt94&t=533s) 戴尔与 NVIDIA 推出覆盖从掌上型 Dell Pro Max（GB10）到 Precision T6（GB300）的桌面算力矩阵，本地最大可支持运行 1 万亿（1T）参数模型。 - [11:29](https://www.youtube.com/watch?v=7gjrjgsAt94&t=689s) Dell AI Factory 采用解耦模块化架构，支持企业构建桌面本地单机、机架共享算力池与公有云弹性结合的多层算力分流体系。 - [13:02](https://www.youtube.com/watch?v=7gjrjgsAt94&t=782s) 第三方研究机构 Futurum 数据显示，在本地桌面端部署智能体最快可在采购后第 3 个月实现全额投资回报（ROI），服务器部署回报周期更短。 - [13:02](https://www.youtube.com/watch?v=7gjrjgsAt94&t=782s) 相比公有云持续消耗的 Token 模式，基于 Dell AI Factory 的本地化方案可实现高达 87% 的整体支出削减，同时确保数据与模型权重的完全自主掌控。 - [15:56](https://www.youtube.com/watch?v=7gjrjgsAt94&t=956s) 戴尔提供基于 PowerEdge 的模块化硬件基石与自动化软件平台，结合 ISV 验证蓝图，在统一物理架构上并行承载代码助手、运营分析与视觉智能体等多类工作负载。 - [18:37](https://www.youtube.com/watch?v=7gjrjgsAt94&t=1117s) 戴尔与 NVIDIA 结合专业服务团队与加速课程，帮助企业从桌面端轻量试点平滑过渡并扩展至企业级生产部署。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=7gjrjgsAt94&t=4s) 欢迎开场与核心议程：构建智能体工作负载的端到端架构来自 NVIDIA 的 AI 解决方案产品营销负责人 Sama Bali 与戴尔科技（Dell Technologies）AI Factory 产品营销经理 Robert Hartman 共同登台。Sama 指出，构建支持智能体（Agentic）工作负载的基础设施必须采取全栈协同视角，从加速计算、软件栈、服务到存储系统进行全局优化，而不是孤立地看待某一个层级。两位演讲者明确了本次分享的核心脉络：从智能体技术拐点切入，剖析它对企业损益表（P&L）产生的隐性成本冲击（即 Token 经济学），并展示如何通过 NVIDIA 与戴尔联合打造的软硬件体系实现算力精准匹配与可预测扩展。 - **演讲阵容**：Sama Bali（NVIDIA AI 解决方案产品营销高级经理）与 Robert Hartman（戴尔科技产品营销经理）。 - **全栈覆盖维度**：加速计算硬件、全栈软件层、数据与存储系统、全流程专业服务。 - **核心议程**： 1. 智能体引发的推理拐点与工作流变革； 2. 智能体时代的 Token 经济学与财务影响； 3. 如何根据智能体负载实现算力“精准定型”（Right-Sized Compute）； 4. 戴尔与 NVIDIA 联合打造的 Dell AI Factory 架构与全生命周期经济模型。 --- ### [01:30](https://www.youtube.com/watch?v=7gjrjgsAt94&t=90s) 从传统聊天机器人到自主智能体：推理拐点的到来 Sama 回顾了从简单生成式 AI 向自主智能体演变的过程。在传统的对话机器人时代，交互模式是线性的单次问答——用户输入 Prompt，模型在数秒内返回响应，Token 消耗是静态且高度可预测的，单次交互成本仅为数美分。然而，这一时代正在迅速翻篇，企业正在跨入持久运行、具备自我学习与自我改进能力的自主智能体时代。智能体拥有思考、规划、工具调用、推理与行动（Reason, Plan, Act）的能力，能够代表人类执行完整的端到端业务流。由于智能体处于不间断的反馈闭环中，其运行时间不再是固定的几秒钟，而是可能持续数分钟、数小时甚至数周，直到任务彻底完成或人工干预终止。这彻底打破了传统算力与成本消耗的线性预测模型。 ```mermaid flowchart LR accTitle: 从单次对话到智能体循环推理的转变 subgraph Traditional["传统 Chatbot 模式 (单向线性)"] A["用户 Prompt 输入"] --> B["单次模型推理"] B --> C["即时生成响应"] end subgraph Agentic["自主智能体模式 (自主闭环)"] D["复杂业务目标"] --> E["规划与思考 (Plan/Reason)"] E --> F["调用工具与行动 (Act)"] F --> G{"任务是否完成?"} G -- 否 / 需进一步迭代 --> E G -- 是 --> H["交付最终成果"] end ``` --- ### [03:27](https://www.youtube.com/watch?v=7gjrjgsAt94&t=207s) Token 经济学与 P&L 财务冲击：为什么便宜的 Token 反而推高账单？ Sama 揭示了许多 IT 管理者在预算规划中的误区：“智能体并不昂贵——直到它们真正出现在你的财务损益表（P&L）上。”在智能体体系下，企业运行的不再仅仅是一个静态模型，而是一台“时刻计费的电表”。智能体进行的每一步规划、反思、重试和外部交互，都在以 Token 的形式飞速计费。针对业界普遍存在的“Token 降价将导致 AI 总成本下降”的观点，Sama 明确予以反驳：这是一种典型的消费模式陷阱。虽然单位 Token 价格在持续走低，但智能体自主循环导致的 Token 消耗总量呈指数级爆发，企业收到的最终云账单反而水涨船高。这使得 Token 管理不再只是财务或传统 IT 的被动报销问题，而是必须由架构团队在顶层设计中解决的系统级工程。 - **计费模式质变**：从离散的单次请求计费演变为持续运行的流式消耗（Consumption Model）。 - **运行时间跨度**：从传统 Chatbot 的数秒级扩展至智能体的分钟级、小时级甚至数周级。 - **财务矛盾点**：单价下降（Token Price Down）与使用量暴增（Usage Up）相互叠加，导致企业总体拥有成本（TCO）激增。 - **责任归属**：从单纯由 IT/财务监控成本，转变为架构设计团队（Architect Team）主导的基础设施全生命周期 Token 治理。 --- ### [05:17](https://www.youtube.com/watch?v=7gjrjgsAt94&t=317s) NVIDIA 智能体工具包与 OpenShell 运行时安全护栏为解决架构碎片化与成本失控问题，NVIDIA 提供了统一的软件架构平台——NVIDIA Agent Toolkit。该平台充当构建、部署、管控和定制智能体的“一站式枢纽”，不仅能接入企业现有的数据资产和工具库，还全面兼容开源大模型（包括 NVIDIA 自研的 Nemotron 系列）以及加速全行业应用的 CUDA 算子库。此外，内置的 AI 数据平台为智能体注入了理解企业业务语境所需的上下文。针对企业普遍担心的“智能体失控乱跑”风险，Sama 重点介绍了开源的安全防护框架 OpenShell。通过 OpenShell 策略引擎，企业可以自定义严格的网络通信、系统安全及数据隐私规则，将智能体的运行时环境完全置于可信边界内，防止非预期行为的发生。 ```mermaid flowchart TD accTitle: NVIDIA 智能体软件架构与部署目标 subgraph SoftwareStack["NVIDIA 智能体核心软件栈"] M["开源与自研模型 (Nemotron / OSS LLMs)"] C["CUDA 底层加速库"] D["企业 AI 数据平台 (业务上下文)"] S["OpenShell 安全沙箱与防护策略 (网络/隐私/安全)"] end subgraph DeploymentTargets["一致性部署目标 (跨全场景无缝迁移)"] T1["桌面工作站 (Desk-side Workstations)"] T2["边缘设备 (Edge Devices)"] T3["企业本地数据中心 (On-Prem Data Centers)"] T4["公有云环境 (Cloud)"] end SoftwareStack --> DeploymentTargets ``` --- ### [08:53](https://www.youtube.com/watch?v=7gjrjgsAt94&t=533s) 从桌面起步：戴尔与 NVIDIA 联合推出的桌面级算力矩阵 Sama 提出建议：企业构建与验证智能体应优先“始于桌面（Start at your desk）”，在本地利用开源工具和本地模型验证业务逻辑，规避公有云高昂的 Token 计费。戴尔与 NVIDIA 联合推出了覆盖不同人员角色与负载规模的三层桌面算力矩阵，且均运行相同的 NVIDIA 软件栈，确保成果能够平滑迁移至数据中心或云端。 ```mermaid flowchart LR accTitle: 戴尔与 NVIDIA 桌面端智能体算力矩阵 subgraph Tier1["入门与探索 (Explore)"] A1["Dell Pro Max (搭载 GB10) • 掌心大小，开发者即时算力 • 128 GB 统一内存 • 支持 200B 参数模型 (FP4)"] A2["Dell Precision T2 塔式工作站 • 面向业务最终用户 / 非开发者 • 支持最高 200B 参数模型"] end subgraph Tier2["专业进阶 (Professional)"] B1["Dell Precision T4 塔式工作站 • 专业级工位部署 • 紧凑高效，支持多类型本地智能体"] end subgraph Tier3["大规模原型 (Massive / Extreme)"] C1["Dell Precision T6 & GB300 方案 • GB10 的高性能进阶版本 • 748 GB 超大内存容量 • 支持运行最高 1 万亿 (1T) 参数模型"] end ``` - **Dell Pro Max 搭配 NVIDIA GB10**：掌心大小的超紧凑形态，配备 128 GB 内存/算力池，可在 FP4 精度下本地运行高达 2000 亿（200B）参数的大模型，专为需要即时本地算力的开发者打造。 - **Dell Precision T2 塔式工作站**：面向日常业务人员及非技术岗，提供本地运行 ~200B 参数模型的能力。 - **Dell Precision T4 塔式工作站**：定位专业级工作站，机身紧凑，适配桌底放置，可承载中高复杂度的本地智能体流。 - **Dell Precision T6 搭配 GB300**：被称为 GB10 的“巨无霸兄弟”，配备高达 748 GB 显存/内存容量，能够直接在本地单机运行高达 1 万亿（1T）参数的前沿模型，供科研人员和核心开发团队进行大型智能体集群的原型验证与深度调优。 --- ### [11:29](https://www.youtube.com/watch?v=7gjrjgsAt94&t=689s) 戴尔 AI Factory 模块化演进：无缝迈向企业级生产当智能体在桌面端完成原型开发并准备毕业（Graduate）投入全天候生产时，企业需要系统性的算力托底。Sama 介绍了覆盖从桌面端系统、Dell PowerRack 密集计算平台到完全模块化设计的 Dell AI Factory with NVIDIA。该方案允许企业根据不同业务场景灵活决策：日常低频任务使用本地开源模型规避 Token 计费；部门级协同任务迁移至共享私有算力池；极致前沿智能场景则对接公有云。统一的软件框架确保代码与策略无需重写即可全域流通。 - **多层算力分流逻辑**： 1. **桌面端单机运行**：利用开源模型处理敏感或高频任务，彻底免除 Token 账单； 2. **本地集群共享**：利用 Dell PowerEdge 与 PowerRack 承载高并发智能体集群； 3. **云端弹性扩展**：按需调用超大规模前沿模型服务。 - **全模块化架构**：算力、网络与存储模块相互解耦且深度优化，支持企业按业务增长逐步按需扩容。 --- ### [13:02](https://www.youtube.com/watch?v=7gjrjgsAt94&t=782s) 本地化部署的经济学账本：3 个月实现 ROI 与 87% 支出缩减戴尔科技的 Robert Hartman 接过话题，深入剖析了企业扩大智能体试点时的成本瓶颈。当企业从单个试点迈向跨部门部署（例如代码助手、项目管理、计算机视觉及供应链智能体）时，公有云的 Token 成本增长往往超出预期。戴尔委托独立分析机构 Futurum 进行了量化经济效益研究，模拟评估了企业在数年周期内运行 3 类典型智能体场景（项目辅助、编程助手、计算机视觉等）的成本回报。研究证实，在本地桌面端开启智能体项目，企业最快在硬件采购后的第 3 个月即可收回全部投资成本（ROI）；若部署至本地服务器机架，ROI 的实现速度甚至更快。 | 评估维度 | 公有云托管 Token 模式 | 戴尔与 NVIDIA 本地化 / AI Factory 方案 | |:--- |:--- |:--- | | **成本控制** | 持续消耗 Token，随着智能体自主循环费用不可控 | 固定资产投入，边际 Token 成本为零，支出削减最高达 **87%** | | **投资回报周期 (ROI)** | 持续性经营费用（OpEx），无硬件回本概念 | 桌面端最快 **3 个月** 回本，服务器端回收期更短 | | **数据与模型所有权** | 数据出境、依赖第三方 API 协议 | 模型权重、私有数据、硬件资产 **100% 自主掌控** | | **安全管控边界** | 需依赖云端外部安全规则 | 智能体在企业内部防火墙内闭环运行，杜绝 Token 预算“野蛮生长” | --- ### [15:56](https://www.youtube.com/watch?v=7gjrjgsAt94&t=956s) 自动化平台与模块化基石：加速智能体从试点到规模化落地 Robert 进一步拆解了 Dell AI Factory 的落地实现路径。戴尔并非要求客户一开始就采购整套昂贵的大型机架与网络设备，而是提供以 Dell PowerEdge 服务器为核心的灵活构建模块（Building Blocks）。在底层硬件之上，戴尔构建了专门的自动化软件平台，并与独立软件开发商（ISV）生态深度集成，提供经过预先验证的“架构蓝图（Blueprints）”。无论企业是要落地代码助手、运营分析智能体、计算机视觉系统还是传统对话机器人，该架构都能利用统一的环境和学习路径支撑多团队、多负载并行运转。 Robert 总结指出，本地化 AI Factory 方案能够帮助企业实现最高 87% 的支出缩减，将原本浪费在失控云端 Token 上的预算重新收回企业内部，为智能化转型构筑可持续的经济壁垒。 - **基础设施解耦构建**：基于 Dell PowerEdge 与 NVIDIA 算力单元按需搭建。 - **自动化平台软件栈**：提供开箱即用的自动化部署工具与 ISV 合作伙伴蓝图。 - **多业务工作负载统合**：在同一套物理基础架构上无缝承载 Chatbot、智能体运营编排、代码协同与视觉分析等多类任务。 - **生产就绪路径**：提供端到端验证方案，消除技术选型试错成本，帮助企业快速跨越从 PoC 试点到全天候生产上线的“死亡之谷”。 ### [18:37](https://www.youtube.com/watch?v=7gjrjgsAt94&t=1117s) 落地路径：从桌面级试点到 Dell 与 NVIDIA 联合赋能的大规模扩展在规划 Agentic AI 的落地路径时，企业面临的首要问题往往是“该从何处起步构建”。针对这一需求，Dell 与 NVIDIA 提供了极其灵活的起点选择——开发团队既可以通过基于 NVIDIA 解决方案的桌面级（Desk-side）工作站方式轻量起步，也可以直接采用经过官方验证的解决方案（Validated Solutions）进行概念验证（PoC）与试点部署。为了帮助企业扫清落地障碍，Dell 配备了专门的加速课程（Accelerator Course）以及专业服务团队（Services Teams），全程协助客户梳理运营流程、评估最佳切入点、明确具体的业务用例，并确保基础设施能够切实交付业务价值，从而平滑迁移至完整的生产级解决方案中。 ```mermaid flowchart LR accTitle: Agentic AI 落地与扩展路径 subgraph S1["1. 起步探索与试点"] A["桌面级工作站试点 Desk-side Approach"] B["Dell 验证解决方案 Validated Solutions"] end subgraph S2["2. 软件堆栈与服务赋能"] C["NVIDIA NGC 蓝图库与软件栈"] D["Dell 自动化平台与标准化蓝图"] E["Dell 专属加速课程与服务团队"] end subgraph S3["3. 生产级扩展"] F["Dell AI Factory with NVIDIA 弹性扩展基础设施"] end A --> E B --> E E --> C E --> D C --> F D --> F ``` 在软件栈与生态支撑层面，开发者拥有充足的选择空间。团队可以直接调用 NVIDIA NGC 平台上提供的全套蓝图（Blueprints）、参考架构及优化软件堆栈，以极高的敏捷度快速、灵活地构建 Agentic 系统；亦或是借助 Dell 的自动化平台与标准化蓝图，走稳健规范的企业级部署路径。正如讲者所言，“世界尽在你的掌握之中”（"the world is the oyster here"），在 Dell AI Factory 与 NVIDIA 坚实底层技术的支撑下，企业探索智能体（Agentic AI）有着极高的容错率与多维选择。在演讲尾声，讲者热情邀请现场观众前往 1307 号展位，与 Sama 和 Robert 展开更深入的一对一交流，共同探讨具体的业务场景需求与架构运营细节。 - **多元化的起步模式**：支持从轻量级桌面端（Desk-side）开发环境起步，亦可直接基于 Dell 验证解决方案（Validated Solutions）搭建试点项目。 - **全流程专业服务支持**：Dell 专业服务团队提供加速课程，协助识别高价值 AI 用例并设计清晰的落地与迁移路径。 - **软硬件协同生态**：全面打通 NVIDIA NGC 软件资产库（含多模态蓝图与软件栈）与 Dell 自动化部署平台，兼顾灵活性与标准化需求。 - **现场交流对接**：讲者邀请参会者前往 1307 展位进行深入技术咨询与定制化架构探讨。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=7gjrjgsAt94&t=4s) 构建支持智能体（Agentic）工作负载的基础设施必须采取全栈协同视角，从加速计算、软件栈、服务到存储系统进行全局优化，而不是孤立地看待某一个层级。 — Sama Bali > [01:30](https://www.youtube.com/watch?v=7gjrjgsAt94&t=90s) 由于智能体处于不间断的反馈闭环中，其运行时间不再是固定的几秒钟，而是可能持续数分钟、数小时甚至数周，直到任务彻底完成或人工干预终止。这彻底打破了传统算力与成本消耗的线性预测模型。 — Sama Bali > [03:27](https://www.youtube.com/watch?v=7gjrjgsAt94&t=207s) 智能体并不昂贵——直到它们真正出现在你的财务损益表（P&L）上。 — Sama Bali > [03:27](https://www.youtube.com/watch?v=7gjrjgsAt94&t=207s) 虽然单位 Token 价格在持续走低，但智能体自主循环导致的 Token 消耗总量呈指数级爆发，企业收到的最终云账单反而水涨船高。 — Sama Bali > [08:53](https://www.youtube.com/watch?v=7gjrjgsAt94&t=533s) 企业构建与验证智能体应优先“始于桌面（Start at your desk）”，在本地利用开源工具和本地模型验证业务逻辑，规避公有云高昂的 Token 计费。 — Sama Bali > [13:02](https://www.youtube.com/watch?v=7gjrjgsAt94&t=782s) 在本地桌面端开启智能体项目，企业最快在硬件采购后的第 3 个月即可收回全部投资成本（ROI）；若部署至本地服务器机架，ROI 的实现速度甚至更快。 — Robert Hartman > [15:56](https://www.youtube.com/watch?v=7gjrjgsAt94&t=956s) 戴尔并非要求客户一开始就采购整套昂贵的大型机架与网络设备，而是提供以 Dell PowerEdge 服务器为核心的灵活构建模块（Building Blocks）。 — Robert Hartman > [15:56](https://www.youtube.com/watch?v=7gjrjgsAt94&t=956s) 本地化 AI Factory 方案能够帮助企业实现最高 87% 的支出缩减，将原本浪费在失控云端 Token 上的预算重新收回企业内部，为智能化转型构筑可持续的经济壁垒。 — Robert Hartman > [18:37](https://www.youtube.com/watch?v=7gjrjgsAt94&t=1117s) 世界尽在你的掌握之中（the world is the oyster here） — Robert Hartman ## Mentioned Sama Bali, Robert Hartman, NVIDIA, Dell Technologies, Dell AI Factory with NVIDIA, NVIDIA Agent Toolkit, OpenShell, Nemotron, CUDA, Dell Pro Max, NVIDIA GB10, Dell Precision T2, Dell Precision T4, Dell Precision T6, NVIDIA GB300, Dell PowerRack, Dell PowerEdge, Futurum, NVIDIA NGC ## Video description Sama Bali - Sr Manager, AI Solutions Product Marketing, NVIDIA Robert Hartman - Product Marketing Manager, AI Factory with NVIDIA, Dell Technologies Overbuilding wastes budget. Underbuilding stalls your roadmap. This session breaks down how the modular architecture of the Dell AI Factory with NVIDIA lets developers start right-sized and expand compute, storage, or networking independently as workloads grow. We'll connect this flexibility to the realities of agentic and distributed inferencing, where demand can shift fast and unpredictably. You'll learn how to match infrastructure to actual workloads and keep your path to scale predictable. [Source](https://www.youtube.com/watch?v=7gjrjgsAt94)

Source quotes

[0:04] Sama Bali: 构建支持智能体（Agentic）工作负载的基础设施必须采取全栈协同视角，从加速计算、软件栈、服务到存储系统进行全局优化，而不是孤立地看待某一个层级。 [1:30] Sama Bali: 由于智能体处于不间断的反馈闭环中，其运行时间不再是固定的几秒钟，而是可能持续数分钟、数小时甚至数周，直到任务彻底完成或人工干预终止。这彻底打破了传统算力与成本消耗的线性预测模型。 [3:27] Sama Bali: 智能体并不昂贵——直到它们真正出现在你的财务损益表（P&L）上。 [3:27] Sama Bali: 虽然单位 Token 价格在持续走低，但智能体自主循环导致的 Token 消耗总量呈指数级爆发，企业收到的最终云账单反而水涨船高。 [8:53] Sama Bali: 企业构建与验证智能体应优先“始于桌面（Start at your desk）”，在本地利用开源工具和本地模型验证业务逻辑，规避公有云高昂的 Token 计费。 [13:02] Robert Hartman: 在本地桌面端开启智能体项目，企业最快在硬件采购后的第 3 个月即可收回全部投资成本（ROI）；若部署至本地服务器机架，ROI 的实现速度甚至更快。 [15:56] Robert Hartman: 戴尔并非要求客户一开始就采购整套昂贵的大型机架与网络设备，而是提供以 Dell PowerEdge 服务器为核心的灵活构建模块（Building Blocks）。 [15:56] Robert Hartman: 本地化 AI Factory 方案能够帮助企业实现最高 87% 的支出缩减，将原本浪费在失控云端 Token 上的预算重新收回企业内部，为智能化转型构筑可持续的经济壁垒。 [18:37] Robert Hartman: 世界尽在你的掌握之中（the world is the oyster here）

Source quotes

[0:00] Overbuilding wastes budget. Underbuilding stalls your roadmap. [0:00] This session breaks down how the modular architecture of the Dell AI Factory with NVIDIA lets developers start right-sized and expand compute, storage, or networking independently as workloads grow. [0:04] Sama Bali: When architecting infrastructure for agentic AI, organizations must optimize across all of these layers holistically rather than focusing on just one layer. [1:23] Sama Bali: These agents do not merely answer questions; they plan, reason, and act in an ongoing execution loop until a task is fully resolved or manually stopped by a human operator. [3:27] Sama Bali: Deploying autonomous agents shifts the cost paradigm from running a discrete model to running a continuous meter. [3:27] Sama Bali: Even as unit costs per token drop, the exponential increase in token volume driven by continuous agent loops results in significantly higher total consumption. [5:17] Sama Bali: OpenShell allows enterprise administrators to explicitly enforce strict networking, security, and data privacy policies directly around the agent's runtime environment. [8:11] Sama Bali: Initiate agentic development directly at the developer's desk rather than immediately provisioning expensive, unconstrained cloud or data center clusters. [13:02] Robert Hartman: By containing and operating these digital agents behind the enterprise firewall on dedicated local infrastructure, organizations maintain direct control over costs, safeguard proprietary data, and achieve a predictable total cost of ownership.

Original transcript

[0:04] So I'm Sama Bali. I lead AI solutions product marketing at NVIDIA and I am going to be tagging [0:09] team with Rob. Rob, why don't you introduce yourself? Robert Hartman, product marketing at [0:13] Dell. It's great to be working with our closest partner in a lot of our developments and our [0:16] innovation going forward. She so get ready to get to a lot of agentic stuff today and I'll pass it [0:21] back to Sama. Thank you. So between Rob and I, we'll talk everything about from accelerated [0:27] compute to software stacks to services, storage systems that you need to be running agentic [0:34] workflows. And that is deliberate, right? Because you when you're thinking about architecting your [0:39] entire infrastructure for running these workflows, you want to make sure that you're optimizing for [0:45] all of these layers, not just one of them. So with that, that's going to be our agenda today. [0:51] We'll get started by talking about, you know, the agentic inflection point. What has changed? [0:56] How is it different? We'll then get into the uncomfortable area of talking about tokconomics, [1:01] right? How is it showing up on your P&L? From there, we'll talk about how you should [1:06] be thinking about making sure that you have the right sized compute based on the agentic workflow [1:12] that you're running. And then we'll introduce the Dell AI factory with Nvidia and then talk about [1:18] the economics, you know, just a mental framework of how you should be thinking about just managing [1:23] all of these workflows end to end. So let's set the scene here. Right here is our the workflow [1:30] that we're seeing with autonomous agents kind of going across the board. We started off with simple [1:36] chatbot error, right? You asked a question, it was a simple prompt, it gave you an answer, [1:42] you moved on. That era is now closing. We are now moving to the world of these autonomous agents. [1:49] These agents are persistent. They are self-arning. They are self-improving. Right? So they actually [1:56] have the ability to reason, act, plan, end up doing the entire workflow on your behalf as well. [2:03] Which gets us to this inference inflection point. Which means that initially when you [2:08] were using your chatbot, it was a very linear curve in terms of your tokens being used, right? [2:14] You asked a question, it gave you an answer. It's a very linear thing. It was predictable. [2:19] Now you have your agents here. They are the ones who are taking the action and they'll continue [2:25] to work on that workflow till it gets completed. They'll keep going on in that loop itself. So this [2:30] is where things get very interesting because now you there is no way to predict what is happening. [2:36] So if you say and I'm going to dig a little bit deeper here as well this is exactly what I'm [2:40] talking about right when you were working with an AI chatbot it was a simple world you asked [2:47] a question it gave you an answer the runtime was a few seconds itself you know you probably [2:52] uh it costed you a few cents right there now we're talking about agents wherein this agent [2:59] has the ability to think plan reason act it has this unending loop of doing that workflow. And [3:07] in some cases, your runtime can be seconds, it can be minutes, it can be hours, it can be weeks. Why? [3:14] Because your agent will continue to work on that task you asked it to finish till it knows it has [3:20] completed or you yourself intervene and stop that task right there. Right? So costs here become very [3:27] interesting because your runtime can be extremely long. Now what we keep hearing from a lot of IT [3:34] managers out there is that oh you know agents are not expensive and that is a true statement agents [3:40] are not expensive till they start showing up on your P&L and here is how that tends to happen [3:47] right you are no longer just running a model you are essentially running a meter which means that [3:55] as your agent continues to think it that meter starts right there for every action it's doing [4:02] including you know planning thinking each of that piece is a token being used right there [4:09] now I do also want to call out that a lot of times when we talk to customers they talk about oh you [4:14] know the token prices are going down so AI costs are going to go down that is not going to be true [4:20] you have to understand the token costs are going down but your usage is going up this is going to [4:28] be a consumption model, right? Because remember it's not a binary way here where you asked a [4:34] question, it gave an answer and it stopped. Here your act agent is acting on your behalf. [4:41] It's going to continue through that thinking loop again and again. So even if your tokens are cheap, [4:47] your consumption is dramatically high. So your total bill is still going to increase right there. [4:54] So now it comes down to, you know, who manages it? Is it going to be the IT people whose job [5:01] is to figure out how much this is going to cost? Is it going to be the finance team? [5:05] It ultimately will come down to your architect team who is thinking end to end about this entire [5:11] infrastructure. It is them who's actually going to have to think about how do we manage these tokens? [5:17] And to make that really easy for you, we have the Nvidia agent toolkit. [5:24] Now this is where we make it really easy for you to think about that one software architecture [5:30] right in terms of thinking about how do I architect my entire infrastructure. [5:35] So let's get started with this that Nvidia agent architecture itself. Now, think of this as [5:42] your one-stop shop for making agents, building agents, but also being able to control agents, [5:49] customize agents. If your organization already has existing agents, data tools, frameworks, [5:56] this has the ability to connect to all of it. Now, this is an open-source platform. It makes use of [6:03] literally any open-source models that you want to use, including Nvidia's own Neotron models. [6:08] It also makes use of our CUDA libraries. These are the same libraries that a lot of organizations [6:15] a lot I'm pretty sure all of you have been using apps that are accelerated by the same CUDA [6:19] libraries itself. The same libraries are available to you to accelerate your agents as well. [6:25] And then the last thing which I do want to call out here is that AI data platform. This [6:29] is what makes gives your agents the context that they need to understand your business as well. [6:37] The most important thing here that I do want to call out what this framework gives you is guard [6:41] rails. We keep talking about, we keep hearing stories about agents going off the rail doing [6:47] whatsoever they wanted. This is where with that open shell policy, we help you make [6:52] sure that your runtime is completely secure. Again, this is open source, so you can make [6:57] use of for any agent that you're trying to build altogether. The idea behind OpenShell is that you [7:04] get to decide the policies for networking, for security, for privacy. So you essentially define [7:11] how your agent is going to work. Now we're going to go back to my question in terms of [7:15] how you should be thinking about the architecture. And this is why this framework helps because uh [7:22] at the bottom of the slide you'll see you actually have the ability to take this entire framework [7:28] and build and deploy agents across any deployment targets. This can be cloud, this can be your data [7:34] center, this can even be edge devices or devices like workstations that are sitting at your desk. [7:41] So it's one architecture and then again the idea is to make it easy for you to think about [7:48] that where should my agents be running right every time that you're building an agent the question [7:54] should not be that this is the infrastructure here's where it runs now I need to figure out cost [7:58] the question should be backwards from thinking where does it make most sense to actually run my [8:04] agents because now I have a security software framework which runs across everything that [8:11] I have in my infrastructure itself. So for most organizations that I speak with, [8:17] my guidance is always that when you're thinking about building these agents, using these agents, [8:23] it should always start at your desk itself, right? And here with Dell, we've partnered to create a [8:29] full portfolio of devices which help you build, run, and deploy agents starting at your desk. [8:37] But I do want to remind you again, it's that same software framework. So if you're starting [8:42] at a desk with any of these devices, you have the opportunity to seamlessly scale it to the [8:47] data center to the cloud of your choice itself. It's that same architecture that runs through it. [8:53] Now let me walk you through this portfolio right here for your desk options. Right, we've got three [8:59] tiers here and again that is deliberate. We want to make sure that based on your agentic workload, [9:06] you have an entry point. So starting first, we'll get started with the left hand side which has the [9:12] explore options. Uh we've got the Dell Pro Max with GB10. As you can see, it's a tiny device [9:18] and I promise it's that small. It fits into the palm of my hands. I have small hands as all of you [9:24] can see. U it is a tiny but a very mighty device. This comes with 128 gigs of compute. You have the [9:32] ability to run models up to 200 billion parameters at FP4. This is designed purely for developers, [9:39] right? We want to make sure that developers have instant compute at their desk to get started. [9:45] And on the other side, in the same frame, you're seeing the Dell precision T2 tower. Now, this is [9:51] designed for your end users. So, any non-developer personas who are using agents for their workflows. [9:58] This is going to be that entry point device again to run a lot of this different kinds of agents [10:03] with about 200 billion parameter models. We keep moving to the next option and that's where we have [10:09] our Dell T4 tower. Again, this is where you're coming into the realm of professional workstations [10:15] to run different kinds of agents. Again, these are compact devices efficient enough to run a [10:21] lot of agents locally right below your desk itself. And in the extreme corner, you're seeing [10:28] your T6 and your GB300. Again, GB300 is, think of this as the big monstrous brother of GB10, [10:37] designed for developers to run massive agents. This has 748 gigs of memory, which means that you [10:45] can run models up to 1 trillion parameters. So, think of, you know, your researchers, your teams [10:52] of developers. This is going to be a great device to again think about prototyping, experimenting, [10:59] testing out that ideas before they actually want to graduate and move it to a bigger target [11:03] altogether. So these are all the options that we have for you to get started on the desk. I do [11:10] want to remind everybody again the same software stack that same open source software stack from [11:16] Nvidia which includes everything from tools to frameworks to SDKs runs. It starts here and helps [11:22] you gradually move to the full DEF factory with Nvidia. I'll quickly give you an overview of what [11:29] that looks like. Right? This is where we provide the entire breadth of solutions with Dell and [11:36] Nvidia. The idea again is based on your workload. When you are thinking about using these agentic u [11:44] agents in action, building these agents in action, here is that full set of options that you have. [11:51] based on that workload right that conversation should always happen where from an economic [11:55] sense does it make sense to actually run these agents can I run them locally today with local [12:00] models and open source tools so that you know there is no token generation problem altogether [12:05] or can I move it to another on-prem device where it can be shared across different teams [12:10] or do I definitely need that you know frontier level intelligence for specific use cases that [12:15] it needs to be in the cloud itself so here is that full stack with Nvidia we spoke about the deskside [12:21] systems altogether. I do want to call out that as you build these agents on these deskite systems, [12:26] they have to move somewhere once you graduate these agents for full-time production. [12:31] And that's where we have the full portfolio with the best-in-class Dell platforms. You also have [12:37] the power rack there. This is designed to host a lot of these dense computationally intensive [12:44] agents altogether. And then we have that fully modular architecture with Delactory with Nvidia. [12:50] Again, you've got all the components, all the pieces, each one of them optimized so that you [12:55] have this end toend solution to build and deploy these different kinds of agents. With that, Rob, [13:02] I'm going to hand it over to you. Thanks, Sama. So as s explained you know there's an architecture of [13:08] solutions that we have here for all kinds of agentic growth because your organization is [13:11] going to wonder you know get one agentic pilot up and running then wonder okay now how can I scale [13:17] this into other operations so this kind of shows you the full breadth as someone mentioned that [13:22] when you get to that point where you're starting to understand what are the economics of the [13:25] operations that I'm running how much it going to cost me and do I need to start bringing that on [13:29] premise do I start need to bring in now frontier models you know up to one trillion parameter [13:34] models on premise so I can decrease the cost to buy tokens. Those are some of the things to [13:39] consider as you start to build your organization, enable more of your workforce teams with the [13:44] agentic tools or with more generative AI or even look at other use cases that your organization may [13:50] be looking at for like computer vision uh chatbot assistance you know more and more agents across [13:55] the board to help with other areas of their business supply chain. This is why we have [13:59] a full-fledged portfolio. But in looking at the approach of thinking about on premise and thinking [14:05] about you know your development team starting at their desk with agentic development with the right [14:10] tools with a full NVIDIA portfolio of software, right? The full stack operation as we call it [14:16] and then understanding what's that going to cost me you know how do I need to do I need to buy into [14:20] a full rack of servers and networking. You don't have to but let's take a look at the economics. [14:25] So we contracted with a vendor called Futurum to have them look at what how much does it [14:31] cost to really understand to run three use cases over a protected projected couple of year terms [14:38] and come up with you know what the ROI is and they came up with some pretty compelling information [14:42] that says if you're starting at the desk side on an agentic project or three use cases project [14:48] assistant a coding assistance of sorts computer vision or other use case you can actually have an [14:54] ROI as soon is 3 months into the purchase cycle of what you just bought. It's even faster if you [15:00] deploy it fully on servers. The agreement there, the powerful compelling value proposition is that [15:06] being on premise gives you not only capability of owning your own equipment, owning your own models [15:13] on your own data, but it also gives you control of your cost. And that's what we're really driving at [15:18] here is that you might have seen with tokconomics and cloud operations, you're getting now to oh my [15:24] budget's got to now account for next year for next year's cloud costs. With an on-remise approach, [15:29] you now are reducing that budget. You're not maximizing. So it's a good approach [15:34] to help you understand that when you unleash agents into your workforce onto operations, [15:40] they are running wild. They're going to be running token budgets. But if you contain them, if you [15:44] manage them and operate them as onboard digital agents for your workforce inside the firewall, [15:49] there's a way to man help manage that and help them develop and do that cost effectively. [15:56] So what does it mean for you? So when you start to look at the larger implementations [16:01] of you know developing the AI factory we start you off with simple flexible building blocks but a lot [16:06] of you have seen with the Dell power edge and the Dell components and on top of that we build [16:11] in what we call the automation platform. We got a specialized software suite that helps you enable [16:15] with ecosystems of our ISV partners that bring in the blueprints these building blocks that [16:20] enable you to do different operations across the board whether it's a gentic whether it's you know [16:26] uh managed approaches whether it's SAS approaches that are all run on the AI factory with MIDI [16:31] architecture and that helps you power what we call the use cases these are where you want to decide [16:36] that a certain team needs a chatbot assistance another team needs an agentic operations manager [16:42] Another team needs a coding system or computer vision. This entire ecosystem is now running [16:48] and supporting an on your infrastructure with multiple workloads, multiple use cases [16:53] for other teams in your organization. So all this comes together and delivers [16:56] you a scalable approach to production that gets you from a pilot stage. It's kicking [17:01] off that agentic approach into a full production stage very quickly and easily. [17:06] So basically one architecture which is the AI factory with NVIDIA starting from the desk [17:11] side all the way to the modular foundations that are validated deliver you full outcome deliver [17:16] you a full portfolio start helping you build multiple use cases on the same infrastructure [17:22] with the same learnings with the same software stack and environment that your teams have grown [17:26] up with in developing as they develop their agentic operations and agentic assistance. [17:34] So what does this mean that you know when you look at the economics of your [17:38] you know cloud uh tokconomics right your cloud operations anything you got in the cloud anything [17:44] you start to develop and look at pile it off you start to look at what is my cost in in drivers [17:49] it's going to be choosing which LLMs it's going to be choosing about which agents are running in [17:54] the cloud on what workloads it's about how much does this cost for software licensing how do I [17:59] integrate you know full stack how do I replicate that on premise how do I do that you know in house [18:05] first or then scale out to the cloud all of these things we take in from the same report we have [18:11] the fact that you can actually still reduce spend as you look at all of these factors with you know [18:17] an on-remise approach and with our AF factory approach with 87% reduce spend so that's going to [18:22] bring that budget back into your team or not have to account for it and help you build but [18:27] you might be asking where do I start how do I get this value quickly right it's all about engaging [18:33] first with a partner that knows understanding what your requirements are you know what are [18:37] you going to build out first how how's going to help you do that and then piloting off with [18:41] one of our validated solutions or starting with our desk side approach with our Nvidia solutions [18:46] and doing that so we've got an accelerator course we've got services teams to help you kind of walk [18:51] you through what the operations are understand you know where do we start best at you know how [18:56] do we deliver value what use cases are you going to look at and then we can help you migrate into [19:00] some of the solutions You got the Nvidia library of of collections of blueprints and approaches [19:05] and software stacks that are all available via NGC that's going to help your teams build out quickly [19:11] and flexibly or you can go very standardized way with our other automation platform and blueprints. [19:17] Again, the world is the oyster here. There's so many choices to start with a Gentic on the [19:21] Dell AI factory, the Nvidia foundations that you can't go wrong. So, I appreciate that. The next um [19:28] approach is come visit us. Stop by our booth in 1307. Oh, it's not on here. And talk to Sama [19:34] and I more about your needs and operations. But until then, thank you. Thank you folks.