Dell Curbs 900 AI Experiments, Finds On-Prem Inferencing 62% Cheaper

戴尔缩减900项AI实验，发现本地推理成本降低62% 企业必须彻底重构失效的业务工作流，而非单纯自动化遗留流程，因为多智能体推理引发了算力暴增，而本地系统处理该需求的成本效益远高于超大规模云厂商。 本次讨论阐述了企业管理者如何通过采用严谨的AI治理与全栈基础设施，告别投机性实验以实现可衡量的投资回报率（ROI）。Dell分享了将内部约900个缺乏协同的AI项目整合为四大核心业务支柱的实践经验，并证明本地推理的成本效益比公有云超大规模环境高出62%。NVIDIA详细介绍了智能体工作负载在物理和算法层面所需的变革——包括机架级计算、解耦服务以及推测解码，以应对中间推理token的组合式爆炸增长。最后，发言者强调必须针对自主智能体原生重构业务工作流，并通过分阶段的工程路线图稳步推进部署。 戴尔（Dell）审查内部运营后，将 800 至 900 个未经协调的 AI“科学实验”整合成四大战略重点领域（客户服务、工程研发、销售与供应链），核心领域仅提升 0.5% 的效率便能创造数百万美元的净收益。 (1:00) 戴尔的财务模型分析显示，在本地部署并运行 AI 推理工作负载，相比使用公有云超大规模服务商（Hyperscalers）的环境可节省 62% 的成本。 (8:40) 随着智能体 AI（Agentic AI）产生大量中间推理 Token（Reasoning Tokens），计算需求呈现组合式暴增，使推理阶段对强劲算力的依赖已变得与模型训练同等重要。 (9:45) 真正的智能体 AI 落地需要从底层重新设计原生业务流程以放大人员产能，而非直接用自主智能体去自动化存在缺陷的遗留工作流。 (11:19)

Dell Technologies and NVIDIA argue that capturing measurable return on investment from agentic AI requires enterprises to consolidate fragmented experimentation, clean underlying data, and redesign workflows rather than automating broken legacy processes. Both companies advocate for modular, full-stack architectures and hybrid deployment models to sustain the heavy compute demands of autonomous intermediate reasoning. Consolidating Internal Experiments into Core Pillars (1:00) Industry data cited by Mindy Cancila indicates that 75% of surveyed organizations tie AI to business outcomes, while 87% of business leaders anticipate major vertical disruption. Despite this high interest, early enterprise adoption often devolves into ungoverned technical sprawl. At Dell, an internal audit uncovered 800 to 900 ad-hoc engineering prototypes. Recognizing that uncoordinated "science experiments" lack enterprise scalability and governance, Dell leadership intervened to focus resources exclusively on four operational pillars: customer service, products and engineering, sales, and supply chain. Cancila reported that shifting efficiency by just 0.5% across any of these four areas yields multi-million-dollar business returns. Dave Salvator emphasized that initial model experimentation is a normal phase of technology absorption, but organizations must swiftly introduce governance. Instead of deploying technology for its own sake, enterprise strategies must pinpoint where workflows are constrained by talent shortages and headcount limits, using generative and agentic systems to amplify human capacity and bring under-resourced projects into active production. Process Redesign Ahead of Autonomous Agent Deployment (11:19) While an internal Dell survey indicated that 92% of organizations claim to have deployed AI agents, Cancila clarified that true agentic systems sit at the far end of the AI maturity spectrum: they autonomously reason, adapt, make decisions, and act across disparate systems. Applying autonomous agents to flawed operational flows creates significant business risk, as automating broken processes simply accelerates failure and wastes capital. To illustrate the foundational data challenge, Cancila noted that an internal audit revealed Dell's HR and finance departments once provided conflicting employee counts, prompting a six-month initiative to clean internal data pipelines before deploying generative tools. To achieve operational returns, organizations must fundamentally redesign processes around agentic capabilities rather than transposing manual human workflows into software. In Dell's commercial sales organization, legacy motions were rebuilt into a multi-agent system dividing labor across engagement, content curation, and pricing agents. In one internal reporting workflow, a data collection and assembly task that previously occupied two analysts for an entire working day was reduced to roughly two autonomous execution hours, allowing staff to pivot to higher-value strategic initiatives. On-Premises Economics and Data-Center Scale Compute (8:40) Evaluating workload placement, Cancila cited Dell internal and external modeling indicating that executing inference workloads on-premises is 62% more cost-effective than running them in public hyperscaler environments. While enterprises operate in multicloud realities, workload placement must be governed by security parameters and strict cost modeling to ensure measurable AI return on investment. Concurrently, the transition from generative models to agentic AI is altering infrastructure requirements. Salvator explained that traditional generative AI moves linearly from input tokens to output tokens, whereas agentic systems generate massive volumes of intermediate "reasoning tokens" and combinatorial agent-to-agent exchanges. This structural explosion in compute demand renders obsolete the paradigm of reserving high-performance silicon strictly for training while using lightweight chips for inference. Modern agentic inference requires heavy computational performance, prompting NVIDIA and Dell to engineer modular rack-scale architectures—such as the Blackwell platform—that treat the entire data center as the fundamental unit of compute. Algorithmic Acceleration and Iterative Deployment Roadmaps (16:26) To handle continuous inference loads, Salvator underscored the role of algorithmic software optimizations deployed across Dell's integrated platforms. Key mechanisms include disaggregated serving—which decouples the compute-heavy prompt pre-fill stage from the memory-bandwidth-constrained token decode stage—and speculative decoding, where smaller draft models generate candidate tokens that foundation models verify in parallel, increasing decode throughput. Because AI capabilities expand significantly every 6 to 12 months, Salvator advised enterprise planners against attempting all-encompassing initial deployments. Teams should implement phased roadmaps, clearly staging basic operational features into a Revision 1.0 release while reserving complex autonomous multi-agent interactions for Revisions 1.5 and 2.0. Additionally, practitioners must track academic preprints on platforms like arXiv to anticipate architectural inflections and avoid committing to architectures that quickly become obsolete. Dell and NVIDIA frame the agentic era not as a software upgrade, but as an architectural and organizational overhaul. Measurable returns depend on solving foundational data discrepancies, fundamentally restructuring legacy processes, and sizing data-center-scale infrastructure to absorb high-density reasoning workloads.

Dell企业战略副总裁Mindy Cancila与NVIDIA加速计算产品总监Dave Salvator指出，企业迈向Agentic AI（智能体AI）的关键在于摆脱分散无序的技术实验，将算力与流程严格收敛至高价值业务支柱；同时，针对多智能体交互引发的推理算力剧增，企业须依托本地部署的成本优势、机架级全栈架构与前沿算法协同，实现可衡量的投资回报。 治理收敛：削减实验并聚焦核心业务支柱 (1:00) 根据Cancila引用的企业调查，75%的受访组织已将AI项目与业务成果直接挂钩，87%的高管预期生成式AI将对自身行业产生颠覆性影响。约两年前，Michael Dell向管理层提出预警，要求公司必须主动拥抱变革。然而Dell随后的内部审计显示，各部门自发构建了800至900个互不相通的AI用例。Cancila解释，工程师在缺乏集中治理与统一技术栈的情况下，倾向于针对孤立任务构建精巧但无法在企业级扩展的原型，极易沦为高消耗、零回报的“科学实验”。为此，Dell管理层通过自上而下的对齐干预，将所有资源强力收缩并聚焦于四大核心业务支柱：客户服务、产品与工程、销售运作以及供应链优化。 Cancila提出了“0.5%效率阈值”概念，即在这四大支柱的任一环节中实现仅0.5%的效能提升，便能换算为数百万美元的可衡量商业回报。但她强调，跨越实验阶段实现规模化落地的先决条件并非盲目配置GPU硬件，而是在基础设施部署前建立起严格的数据清洗流水线、数据治理体系以及以业务成果为导向的管理规则。 流程重构：以智能体放大产能并革新工序 (11:19) 针对概念混淆，Cancila援引Dell调查指出，虽然92%的企业声称已部署AI智能体，但多数仅属于固定规则的脚本或常规生成式自动化。Dell所定义的真正Agentic AI处于成熟度光谱最右端，具备自主推理、自适应学习并在多系统间协同行动的能力。企业若直接在未优化的老旧工序上套用自主智能体，只会加速暴露运营缺陷并造成资本浪费。Cancila分享了Dell早期的内部案例：首席运营官曾就公司员工总数分别询问人力与财务主管，却因数据源未打通而得到两个不同答案，促使Dell花费六个月时间重新梳理核心业务流程并治理底层数据。因此，引入智能体不能简单复制人类手工工序，必须围绕智能体的自主运作模式进行流程原生重构。 Salvator与Cancila一致强调，Agentic系统的本质是“放大人类产能”而非单纯削减员工，其战略价值在于激活过去因专有人才不足、编制受限或带宽短缺而搁置的高价值项目。在Dell商业销售团队的重构实践中，工作流被细分为负责日程排期的互动Agent、梳理技术资料的内容推荐Agent，以及处理繁琐测算的报价Agent。这种分工将员工从低效行政事务中解脱出来。例如在一项内部管理报表任务中，以往需要两名分析师耗费整整一天的手工数据清洗工作，在多智能体系统协同下缩短至约两小时自动完成，使分析人员能够将精力投入到更高阶的战略分析中。 部署经济学：本地推理成本优势与开放生态 (8:40) 结合过往担任云分析师的经验，Cancila指出尽管多数企业已进入多云部署格局，但底层架构的选择必须严格权衡安全边界、成本控制与实际投资回报率。根据Dell的财务模型与评估结果，在本地基础设施上运行AI推理工作负载的成本，比在公有云超大规模服务商（Hyperscalers）环境中低62%。Cancila强调，实现AI商业回报必须避免在各层技术栈产生非必要支出，企业应根据工作负载特征精准配置基础设施，而非默认将公有云作为唯一承载平台。 在落地实践上，Dell联合NVIDIA推出了“Dell AI Factory”，践行将自身作为验证试验场的策略。Cancila指出，裸算力与GPU集群脱离端到端堆栈便无法创造商业价值，必须搭配硅芯片多样性、预调优模型、系统编排及咨询服务。鉴于不同业务场景（如低延迟销售助手与大上下文客服智能体）对模型架构的要求截然不同，Dell与NVIDIA倡导涵盖红帽等软件伙伴的开放模块化生态，以保障部署灵活性并防止企业被特定技术体系深度锁定。 算力演进：机架级计算与算法协同应对推理 (9:45) Salvator指出，AI范式向Agentic演进打破了传统的计算资源模型。常规生成式AI遵循单向的“输入Token生成输出Token”，而智能体在多Agent交互与自主决策过程中，会在中间层产生海量的“推理思考Token”（Reasoning Tokens），导致计算负载呈组合式爆炸。这使得“旗舰芯片专用于训练、轻量平台用于推理”的历史架构彻底失效，推理管道同样需要极高的算力密度。现代基础设施因而将整个数据中心视作基本计算单元，推动了NVIDIA基于Blackwell架构的机架级全栈协同设计。在算法层面，NVIDIA与Dell通过软件协同部署了两项加速技术：一是将Prompt处理（Pre-fill阶段）与Token生成（Decode阶段）解耦的“解耦式服务”（Disaggregated Serving），以适应各自不同的计算与显存带宽特征；二是通过轻量小模型预测候选Token并由基座模型并行验证的“投机解码”（Speculative Decoding），大幅提升了解码周期吞吐率。 面对AI系统性能与自主能力每6至12个月便产生重大跃迁的现实，Salvator强调工程团队在制定规划时必须持续追踪arXiv等学术前沿预印本，避免因技术视界狭隘而陷入“打靶射在鸭子身后”的工程被动。在产品落地机制上，企业应坚持分阶段版本路线图，明确划定1.0版本发布范围并预留向1.5和2.0版本演进的架构弹性，以敏捷迭代姿态适应模型能力演化周期的快速轮动。 本场研讨系统梳理了企业从分散实验转向Agentic AI规模化生产的路径，确立了以四大业务支柱为靶向、本地推理降本增效及机架级软硬件协同的工程框架。然而，对于自主多智能体在跨系统执行任务时的授权边界与合规风险，发言者未作深入展开；同时，本地部署相较公有云62%的推理成本优势仍高度依赖具体工作负载规模与技术栈架构，企业在实践中仍需结合自身数据体量与运营条件进行实证测算。

## Background 本内容由频道 Ai4 于 2025 年 8 月 24 日发布在 YouTube 平台上。视频记录了戴尔科技集团与 NVIDIA 联合开展的主题研讨，重点探讨两家企业如何携手构建实用、可扩展的解决方案，引领企业自信迈入具备自适应与目标驱动特性的智能体 AI（Agentic AI）时代。 ## TL;DR 戴尔科技集团的 Mindy Cancila 与 NVIDIA 的 Dave Salvator 共同剖析了企业在落地 Agentic AI 时的战略收敛与技术底座建设。戴尔分享了从内部 900 个零散用例向四大核心业务聚焦的经验，强调业务重塑与本地推理相较云端节省 62% 成本的优势。NVIDIA 则从 Blackwell 架构、解耦式服务与投机解码等软硬件创新切入，阐明智能体带来的推理算力激增挑战，并建议企业以分阶段路线图与前瞻学术视野积极拥抱技术演进。 ## Atmosphere & style 现场对话务实严谨且不失风趣，两位技术领袖以“饮用自家香槟”的坦诚态度拆解了企业落地 AI 容易踩坑的真实教训。整场分享节奏明快、逻辑紧凑，从顶层商业 ROI 的拷问自然过渡到数据中心级算力与前沿算法的深度剖析，洋溢着务实落地与前瞻洞察兼备的行业探索热情。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=I4FKvev55E8&t=4s) 企业引入 AI 的核心已从探讨“是否会成功”转向“如何实现并衡量实际 ROI 与业务增益”。 - [01:00](https://www.youtube.com/watch?v=I4FKvev55E8&t=60s) 自下而上野蛮生长出的数百个孤立用例难以规模化，戴尔将重心聚焦在客户服务、产品研发、销售与供应链四大核心职能以撬动全局价值。 - [04:40](https://www.youtube.com/watch?v=I4FKvev55E8&t=280s) AI 的核心价值在于打破产能瓶颈、放大现有团队的专业能力，将积压的高价值项目转化为市场增量收入。 - [05:49](https://www.youtube.com/watch?v=I4FKvev55E8&t=349s) 落地 AI 需坚持业务结果先行，戴尔携手 NVIDIA 打造包含芯片算力、模型、平台软件及咨询服务的全栈 Dell AI Factory 生态。 - [08:40](https://www.youtube.com/watch?v=I4FKvev55E8&t=520s) 戴尔测算显示本地运行 AI 推理比超大规模云环境便宜 62%，企业需在多云架构中结合安全与成本权衡工作负载位置。 - [09:45](https://www.youtube.com/watch?v=I4FKvev55E8&t=585s) Agentic AI 时代智能体间协作会产生海量中间推理 Token，NVIDIA 主张将数据中心作为基本计算单元来支撑推理算力爆发。 - [11:19](https://www.youtube.com/watch?v=I4FKvev55E8&t=679s) 真正的 Agentic AI 具备自主学习、适应与独立决策能力；自动化有缺陷的旧流程只会加速混乱，引入前需从零重构业务流程。 - [14:15](https://www.youtube.com/watch?v=I4FKvev55E8&t=855s) 在销售等业务场景中，通过日程排期、内容策展与报价核算等多智能体协作，可大幅解放员工并聚焦于高价值业务。 - [16:26](https://www.youtube.com/watch?v=I4FKvev55E8&t=986s) NVIDIA 正将解耦式服务（Disaggregated Serving）与投机解码等推理端算法创新集成到底层软件栈，以加速智能体运算吞吐。 - [17:34](https://www.youtube.com/watch?v=I4FKvev55E8&t=1054s) 企业应建立分阶段迭代的版本路线图，并持续关注前沿学术研究，以适应 6 至 12 个月内 AI 能力的高速跃迁。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=I4FKvev55E8&t=4s) 拥抱智能体 AI 时代与衡量真实 ROI 会议由戴尔科技集团的 Robert Hartman 开场主持。他向到场观众致谢并指出，当前企业探索人工智能（AI）的关键早已不是“AI 是否会成功”，而是“AI 能为组织和具体项目带来多大程度的成功”。在如今算力性能呈爆发式增长、各类创新层出不穷的技术浪潮中，企业管理者面临着纷繁复杂的选择与决策压力。因此，企业必须直面几个核心拷问：投资 AI 能为企业带来怎样的实际投资回报率（ROI）？该如何科学度量其业务表现？当前 AI 的发展现状究竟如何？ 随后，主持人介绍了本场演讲的两位重磅嘉宾——来自戴尔科技集团（Dell Technologies）的企业战略副总裁 Mindy Cancila，以及来自 NVIDIA 的加速计算产品总监 Dave Salvator。戴尔与 NVIDIA 长期保持着紧密的战略合作关系，本次对话将聚焦于智能体 AI（Agentic AI）的发展趋势，并深入剖析当前企业落地 AI 的现状与实践路径。 --- ### [01:00](https://www.youtube.com/watch?v=I4FKvev55E8&t=60s) 从 900 个用例的野蛮生长到聚焦四大核心业务 Mindy Cancila 接过话茬并提到，在近期的多场行业交流中，“智能体（Agentic）”已成为最高频的讨论焦点，企业普遍迫切想知道该如何开启自己的 AI 转型之旅。她首先分享了戴尔调研中的几组关键数据：不仅有 75% 的企业已经将 AI 紧密绑定至实际业务成果，更有高达 87% 的受访企业高管明确表示，以生成式 AI 为代表的技术将对其所在行业（无论是制造业、医疗保健还是其他垂直领域）带来全面而深刻的颠覆。 面对这场历史性变革，戴尔科技集团首席执行官 Michael Dell 早在两年前便展现出了技术派领袖的敏锐洞察，他当时对团队明确强调：“这场变革对所有人来说都将是颠覆性的，我们要么被它颠覆，要么由我们自己来主导这场颠覆，并进而赋能我们的客户实现同样的变革。”（"This is going to be massively disruptive for everybody. And we can either let it disrupt us or we can drive the disruption for oursel and then help our customers do the same."） ```mermaid flowchart TD accTitle: 戴尔生成式 AI 落地策略收敛路径 subgraph S1["初期探索阶段（自下而上）"] A["工程团队自发探索"] --> B["涌现 800-900 个离散用例<br/>（单一目标、无法横向扩展、缺乏统一治理）"] end subgraph S2["战略聚焦阶段（自顶向下）"] B --> C["管理层顶层梳理与收敛"] C --> D1["客户服务<br/>(Customer Service)"] C --> D2["产品与工程研发<br/>(Products & Engineering)"] C --> D3["销售赋能<br/>(Sales)"] C --> D4["供应链协同<br/>(Supply Chain)"] end subgraph S3["商业价值释放"] D1 & D2 & D3 & D4 --> E["单点效率提升 0.5% 即可带动全局显著 ROI"] end ``` 戴尔作为一家拥有深厚工程师文化的企业，起初对内部进行了摸底调查，结果发现各团队自发构建的生成式 AI 用例竟然多达 800 至 900 个。Mindy 坦言，工程师拿到新工具后往往热衷于针对单一场景构建精巧但孤立的方案，由于技术尚处萌芽期，这些项目根本无法规模化推广，更缺乏统一的协同调度和治理机制，最终极易沦为无疾而终的“科学实验”。戴尔高管层随后迅速推行自顶向下的战略对齐，将重心收敛到决定企业核心竞争力的四大关键职能：客户服务（Customer Service）、产品与工程研发（Products & Engineering）、销售（Sales）以及供应链（Supply Chain）。只要在这四大领域中实现哪怕仅仅 0.5% 的效率提升，就能产生撼动全局的量化 ROI。 - **调研核心数据**：75% 的组织将 AI 投资与业务成果直接挂钩；87% 的企业高管预计生成式 AI 将彻底颠覆其所属行业。 - **内部用例过载教训**：内部早期自发涌现出 800–900 个 AI 用例，由于缺乏统一治理与规模化能力，存在极高的试错成本与失败风险。 - **四大核心聚焦领域**：客户服务、产品工程、销售体系、供应链管理。 - **企业普遍痛点**：缺乏启动路径规划、治理体系缺失、数据未清洗准备好、难以精准挂钩量化业务目标。 --- ### [04:40](https://www.youtube.com/watch?v=I4FKvev55E8&t=280s) 技术必须服务于商业约束与产能放大 NVIDIA 加速计算产品总监 Dave Salvator 对 Mindy 分享的内部探索过程深表认同。他指出，工程师在初期自发尝试数百个用例是理解 AI 潜力的必经阶段，但在探索之后，企业必须迅速聚焦于自身面临的真实业务挑战。Dave 强调，“为了技术而技术并不能真正解决问题”（"Technology for its own sake doesn't really solve problems."），企业应当反思自身的瓶颈到底在哪里。 在许多企业内部，常常存在大量由于人手短缺、缺乏专属专业知识而被迫搁置、甚至长期停滞的高价值项目。AI 的核心价值在于赋能和放大（amplify）现有团队的专业能力，将那些原本沉没在底层的边缘项目推升到基准线之上，使其得以顺利推向市场并直接转化为增量收入（incremental revenue）与实际利润，这才是企业引入 AI 的根本目的。 - **核心价值定位**：AI 的首要任务是赋能现有团队并打破产能瓶颈，而非盲目堆砌技术。 - **盘活滞后项目**：借助 AI 补齐专业知识和人力缺口，将积压项目转变为可推向市场的创收业务。 - **终极商业逻辑**：通过提升生产力与效率，为企业创造实打实的增量营业收入与净收益。 --- ### [05:49](https://www.youtube.com/watch?v=I4FKvev55E8&t=349s) 践行自身方案：构建 Dell AI Factory 全栈生态与开放选择 Mindy 幽默地借用“饮用自家的香槟”（Drink your own champagne，即吃自家狗粮）来概括戴尔的战略实践，并打趣称身处拉斯维加斯当然更要“畅饮香槟”。她指出，戴尔在内部推行 AI 所踩过的坑和积累的经验，对广大企业客户具有极高的复用价值。任何企业都拥有产品研发、销售体系、客户支持和供应链交付这套基础底座，戴尔的成功路径完全可以无缝映射到其他行业。 在落地方法论上，绝不能从底层技术切入，而必须从希望解决的业务结果出发。尽管 AI 领域的创新迭代速度超越了过去数十年的 IT 发展周期，但只要锚定业务目标，技术架构的搭建就会顺理成章。以戴尔与 NVIDIA 联合打造的“Dell AI Factory”（戴尔 AI 工厂）为例，光有算力基础设施是无法直接支撑业务的，企业需要的是贯穿硅芯片、模型层、全栈解决方案、生态系统以及配套咨询服务的端到端能力。 ```mermaid flowchart LR accTitle: Dell AI Factory 端到端全栈能力架构 subgraph L1["算力与基础设施层"] A1["多样化芯片与 GPU 算力"] A2["基础 IT 设施"] end subgraph L2["模型与平台层"] B1["NVIDIA AI 基础模型与软件栈"] B2["Red Hat 等开放生态组件"] end subgraph L3["全栈方案与咨询服务"] C1["Dell 咨询与端到端集成服务"] C2["基于自身实践沉淀的最佳实践"] end subgraph L4["场景化业务应用"] D1["销售智能体应用"] D2["客服智能体应用"] end L1 --> L2 --> L3 --> L4 ``` Mindy 特别强调了“开放生态系统”与“客户自主选择权”的重要性。因为“一种规格无法适配所有需求”（"One size is not going to fit all."），销售场景所需要的模型和技术栈，与客户服务场景截然不同。为此，戴尔在全栈各个层面与 NVIDIA 开展深度整合，将 NVIDIA 的前沿模型与软件完全纳入戴尔技术栈，并携手 Red Hat 等生态伙伴，为客户提供兼具灵活性与针对性的软硬件一体化交付方案。 - **业务先行原则**：不以技术选型为起点，而是优先定义业务目标，驱动技术架构自然成型。 - **Dell AI Factory 全栈要素**：涵盖多样化硅芯片与 GPU 硬件、模型层、端到端解决方案集成以及专业咨询服务。 - **开放生态理念**：反对专有封闭绑定，强调跨芯片、模型与平台的多样化选择（如整合 NVIDIA 模型资产、Red Hat 平台等）。 - **差异化场景适配**：销售、客服、研发等不同业务场景需匹配不同规模与特性的模型及支撑软件。 ### [08:40](https://www.youtube.com/watch?v=I4FKvev55E8&t=520s) 本地推理的成本优势与全栈多云架构考量 戴尔企业战略副总裁 Mindy Cancila 回顾了戴尔启动相关项目时的历程：迈克尔·戴尔（Michael Dell）曾要求团队进行深入评估，探究这些 AI 工作负载究竟应该部署在哪里、未来的基础设施格局会是什么样，以及客户最终会做出怎样的选择。团队完成了一份非常详尽的分析报告（Mindy 表示很乐意分享公开报告及内部细节），尽管其中有许多需要权衡的权衡点（puts and takes），但当时的测算得出了一个明确结论：在本地（on-prem）运行模型推理比在超大规模云厂商（hyperscalers）的环境中运行要便宜 62%。 作为拥有多年云计算分析师背景的专业人士，Mindy 强调自己始终保持务实。她认为绝大多数组织已经并将继续处于多云（multicloud）环境中，而 AI 基础设施正是这一多云旅程的自然延伸。不同用例适合运行在不同位置，企业必须权衡安全形态与成本支出；如果在技术栈中不必要的环节挥霍预算，就根本无法实现预期的 ROI。因此，针对具体业务用例精准匹配底层架构，是实现商业成果的关键前提。 - 戴尔的详细评估显示，本地进行 AI 推理比超大规模云环境成本效益高出 62%。 - 企业基础设施正处于多云格局的延伸期，需根据具体用例的安全性与成本要求选择运行位置，避免不必要的堆栈开销侵蚀 ROI。 --- ### [09:45](https://www.youtube.com/watch?v=I4FKvev55E8&t=585s) 数据中心即计算单元：Blackwell 架构与智能体带来的算力变革 NVIDIA 加速计算产品总监 Dave Salvator 对此表示高度赞同，指出 NVIDIA 与戴尔以及众多生态伙伴的合作同样采用这种全栈（full-stack）视角。随着行业向智能体 AI（Agentic AI）乃至物理 AI（Physical AI）演进，企业在部署技术方案时，眼光必须超越单一芯片、单台服务器、甚至单个机柜的层面，而必须直接从“数据中心级别”进行全局考量。在 NVIDIA 的设计哲学中，“数据中心本身就是一个计算单元”（the data center is actually a unit of compute）。NVIDIA 基于最新的 Blackwell 架构构建机架级（rack-scale）系统，正是为了解决这一代际转变带来的计算瓶颈。 生成式 AI（Generative AI）最初主要涉及输入 Token（input tokens）和输出 Token（output tokens）的转换；然而迈向 Agentic AI 时代后，系统在中间阶段会产生海量的“推理/思考 Token”（reasoning tokens）。随着不同智能体之间相互交互与调用，还会引发算力需求的组合爆炸效应（combinatorial effect）。过去业界普遍认为最繁重的算力负载只发生在模型训练阶段，而在推理端可以采用非旗舰级的硬件平台；但智能体技术的兴起彻底打破了这一假设，它对推理算力的要求急剧攀升。企业在制定规划周期时，必须认识到重度算力开销将同时贯穿训练与推理两端。 ```mermaid flowchart LR accTitle: 生成式 AI 与智能体 AI 计算范式对比 subgraph GenAI ["传统生成式 AI"] direction LR G_In["输入 Token"] --> G_Out["输出 Token"] end subgraph AgenticAI ["Agentic AI 计算范式"] direction LR A_In["输入 Token"] --> A_Reason["海量推理/思考 Token<br/>(智能体间多轮协作)"] A_Reason --> A_Out["最终输出 Token"] end ``` - NVIDIA 主张将“数据中心作为计算单元”，其 Blackwell 架构针对机架规模计算进行了专门优化。 - 智能体 AI 在生成最终输出前需要处理海量的中间推理 Token（reasoning tokens），导致推理阶段的算力消耗大幅激增，必须采用顶级算力平台支持。 --- ### [11:19](https://www.youtube.com/watch?v=I4FKvev55E8&t=679s) 厘清 Agentic 的真正定义与戴尔的业务流程重塑 进入对 Agentic 的探讨时，Mindy 首先指出了当前业界的普遍认知混淆。戴尔在当年早些时候进行的一项调研显示，高达 92% 的受访企业声称自己已经部署了智能体（agents），团队看到这个数据时不禁感到诧异。在戴尔的技术成熟度光谱中，“Agentic”位于最右端的极高维度——它指代的是系统具备自主学习、自我调整并能独立做出决策的能力（"the ability to learn and adapt and to make decisions autonomously"）。 Mindy 澄清，这并不是要贬低现有应用中成熟的自动化流程，如果既有自动化能产生商业价值就应该继续推进；但真正的 Agentic AI 必须能够自主推理和执行动作，并且能在跨系统环境中以多智能体（multi-agent）形态协作运转。戴尔在开启自身 AI 旅程时，最初花了约整整六个月时间深入梳理业务流程。她告诫道：“自动化有缺陷的流程只会让糟糕的事情崩溃得更快”（automating broken processes are just going to break things really quickly）；如果在糟糕的流程之上堆砌智能体，不仅会在极短时间内造成更大规模的混乱，还会烧掉巨额资金，彻底毁掉 ROI。 作为一家拥有 40 年历史、历经多次并购的大型企业，戴尔内部也曾面临系统拼接与数据孤岛带来的巨大挑战。Mindy 分享了一则内部轶事：戴尔首席运营官（COO）曾问起公司到底有多少名员工，结果人力资源主管和财务主管给出了两个截然不同的数字。戴尔在生成式 AI 建设初期投入大量人工精力清理底层数据源。而面对智能体时代的到来，团队决定不再拘泥于原有的“人类工作流程”，而是全面推翻重想：如果站在全新的起点、并且预先知晓智能体技术存在，流程应该如何从零构建？ - 调研显示 92% 的受访机构声称已部署智能体，但多数仅为传统自动化，并非真正的自主智能体。 - 戴尔将真正的 Agentic 定义为具备自主学习、适应与独立决策能力，并在跨系统场景下依托协议协作的多智能体网络。 - 戴尔耗时 6 个月梳理核心流程，防止将智能体强行嫁接到损坏的旧流程上导致错误与成本按指数级扩散。 --- ### [14:15](https://www.youtube.com/watch?v=I4FKvev55E8&t=855s) 销售场景下的多智能体协作实践与高价值赋能 以戴尔销售团队与客户的互动流程重塑为例，戴尔的目标不是简单地拿既有流程套用智能体，而是探讨如何利用 Agentic 释放销售代表的生产力，使其摆脱繁琐事务，专注于深度客户沟通。如果重新设计整个架构，业务流将被拆解并分配给各自专精的多个智能体： - **日程协同智能体**：专门负责与客户之间的沟通排期与日程对接； - **内容策展智能体**：自动筛选整理针对该客户的高价值物料，供销售代表决定沟通时机与策略； - **报价与核算智能体**：负责在后台与客户需求之间进行反复的测算、定价与生成报价单。 这种多智能体系统协同工作，目标绝不是取代销售人员或客户，而是作为赋能工具让人员的工作更敏捷、更高效。Mindy 观察到，智能体最容易取得成功的切入点，往往是那些既能创造业务价值、又是员工日常极力想要摆脱的枯燥事务。例如戴尔每天需要向高管团队报送一份关键数据报告，以往需要两名全职员工每天花费大量时间搜集、整理、清洗并发送；引入多智能体系统后，整个流程在两小时内即可自动完成。这两名员工并未被裁减，而是得以抽身去从事更高价值的工作。 --- ### [16:26](https://www.youtube.com/watch?v=I4FKvev55E8&t=986s) 全新构建与推理端算法创新：投机解码与解耦式服务 Dave Salvator 对 Mindy 的流程重塑理念深表认同，重申在引入智能体时不应采取“外挂/拼凑”（bolt-on）的思维，而应采用全新构建（net new）的眼光，并建立中长期的技术演进路线图。Dave 指出，尽管技术演进极快让人难以看清所有发展拐点，但无论是平台硬件能力还是底层算法创新都在突飞猛进，尤其是在模型推理端。 Dave 重点提到了两项关键的推理优化前沿技术： 1. **解耦式服务（Disaggregated Serving）**：在讲解过程中 Dave 一度将其与投机解码口误混淆，随后迅速纠正说明。该技术将大语言模型推理中的预填充（pre-fill）阶段与解码（decode）阶段彻底分离成不同的独立操作，因为这两个阶段在计算敏感度、内存带宽与性能瓶颈上各不相同，解耦部署能够极大提升整体资源利用率与吞吐表现。 2. **投机解码（Speculative Decoding）**：旨在通过小模型预先推测、大模型批量验证等机制，大幅加快推理中耗时最长的解码阶段。 NVIDIA 正在加速将这些前沿推理算法集成到底层核心软件库中，并随后全面落地到戴尔的基础设施平台解决方案里，帮助企业应对智能体落地带来的计算挑战。 ### [17:34](https://www.youtube.com/watch?v=I4FKvev55E8&t=1054s) 紧跟前沿研究并制定迭代路线图 Dave Salvator 强调，在迈向智能体 AI（Agentic AI）和生成式 AI（Generative AI）的浪潮中，企业必须时刻保持对前沿技术研究管线的敏锐扫描与理解。NVIDIA 内部拥有实力雄厚的深度研究团队，不断探寻“下一步能做什么”以及“如何将技术推向更高水平”。他提到，业界应该密切关注像康奈尔大学托管的 arXiv 这类预印本论文平台，时刻洞察下一轮驱动技术能力跃升的创新。企业在进行规划时必须尽可能具备前瞻性视野，切忌目光短浅，否则“三年后再回头看，就会发现自己落后于时代（shoot behind the duck）”。 在将现有业务流程升级到智能体时代的过程中，Dave 建议企业管理者和开发者树立清晰的路线图思维。软件开发领域有一条基本常识：不可能在 1.0 版本中就实现所有理想功能。因此，务必区分清楚“1.0 版本要落地什么”与“1.5、2.0 乃至更长远的未来规划是什么”。鉴于 AI 技术演进的惊人速度，行业正在经历持续的集体学习——即使在短短的 6 到 12 个月内，AI 的能力都将发生巨大跃迁，企业唯有保持终身学习与灵活演进，才能把握住未来的先机。 - **追踪顶尖学术动态**：持续关注 NVIDIA 深度研究团队的探索方向以及 arXiv 等平台的最新论文，提前布局下一代核心能力。 - **避免技术滞后风险**：前瞻性规划可防止数年后出现方案落伍、被迫推倒重来的窘境。 - **分阶段版本演进**：将功能拆解为 1.0 版（基础落地）与 1.5/2.0+ 版（深度智能化），避免因追求一步到位而阻碍项目启动。 - **应对高速迭代周期**：以 6 至 12 个月为一个关键观察窗口，针对模型能力的飞跃持续调整与升级系统架构。 ```mermaid flowchart LR accTitle: 智能体 AI 演进规划与迭代路径 subgraph Track["前沿输入"] A["学术研究与论文<br/>(arXiv 等)"] B["NVIDIA 深度研究探索"] end subgraph Plan["版本演化路径"] C["1.0 版本<br/>核心流程接入智能体"] D["1.5 / 2.0+ 版本<br/>高阶自主与全链路协同"] end subgraph Goal["最终目标"] E["保持技术领先<br/>应对 6-12 个月快速变革"] end Track --> C C --> D D --> E ``` ### [18:54](https://www.youtube.com/watch?v=I4FKvev55E8&t=1134s) 峰会总结与展台交流指引 主持人 Robert Hartman 对 Mindy Cancila 与 Dave Salvator 的深度分享表示衷心感谢。他总结道，本次演讲深入展示了 Dell Technologies 与 NVIDIA 双方如何合力布局 Agentic AI，这不仅为各大企业制定智能化战略指明了方向，也提供了兼具极致性能与高度确定性的落地信心，确保智能体能够在实际业务中输出更强的洞察力与生产力。 由于现场演讲时间所限，未能安排单独的 Q&A 环节，主持人向在场嘉宾展示了展位信息的幻灯片，热情邀请大家次日前往 Dell 与 NVIDIA 的联合展台参观交流，并在会议室外直接与 Dave、Mindy 展开面对面的深入沟通与答疑。最后，三位演讲嘉宾在全场的致谢与道别声中圆满结束了本次专题研讨。 ## Notable quotes > [01:00](https://www.youtube.com/watch?v=I4FKvev55E8&t=60s) This is going to be massively disruptive for everybody. And we can either let it disrupt us or we can drive the disruption for oursel and then help our customers do the same. — Michael Dell (quoted by Mindy Cancila) > [04:40](https://www.youtube.com/watch?v=I4FKvev55E8&t=280s) Technology for its own sake doesn't really solve problems. — Dave Salvator > [05:49](https://www.youtube.com/watch?v=I4FKvev55E8&t=349s) Drink your own champagne — Mindy Cancila > [05:49](https://www.youtube.com/watch?v=I4FKvev55E8&t=349s) One size is not going to fit all. — Mindy Cancila > [09:45](https://www.youtube.com/watch?v=I4FKvev55E8&t=585s) the data center is actually a unit of compute — Dave Salvator > [11:19](https://www.youtube.com/watch?v=I4FKvev55E8&t=679s) the ability to learn and adapt and to make decisions autonomously — Mindy Cancila > [11:19](https://www.youtube.com/watch?v=I4FKvev55E8&t=679s) automating broken processes are just going to break things really quickly — Mindy Cancila > [17:34](https://www.youtube.com/watch?v=I4FKvev55E8&t=1054s) shoot behind the duck — Dave Salvator ## Mentioned Dave Salvator, Mindy Cancila, Robert Hartman, Michael Dell, Dell Technologies, NVIDIA, Dell AI Factory, Red Hat, Blackwell, Cornell University, arXiv, Ai4 ## Video description Dave Salvator, Director of Accelerated Computing Products, NVIDIA Mindy Cancila, VP of Corporate Strategy, Dell Technologies Robert Hartman, AI Factory with NVIDIA Solutions, Dell Technologies Abstracts Explore the transformative power of AI in business and productivity at this engaging session. Discover how Dell and NVIDIA collaborate to deliver practical, scalable solutions built for the next era of autonomy with real-world applications in action. Connect with thought leaders driving innovation forward and learn how today’s intelligent systems are paving the way for Agentic AI—adaptive, goal-driven technologies that elevate decision-making and unlock new potential. If you’re ready to lead your organization confidently into the future of AI, don’t miss this opportunity to gain actionable insights and chart your path toward progress. [Source](https://www.youtube.com/watch?v=I4FKvev55E8)

[1:00] Michael Dell (quoted by Mindy Cancila): This is going to be massively disruptive for everybody. And we can either let it disrupt us or we can drive the disruption for oursel and then help our customers do the same. [4:40] Dave Salvator: Technology for its own sake doesn't really solve problems. [5:49] Mindy Cancila: Drink your own champagne [5:49] Mindy Cancila: One size is not going to fit all. [9:45] Dave Salvator: the data center is actually a unit of compute [11:19] Mindy Cancila: the ability to learn and adapt and to make decisions autonomously [11:19] Mindy Cancila: automating broken processes are just going to break things really quickly [17:34] Dave Salvator: shoot behind the duck

[1:00] Mindy Cancila: didn't have any business doing eight or 900 use cases [5:49] Mindy Cancila: drinking your own champagne [11:19] Mindy Cancila: Automating broken processes are just going to break things really quickly [11:19] Mindy Cancila: There was a point in time where our COO had said if he wanted to know how many employees were in the company, and he asked our HR leader and our finance leader, he got two different answers. [17:34] Dave Salvator: boy did we shoot behind the duck.

[0:04] All right folks um as you say and thanks [0:06] for attending our session this evening. [0:08] Hopefully it's been a great day for for [0:09] you. One of the things that come you [0:11] come to the show for is understanding [0:13] about AI and AI obviously is not a [0:14] question about if AI will be successful [0:16] but how successful can it be for your [0:18] organization and all your projects you [0:21] know with the amount of performance [0:22] innovation and everything that's thrown [0:24] at you. There's a lot to consider and [0:25] work through. So, you got to ask [0:27] yourself some questions about what is [0:29] the real ROI and return that you're [0:31] going to get from AI, from investing in [0:32] your corporation. How are you going to [0:34] measure that performance? What's [0:36] happening with the state of AI? And [0:37] that's what we're here to do today with [0:38] Dave and Mindy. And we're introducing [0:40] them. Mind is our VP of of strategy and [0:43] at Dell Technologies. And Dave is [0:44] director of accelerated compute at [0:46] NVIDIA. They're both we're all strong [0:48] partners here. We work together. I want [0:50] to introduce the topic and bring it on [0:52] turn it over to Mindy to talk about the [0:54] question about Agentic AI and also [0:56] delivering what the state of AI is [0:58] today. Take it away Mindy. [1:00] >> Yeah, thanks. Happy to be here. I had [1:02] the pleasure of sitting in a couple of [1:03] sessions and no surprise there's a lot [1:05] of talk about agentic um and a lot of [1:07] interest in how do we actually get [1:09] started on our journey. So I thought [1:11] today between Dave and I we'd take a few [1:13] minutes and talk about our journey and [1:14] the things that have forged the [1:16] partnership between Dell and Nvidia. on [1:18] my side. You know, I think it's [1:19] interesting. We fielded a bunch of [1:21] surveys, and I'm sure you've seen a [1:22] bunch of survey data already today. Um, [1:24] one of the stats that I found really [1:26] interesting is not just the 75% of [1:28] organizations are tying AI to their [1:30] business outcomes, but 87% of leaders [1:34] that we surveyed, [1:36] expect AI, generative AI in particular, [1:39] to disrupt their industry. Whether it's, [1:42] you know, manufacturing, healthcare, [1:44] whatever that vertical or industry is, [1:46] it's everywhere. It's pervasive across [1:48] everything. And as as mentioned, I work [1:50] at Dell Technologies and we've really [1:53] leaned into AI as a company. You know, [1:56] Michael is a technologist at heart. He [1:58] loves to think through technologies and [2:00] be on the forefront of what's emerging. [2:03] And about two years ago, he came in and [2:05] he said, "This is going to be massively [2:07] disruptive for everybody. And we can [2:08] either either let it disrupt us or we [2:12] can drive the disruption for oursel and [2:14] then help our customers do the same. [2:16] And so when we started on our journey, [2:18] the first thing that we did was to put a [2:21] um a toll out to kind of see how much [2:24] generative AI is actually out there [2:25] today. What are people actually doing [2:27] within our own four walls? And we found [2:30] a staggering number of use cases. In [2:33] that survey that we had done internally, [2:35] we found that we had more than eight or [2:37] 900 use cases. Again, this is two years [2:39] ago. We're an engineering organization. [2:42] We like to play with technology. I'm an [2:44] engineer by trade. I find if you give [2:46] engineers tools, we'll build things that [2:47] are amazing and purpose-built for a [2:50] singular use case and probably not [2:52] terribly scalable across a bunch of [2:54] others, at least not initially, and not [2:56] when the technologies in its infancy. [2:58] Um, I can tell you we didn't have any [3:00] business doing eight or 900 use cases. [3:02] They weren't going to be successful. [3:03] They weren't going to drive a meaningful [3:05] ROI. They weren't going to be [3:06] orchestrated or governed. And they [3:08] weren't actually going to tie back to [3:09] those business outcomes that were most [3:11] important for us. And so Michael and the [3:14] leadership team really leaned in and [3:15] said, "We're going to align on the key [3:17] use cases for our business." And the way [3:19] that we did that was to think about what [3:21] it is that makes us special as a [3:23] company. And we came back to four key [3:25] functions. It was customer service, all [3:27] about our customers, our products, the [3:30] engineering, and what we build, our [3:32] sales, and our supply chain. And if we [3:35] could become even just half a percent [3:38] more performant in any of those four [3:40] areas, we could move the needle and have [3:42] a measurable ROI. Conversely, if we let [3:46] the organization just sort of run ad hoc [3:49] with eight or 900 use cases, we were [3:52] pretty likely to not be successful and [3:54] have that same outcome. So it became [3:56] really important to us that we align [3:58] tops down on a set of use cases that [4:00] were going to be critical for us and [4:02] that we tied those back to things that [4:03] were going to have measurable business [4:05] outcomes. Otherwise, they become science [4:07] experiments and they're going to get [4:08] shut down. So that was really what [4:10] started us down the journey and sort of [4:12] what we see happening in industry. I see [4:14] so many enterprise organizations asking [4:16] those same questions. Where should we [4:18] get started? How do we go about this? [4:20] How do we stand up the right governance? [4:22] Oh, hey, by the way, what about the data [4:24] that we need? What state is the data in? [4:27] How do we get it prepared, cleansed, [4:28] ready to actually go drive the kind of [4:30] outcomes we need? And then most [4:32] importantly, back to that stat, how do [4:34] we tie it back to our business outcomes? [4:36] So, I don't know, Dave, what are your [4:37] thoughts on that? And what do you see [4:38] from your side? [4:40] >> Well, I mean, I think you've summarized [4:41] it well. um you know and it's [4:43] interesting to hear about kind of the [4:45] those hundreds of use cases you had [4:46] internally because part of the process [4:48] of getting to what is it that AI can do [4:50] for us is almost that notion of kind of [4:52] playing with it as engineers do and some [4:54] of those you know you're going to [4:55] discover things along the way with that [4:56] but then to your point at a certain [4:58] point you have to focus and say what are [5:00] the actual business problems we face as [5:01] an organization you know uh and so one [5:04] of the things of course always is how [5:06] can we be more efficient how can we [5:07] amplify what our teams do and again [5:09] looking at through the lens of business [5:11] problems because you know technology for [5:13] its own sake doesn't really solve [5:15] problems. So you really have to begin [5:16] with, you know, where are we [5:17] constrained? Where are their projects [5:19] that are languishing because we simply [5:21] either don't have the people, we don't [5:23] have the we don't have the internal [5:25] expertise or what have you to actually [5:27] take those projects on which could [5:28] ultimately be a creative for the [5:30] business. And so if you look at those [5:32] and start thinking about how can AI help [5:34] amplify our teams and what they can do [5:36] to bring some of those things above the [5:38] line, bring them into the business, [5:39] actually bring them to market and [5:41] hopefully generate incremental revenue, [5:43] incremental actual earnings for for the [5:45] company, which at the end of the day is [5:47] what businesses are trying to do. [5:49] >> Yeah. And I mean, I feel like we took [5:51] that and sort of segueing it into how do [5:53] we make these things become real? From [5:56] my standpoint, I think it's the best [5:58] case of eat your own dog food or drink [6:00] your own champagne. I I really prefer [6:01] the latter. Um, you know, when in Vegas, [6:04] I guess, is is the mantra. So, um, [6:07] kidding aside, I really have found that [6:10] in our journey, the things that we [6:12] really set out to do within Dell aren't [6:15] that different than most enterprise [6:17] organizations. Obviously, we're [6:18] different and that we're a vendor, but [6:19] when you go think back to the four use [6:21] cases and the areas of focus that we [6:24] decided to lean into, they're pretty [6:26] transferable to pretty much any p any [6:28] enterprise organization. Most of you [6:30] have a product you're building. You have [6:32] a sales function. You have to support [6:34] your customers and you have a way to get [6:36] the product to your customers. So, we're [6:38] talking about core blocking and [6:40] tackling. We're not talking about things [6:41] that are really sophisticated and unique [6:44] to a vendor in particular. And so when [6:46] we went through our journey and we [6:47] thought about how are we actually going [6:49] to stand these use cases up and make it [6:52] real, the way that we approached that [6:55] was not to start with the technology, [6:57] but rather what were those business [6:59] outcomes that we were trying to solve [7:01] for. And the technology, I'm going to [7:03] oversimplify this, kind of works itself [7:05] out. Now, it's really difficult in AI [7:06] because everything is moving so quickly. [7:09] The pace of innovation is not like [7:10] anything I've ever seen in my couple [7:12] decades in the IT industry. But when we [7:15] thought about it, you know, if you if [7:16] you look at like the Dell AI factory as [7:18] example, there's infrastructure. You [7:20] don't do a whole lot of AI with that [7:21] with just core infrastructure. You need [7:23] some things that sit around that. You [7:25] need a full stack of capabilities. [7:27] Again, starts really with the models and [7:30] the endtoend stack that we've brought [7:33] together with an ecosystem of services, [7:35] the consultative services wrapped around [7:37] that from what we learned through our [7:39] own journey. That really is how we've [7:41] worked to bring these capabilities to [7:43] life. And Dell's always been a believer [7:46] that open ecosystems are really [7:48] important so that you as a customer have [7:50] choice in what it is you want to deploy. [7:52] And it's why you see us forge [7:53] partnerships at every layer from the [7:55] silicon layer and diversity all the way [7:57] up to the the models and the different [8:00] types of solutions. One size is not [8:02] going to fit all. So the solution and [8:03] the and the stack that you bring to bear [8:06] for that sales use case in our example [8:08] is going to look different than what you [8:09] might need for customer service and as a [8:11] result you're probably going to choose [8:12] different models and that's where we [8:14] really strengthen the partnership with [8:16] Nvidia being able to bring their models [8:18] completely into our stack. We're doing [8:19] the same with other partners like Red [8:21] Hat um and really bringing an endto-end [8:23] set of capabilities so that you have [8:25] choice in how you make that deployment [8:27] and again playing out our journey. [8:29] You're going to need some GPUs. You're [8:31] going to need some models. You're going [8:32] to need some services to tie all that [8:34] together and an ecosystem of partners [8:36] wrapped around that with the idea that [8:38] you're going to want choice in how you [8:40] do those deployments. When we got [8:42] started, Michael asked my team to do [8:44] some work to assess where should we run [8:46] these things and what is that going to [8:47] look like and what do we think our [8:48] customers will end up doing. We had done [8:51] some early analysis and it was quite [8:53] detailed. I'm happy to share the public [8:56] report we ended up doing as well as the [8:58] internal details because there are a lot [8:59] of puts and takes but at that time we [9:01] found that it was 62 more cost 62% more [9:05] cost effective to do inferencing on prem [9:08] than it was in a hyperscalers [9:09] environment. I like to be pretty [9:11] pragmatic those who know me um and have [9:14] followed me for years. I was a cloud [9:15] analyst. I think most organizations [9:17] already are and will continue to be [9:18] multicloud and this is an extension of [9:21] that multicloud journey. But you're [9:22] going to have some use cases that make [9:24] sense to run in different places. And [9:26] the choice that you have to decide the [9:28] right type of security, the cost, again, [9:31] you're not going to hit that ROI if [9:33] you're spending money in in in pockets [9:35] of the stack where you don't need to. [9:37] So, we think it's critically important [9:38] to really map the right architecture for [9:41] the use case to get to that business [9:43] outcome. [9:45] >> Yeah. And again, I think you you've said [9:46] it well. I mean and the way that we work [9:48] with Dell and the way we work with many [9:51] of our partners is we also take that [9:52] very full stack approach right and [9:54] increasingly and we're going to talk [9:55] about in just a moment um what we're [9:57] seeing now as we evolve towards agentic [9:59] and even physical AI that as you think [10:01] about deploying technologies to solve [10:04] for these problems you have to get past [10:06] the chip even the server level even in [10:09] some cases the rack level of detail and [10:11] think at the data center level of detail [10:13] right in Nvidia we talk about the notion [10:15] that the data center is actually a unit [10:16] of compute, you know, and we're now [10:18] building at rack scale um with with our [10:20] current generation of products based on [10:22] the black wall architecture precisely to [10:24] solve for these problems because as we [10:26] move towards Agentic, Agentic is going [10:27] to raise the stakes that much more in [10:30] terms of you know initially with it with [10:32] uh uh generative AI which has done some [10:35] amazing things for us. You basically had [10:36] input tokens and output tokens. Well, [10:38] now as we go to Agentic, you're going to [10:40] have this massive number of reasoning [10:42] tokens in the middle. And uh from the [10:44] last session that was just in this room, [10:46] they were talking about as agents [10:47] continue to talk to agents, you'll have [10:49] this combinatorial effect on top of [10:51] that. And so when you think about scale, [10:53] you know, we've historically thought, [10:54] well, the really heavy lifting has to [10:55] happen on training. And then when we [10:57] deploy on inference, well, maybe we can [10:59] use, you know, uh maybe not quite [11:01] flagship level uh platforms. Uh I think [11:04] Agentic very much changes that. So as [11:06] you think about your planning cycles and [11:07] think about what you're looking to [11:08] deploy and when you're looking to deploy [11:10] it, bear in mind that that heavy lift [11:13] aspect is going to affect both training [11:15] and inference and you should probably [11:16] plan accordingly. [11:19] >> Probably a good time to talk about [11:20] agentic. So I have to start with just [11:23] sort of a definition in the spectrum of [11:26] all things agentic. I find that a lot of [11:29] organizations are maybe a little bit [11:31] confused. We had done a survey earlier [11:34] this year and we found that 92% of [11:37] organizations said they were already [11:38] deploying agents and we kind of looked [11:40] at that and went wait what? And so at [11:43] least within Dell I'll share our [11:45] definition on the spectrum of maturity [11:47] is really far off to the right when we [11:49] talk about agentic we're talking about [11:52] the ability to learn and adapt and to [11:54] make decisions autonomously. [11:56] And that's a really important [11:57] distinction. And I'm not trying to [12:00] downplay the importance of the things [12:01] that we've automated that already live [12:03] in applications and are there today. You [12:05] should continue doing those. They're [12:06] great. And if they drive business value [12:08] for you, that's fantastic. But when we [12:11] talk about Agentic, we're truly talking [12:13] about things that are able to action and [12:16] reason on their own. They learn, they [12:18] adapt, and they make those decisions. [12:20] And not only do they do that [12:21] autonomously, they do them across [12:23] systems. So that multi-agogentic and the [12:26] protocols associated with that become [12:27] incredibly important and they do forge a [12:30] different architecture. So when we think [12:32] about that again I'll come back into the [12:34] Dell story when we started to map out [12:37] our our AI journey the first thing we [12:40] did we spent about six months to really [12:42] understand our processes within those [12:44] four areas those four use cases that I [12:47] had mentioned and the idea was you've [12:49] all heard it if you live through the [12:50] automation phases you don't want to [12:52] automate it things automating broken [12:54] processes are just going to break things [12:55] really quickly and if you put agents on [12:57] top of that you can do it at scale [12:59] really really quickly and you can spend [13:00] a lot of money doing it. So, you're [13:02] really not going to hit the ROI. So, if [13:03] you really want to focus on the right [13:05] types of use cases, you've got to get [13:06] your processes mapped out and you have [13:08] to streamline those. That was really [13:09] hard for us. We're a large company. [13:11] We're a 40-year-old company. We've made [13:13] acquisitions. We've bolted systems [13:16] together. We've b bolted data sets [13:18] together. Um I a funny story, there was [13:20] a point in time where um our COO had [13:23] said if he wanted to know how many [13:24] employees were in the company, and he [13:27] asked our HR leader and our finance [13:29] leader, he got two different answers. [13:31] That's a true story. And again, sharing [13:34] in the spirit of of of sharing what [13:36] we've learned through our journey, those [13:37] are the types of things that we we are [13:39] fixing. We've gone back into our data [13:41] sources and really clean them up. [13:42] There's some manual work that has to be [13:44] done with that. We did that as part of [13:46] our generative AI journey to really get [13:49] the right use cases and the efficiency [13:50] that we needed. Now that we think about [13:53] applying agentic to that, we're kind of [13:55] rethinking these all over again because [13:57] what you don't want to do is map out a [13:59] process and apply agentic processes to [14:02] your existing processes for the way [14:04] things were defined in a human world. [14:07] What you want to do is rethink those if [14:09] you had agents in the mix. So [14:12] specifically within our sales [14:13] organization, [14:15] we've we've mapped out the way that our [14:17] sellers engage with customers and we [14:19] want to use Agentic to make that better [14:21] for our customers to give a better [14:23] experience to enable our sellers to [14:25] spend more time with their customers and [14:27] to have an easier time getting the right [14:29] information and ultimately to drive [14:31] growth in those customer engagements. [14:34] And so if you think through what that [14:35] looks like in agentic, you might apply [14:38] one set of agents to the process. But [14:40] what we did instead is sort of rethought [14:42] that and started with if we started over [14:45] today what what is the work that needs [14:47] to be done and if I knew Agentic existed [14:50] when I define that, which I now do, how [14:53] would I define it differently? And I [14:55] might have an agent that was responsible [14:58] for coordinating the engagements back to [15:00] a customer, like literally the the [15:03] scheduling. I might have an agent that [15:05] curated content and gave it to the [15:08] seller so that they could determine when [15:10] and how to engage the customer. I might [15:12] have an agent that would go through and [15:14] actually do the pricing and the quoting [15:17] back and forth with the customer. [15:18] There's some facets that just make a lot [15:20] of sense that you would push off to [15:22] Agentic Systems and they would be [15:25] multiple agents that would be working [15:26] together not to replace the seller or [15:29] the customer but to enable them to be [15:32] more effective more quickly. And so [15:34] that's the way that we're thinking about [15:36] Agentic across the company. It becomes [15:38] something that augments the teams. And [15:40] by the way, where I see really great [15:42] successes is if you lean into things [15:44] that not only are going to drive [15:45] business outcomes, but they're things [15:47] that the people in your organization [15:48] don't want to do. We had a um an [15:52] internal study that we had done and we [15:54] pull some data together that we send out [15:56] to our senior leaders daily and we had a [15:59] team of two people that used to do the [16:01] research on this daily, pull it [16:03] together, clean it up and send it out to [16:05] that leadership team. that now gets done [16:07] through a multi-aggentic system in like [16:10] two hours and so the time it saved and [16:13] those were things that the team they're [16:15] still there they're doing different [16:16] things now that are higher value so I [16:19] think there really is a good story [16:20] around mapping out the right processes [16:22] and the right areas of focus within [16:24] Agentic in particular [16:26] >> yeah um amen I mean you know I think you [16:30] you said it well I'm not sure I can add [16:32] a lot to that um no but but again the [16:35] point about thinking about how you [16:36] implement a gentic in sort of a net new [16:38] fashion as opposed to you know thinking [16:40] of it as sort of a bolt-on that that's [16:42] actually a very important concept and [16:44] it's it's also again important to think [16:45] about not just this project but to try [16:47] and think in terms of a roadmap and it [16:49] is in some ways a little bit hard to [16:51] know what's just around the corner I [16:52] mean we are making advances so quickly [16:54] uh not only in terms of platform [16:56] capabilities but also an algorithmic [16:57] innovation especially on the inference [16:59] side of things examples being things [17:01] like speculative decoding which which [17:03] holds great promise in terms of you [17:05] separating pre-fill and decode into [17:07] different operations because each one [17:09] has sort of a different performance [17:10] sensitivity as one example. Uh, another [17:12] one is disagregated. I'm sorry. Yeah. [17:14] And the other one is u I'm sorry that's [17:16] actually disagregated serving. Sorry, I [17:18] got my wires crossed. Uh, but then also [17:21] speculative decode is another one where [17:22] again we're looking at ways to do the [17:24] decode uh phase faster and we're [17:26] implementing those things into our [17:27] software libraries which then find their [17:29] way into the Dell platform. So you know [17:32] it's hard to see around the corners but [17:34] do be thinking about you know and this [17:36] is why it's important to kind of as best [17:38] we can you know using uh agentic and a [17:40] generative AI to really keep scanning [17:43] and understanding what are what's in the [17:44] research pipelines you know we have a [17:46] deep research team here at Nvidia who's [17:48] constantly probing and saying what can [17:50] we do next how can we take this to that [17:52] next level and you know there's sites of [17:53] course like archive which is hosted by [17:55] Cornell where a lot of papers get [17:56] published and just sort of keeping that [17:58] eye on on what's the next round of [18:00] innovation that's going to enable that [18:02] next level of capabilities and to try [18:05] and to the extent you can look that far [18:07] ahead because again you don't want to [18:09] have to come back to this 3 years from [18:10] now and go a boy did we shoot behind the [18:12] duck right so things to think about as [18:15] you're going through the planning cycle [18:16] of taking those existing processes and [18:19] trying to kind of bring them into that [18:20] agentic era do try to think about the [18:22] there's what we're going to do in rev [18:23] 1.0 versus what do we want to do in rev [18:25] 1.5 2.0 and beyond because you may not [18:28] get it all into rev 1.0 I know I mean [18:30] that's a very common story in software [18:32] development right so um so do be [18:34] thinking in terms of you know roadmap [18:36] and as best we can you know we have to [18:39] we all have to continue studying you [18:40] know it's just this constant learning [18:42] that we're all doing around what can we [18:44] do with AI 6 to 12 months from now [18:46] because we all know it's going to change [18:48] uh and be that much more capable within [18:51] even that short period of time [18:54] >> all right well Mindy and Dave thanks [18:56] again for your opinions here great [18:57] insights into how both Dell anybody you [19:00] know approached Agent AI which is [19:01] important for all your organizations and [19:03] businesses in terms of your strategy and [19:05] getting you the fast performance as well [19:07] as giving you that confidence that Agent [19:08] AI is going to deliver you know better [19:12] performance and insights here's a quick [19:14] slide on on our booth tomorrow and [19:16] please come by and visit us while we [19:17] don't have much time for questions you [19:19] know you're welcome to engage Dave and [19:21] Mindy right outside this session for any [19:23] additional questions and we can move on [19:24] from there thanks again for attending [19:26] >> thank you [19:27] >> thank you everybody