Nvidia Accelerates Supply Chain Replanning Cycles to Seconds With Agentic AI

NVIDIA利用智能体AI将供应链重规划周期缩短至秒级 将GPU加速优化求解器与智能体推理引擎相结合，把复杂的供应链情景规划时间从数小时缩短至数分钟，从而在高管规划会议期间实现实时决策。 尽管拥有海量数据，但受制于复杂的运营权衡、碎片化的数据以及行业特有的约束，AI在供应链规划中的渗透率仍然显著不足。NVIDIA通过其AI Planner架构解决了这些障碍——该架构由NVIDIA Neotron推理引擎和cuOpt GPU加速数学求解器提供支持，用于执行复杂的线性规划和混合整数规划模型。通过将自然语言界面与高速优化相结合，供应链团队可以将情景规划周期从长达一整天缩短至数分钟或数秒。这使得在高管S&OP（销售与运营规划）会议期间进行现场实时情景分析成为可能，并大幅提升了整体运营敏捷性。 与软件工程等其他领域相比，AI在供应链规划中的渗透率严重不足，即便在大型科技峰会上也极少有专门的探讨。 (0:04) 由于现实世界中严格的物理约束，以及相较于人类基准评估智能体性能时存在模糊性，将自主智能体应用于实体供应链带来了独特的挑战。 (3:05) 需求规划、库存和供应分配等各个分散的数据孤岛缺乏自主AI决策所需的语义互联元数据和上下文本体。 (7:05) NVIDIA依赖完全外包的无晶圆厂制造模式，形成了涵盖晶圆代工、封装和高带宽内存（HBM）的复杂多级供应约束。 (8:30) NVIDIA开发了专有的GPU加速优化求解器cuOpt，其在线性规划（LP）和混合整数规划（MIP）上的性能比主流CPU求解器快10倍以上。 (12:00)

Russell Halper of Insight Kitchen argues that artificial intelligence remains severely underpenetrated in enterprise supply chain planning, while NVIDIA's Anthony Noguera demonstrates how combining proprietary mathematical solvers with agentic reasoning can compress multi-hour planning cycles into minutes. The Supply Chain AI Adoption Deficit (0:04) Russell Halper states that artificial intelligence has seen far less adoption in supply chain planning than in domains such as software engineering. Halper notes that an informal search of the Ai4 conference program yielded only two supply chain sessions out of more than 1,000 speakers, underscoring how many supply chain leaders struggle to begin implementing the technology. The discussion focuses specifically on planning tasks—such as forecasting, capacity, and allocation—rather than physical logistics or execution. Operational Trade-Offs and Architectural Obstacles (3:05) Halper highlights that applying agentic AI to supply chain planning introduces distinct challenges because physical supply chains directly underpin societal essentials, from fuel networks to data centers. Unlike traditional data science, where models are measured against clear mathematical benchmarks, evaluating an agent operating at approximately 90% of human capability involves ambiguity. Organizations face a strategic dilemma regarding whether autonomous speed introduces systemic risks if bad decisions are executed at scale across varied industries. Technically, organizations remain hindered by fragmented islands of data spanning demand forecasting, inventory, and supply allocation. Halper notes that companies struggle to choose between single-vendor monolithic platforms and modular best-of-breed architectures. Furthermore, even when enterprise data is centralized in cloud repositories, it often lacks the contextual ontology, operational logic, and metadata necessary for autonomous agents to make context-aware decisions. NVIDIA's Fabless Constraints and AI Planner (8:30) Anthony Noguera details NVIDIA's fabless manufacturing model, which relies on external partners such as TSMC for wafer manufacturing and third-party contractors for die packaging, assembly, and testing. This global footprint faces critical bottlenecks, including high bandwidth memory (HBM), interposers, and substrates. To resolve dynamic constraints involving multi-level bills of materials and yield changes, NVIDIA moved away from legacy software like SAP to develop its internal AI Planner architecture. Accelerated Solvers and Live S&OP Meetings (16:23) The AI Planner pairs the NVIDIA Neotron reasoning engine and specialized agents with cuOpt, a proprietary linear and mixed-integer programming solver that Noguera reports runs more than 10 times faster than leading CPU-based alternatives. Through conversational queries or custom dashboards, the system automates intent selection, side-by-side scenario modeling, and executive reporting. Operational scenario runs that historically required 15 hours to a full day of manual effort across global teams now execute in 10 to 15 minutes, with NVIDIA CEO Jensen Huang noting that replanning cycles take seconds. This rapid turnaround allows leadership to run complex allocation scenarios live during Sales and Operations Planning (S&OP) meetings instead of waiting for offline iterations. While Noguera cautions that broader service metrics like on-time delivery cannot be credited to a single platform alone, the tool serves as a primary contributor to responsiveness. Noguera and Halper advocate for an incremental, modular adoption of available AI tools rather than waiting for an entirely autonomous end-to-end architecture. While NVIDIA demonstrates significant cycle-time reductions by pairing optimization solvers with reasoning agents, Halper points out that the broader industry faces persistent obstacles, including ambiguous evaluation metrics for agents and fragmented enterprise data lacking business context.

Insight Kitchen创始人Russell Halper与NVIDIA业务应用AI总监Anthony Noguera指出，尽管人工智能在软件工程等领域已实现规模化应用，但在企业供应链计划领域的渗透严重不足；通过将混合整数规划求解器与大语言模型推理智能体相结合，NVIDIA成功将复杂的供应链排产模拟时间从原先的数小时乃至一整天缩减至数分钟内，推动产销协同从滞后的离线测算迈向实时决策。 供应链智能体的落地挑战与评估困境 (3:05) Halper指出，供应链计划领域（如需求预测、产能规划与资源分配）的AI渗透水平极低，在拥有上千名演讲者的Ai4大会中，涉及供应链的专题仅有两场。物理供应链直接维系着社会运转中的能源、消费品和数据中心基础设施，其运作对现实扰动的敏感性显著高于抽象的纯软件系统。这使得引入自主智能体所伴随的业务权衡更加严峻。Halper强调，传统数据科学具有清晰的量化评价标准，只要预测模型的数学指标胜过人工基准即可确认其有效性；然而智能体系统的评估则具有明显模糊性，若智能体达到人类90%的水平，企业不仅难以精确衡量剩余10%的缺陷风险，更面临加速决策以便快速纠偏与错误决策被大规模执行之间的两难选择。此外，不同产业的物理特性差异巨大，消费品高周转的分销模式与半导体制造动辄数十亿美元的资本支出及严苛产能瓶颈截然不同，无法直接套用统一的通用模型。 企业数据孤岛与业务语义上下文缺失 (7:05) 在技术架构层面，阻碍智能体落地的主要障碍在于企业内部需求规划、库存管理和供应分配等职能之间长期存在的数据孤岛。Halper分析称，企业在技术选型上面临两难困局：究竟是迁移至单一供应商提供的大型互联规划平台，还是采用针对特定场景的最佳模块化组合架构；同时，企业普遍难以平滑连接宏观规划系统与底层执行软件。更为核心的症结在于，即使企业将各个系统的数据统一清洗并归并至云端存储，这些数据通常依然缺乏必要的业务本体模型、行业运作逻辑以及语义关联元数据。自主智能体如果无法理解数据背后的业务情景与约束关系，便无法做出符合真实运营环境的决策判断。 NVIDIA无晶圆厂网络与AI规划架构 (8:30) Noguera介绍了NVIDIA庞大且复杂的全球供应链运作体系。NVIDIA采用完全外包的无晶圆厂芯片制造模式，晶圆代工依赖台积电（TSMC）等合作伙伴，后续的晶圆切割、测试与先进封装则由专业外协厂商承担，生产流程受制于高带宽内存（HBM）、中介层（Interposers）和封装基板等关键物料的产能供给。为应对多层级物料清单、动态波动的良率以及替代料调配等复杂约束，NVIDIA逐步从SAP等传统企业级平台转型为自主研发的优化架构，推出了数学求解器cuOpt，在线性规划与混合整数规划上的求解速度较市场主流CPU求解器提升10倍以上。在此基础上，NVIDIA构建了以Neotron推理引擎为核心的AI Planner，整合NeMo Retriever检索管道、外部工具与沙盒环境，调度专门的任务优化、深度分析与自动化编码智能体，不仅执行数据检索，更负责推导端到端的排产决策。 产销协同规划从离线测算迈向即时交互 (16:23) 在实际业务操作中，AI Planner通过定制的React仪表板和基于自然语言的对话界面交付，涵盖从意图识别与技能选取、多场景比对、自动化洞察报告到动态生成交互看板的完整自动化工作流。Noguera说明，以往跨全球多个计划团队协同完成一次复杂的供应链情景推演需要花费15个小时乃至一整天的人工劳动，现在缩短至10到15分钟；NVIDIA首席执行官黄仁勋（Jensen Huang）评价称，针对大规模复杂供应链的重新规划目前仅需数秒即可完成。这一时效突破改变了传统的销售与运营规划（S&OP）流程，决策层得以在会议进行过程中实时输入变动参数，动态推演台积电晶圆分配等关键产能瓶颈方案，而无需在会后经历漫长的数据重新录入与离线测算等待。 构建完全自主的端到端智能体供应链仍是一项长期工程，且订单交付率和准时履约率等核心业务指标的维系受多重客观运营因素影响，无法单独归功于某单一系统。然而NVIDIA的落地实践表明，企业无需等待未来理论技术的突破，现阶段通过模块化引入现有的数学求解引擎与对话式AI智能体，逐步改造具体业务痛点，便能在真实的产销协同中获得可观的效率与响应敏捷度。

## Background 本视频由 Ai4 频道发布于 YouTube 平台（标注发布日期为 2026 年 8 月 21 日），由 Insight Kitchen 创始人 Russell Halper 与 NVIDIA 业务应用 AI 总监 Anthony Noguera 联合呈现。视频简介指出，供应链数据的获取已不再是核心难题，关键在于如何将其转化为果断且优化的决策行动；演讲通过 NVIDIA 的真实业务实践，展示了智能体如何跨越预测、产能约束与异构系统孤岛，推动 AI 从被动的分析报告工具转变为深度参与战略决策的智能伙伴。 ## TL;DR 演讲首先剖析了智能体（Agentic AI）在供应链计划领域落地艰难的三大症结：成效量化与决策风险高、行业链条异质性极大，以及数据孤岛与业务上下文的缺失。针对这些痛点，NVIDIA 展示了其为应对极其复杂的半导体代工与封装测试链路所打造的端到端“AI Planner”架构，该系统深度融合了自研 GPU 加速求解器 cuOpt、Nemotron 推理大脑、技能库以及多 Agent 协作网络。通过这套方案，NVIDIA 将原本耗时 15 小时的复杂跨团队假设排程分析缩短至 10 到 15 分钟，甚至实现数秒级的重排计算。最后，讲者指出运筹优化求解器与 AI Agent 的结合使得在产销协同（S&OP）会议现场进行实时情景模拟成为可能，呼吁企业采取模块化、渐进式的策略在当下创造切实的业务价值。 ## Atmosphere & style 整场演讲洋溢着顶级科技峰会兼具专业度与实操深度的硬核氛围，风格务实、条理清晰且节奏紧凑。Russell Halper 以运筹学专家的敏锐与幽默开场，直面供应链领域 AI 渗透滞后的痛点；Anthony Noguera 随后以服务 NVIDIA 24 年的深厚积淀，冷静从容地拆解先进半导体制造中的重重物理瓶颈与底层技术架构。两位讲者的互动默契流畅，从技术挑战的真实困境层层推进至秒级仿真重排的震撼成效，让听众在严谨扎实的工程逻辑中切身感受到供应链决策范式被彻底重构的敏捷与冲击力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=4s) 相较于软件开发等爆发式落地的领域，AI 在供应链计划中的渗透率仍然严重不足，缺乏清晰的落地路径。 - [03:05](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=185s) 供应链计划难以落地 Agentic AI 的首要障碍在于评估标准模糊与决策风险，无法单纯套用传统数据科学的准确率指标衡量辅助决策的成效。 - [03:05](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=185s) 供应链类型差异巨大（如短链快消与长周期重资产半导体），且企业普遍面临需求、库存和供应系统割裂与数据本体模型（Ontology）缺失的问题。 - [08:30](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=510s) NVIDIA 具备极度复杂的 Fabless 半导体供应链网络，涉及晶圆、封测、HBM 显存、基板等重重物理约束以及动态良品率，传统优化工具难以承载。 - [11:56](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=716s) NVIDIA 打造了基于 GPU 加速的运筹优化求解器 cuOpt，其 LP/MIP 求解性能相比市场上最快的 CPU 求解器提升了 10 倍以上。 - [11:56](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=716s) NVIDIA AI Planner 系统以 Nemotron 推理引擎为大脑，集成了 NeMo Retriever 结构化检索、MCP 外部协议与沙盒安全运行环境，构建了多 Agent 协同体系。 - [16:23](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=983s) 系统结合了现代化 React UI 与对话式交互，通过意图解析、情景模拟、智能报告与共享仪表盘实现四步闭环决策编排。 - [16:23](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=983s) AI Planner 将过去耗时 15 小时甚至一整天的复杂供应排程假设分析压缩至 10 到 15 分钟，支持在高管 S&OP 会议中即时推演决策。 - [19:25](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=1165s) 黄仁勋评价指出，超大规模复杂供应链的重排仿真已实现秒级计算，极大释放了计划员生产力并提供了应对波动的敏捷性。 - [20:19](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=1219s) 运筹优化与 AI Agent 的深度融合彻底重塑了 S&OP 会议模式，使其从离线等待数天的低频评审演变为现场秒级交互推演。 - [20:19](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=1219s) 企业实施 Agentic 供应链无需追求一步到位，应采用模块化与渐进式落地策略，利用当下成熟可用的 AI 与优化工具切实解决业务痛点。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=4s) 供应链中 AI 应用的现状与演讲背景 Insight Kitchen 创始人兼董事总经理 Russell Halper 开启了演讲，向在场观众致谢。在询问现场有多少人从事供应链工作后，他提出了一个核心观点：相较于软件开发等爆发式落地的领域，AI 在供应链中的渗透率仍然严重不足，许多业内人士甚至不知道该如何迈出第一步。本次分享聚焦于“智能体（Agents）如何应用于供应链计划（Supply Chain Planning）”，并联合 NVIDIA 展示前沿落地实践。 演讲的主线内容包括：讲者背景、AI 在供应链计划中的困境与机遇、NVIDIA 内部正在构建的创新能力，以及关键启示。 - **Russell Halper**：拥有混沌理论数学博士学位，在商学院接触运筹学后深耕供应链与数据科学领域。曾任职于雀巢，后加入 End-to-End Analytics 成为合伙人（该公司后被埃森哲收购）。在埃森哲担任董事总经理期间，恰逢 ChatGPT 问世，他受命牵头组建了西海岸 GenAI 业务线；数年前创立专注 AI、供应链及收益管理的咨询机构 Insight Kitchen。 - **Anthony Noguera**：NVIDIA 业务应用 AI 总监，拥有加利福尼亚大学伯克利分校工业工程硕士学位。在 NVIDIA 主导将 AI 深度融入极端复杂的供应链场景，推动了包括文本优化及前沿求解器在内的诸多创新。 --- ### [03:05](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=185s) 供应链计划为何难以落地 Agentic AI Russell 指出，在拥有上千位演讲嘉宾的 Ai4 大会议程中，搜索“供应链”竟然只能检索到两场演讲，这印证了 AI 在该领域的渗透滞后。现代社会的运转离不开供应链，从原油波动、疫情冲击到数据中心硬件，处处受其影响。供应链虽然是高度量化、数据驱动的领域（运筹学与预测正是数据科学的重要源头之一），但引入 Agentic 能力依然困难重重： 1. **评估标准模糊与决策风险**：传统数据科学（如需求预测）目标明确，只需产出准确率优于人类专家的数值指标即可。但当引入智能体替代或辅助人工决策时，若 Agent 的表现只有人类的 90%，该如何量化衡量？如果决策速度大幅提升，是在帮助企业快速修正方向，还是在以极高速度做出更糟糕的决策？ 2. **供应链类型差异巨大**：快消饮料的短链高频低货值流通，与半导体制造动辄巨额资本投入、长跨国周期、强产能约束的运作逻辑天差地别，不存在通用的一刀切方案。 3. **系统与数据孤岛**：需求计划、库存管理和供应计划往往割裂在不同技术架构中，企业在一体化平台与最佳单项组合（Best-of-breed）之间反复摇摆。执行层与计划层系统分立，即使做了大规模数据清洗与归一化，仍普遍缺乏支撑 AI 规模化运作所需的业务上下文与本体模型（Ontology）。 --- ### [08:30](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=510s) NVIDIA 复杂的半导体供应链全景 Anthony Noguera 分享了自己在 NVIDIA 任职 24 年的历程。他最初从 SAP 加入 NVIDIA，负责基于 SAP 搭建供应链系统。十年前，NVIDIA 敏锐意识到用于游戏图像渲染的并行计算架构正是 AI 计算的基石，从而推动公司转型为全球 AI 训练与推理基础设施的核心支柱。 ``` [ 晶圆代工 (如 TSMC) ] │ 制造 Wafer ▼ [ 晶圆切割与测试 ] (外包代工伙伴) │ Die 切割 / 测试 ▼ [ 芯片级封装 ] ── 整合高带宽内存 (HBM)、中介层 (Interposer)、基板 (Substrate) │ ▼ [ 板卡与模组制造 ] (合同制造伙伴，如 Turnkey Modules / Boards) │ ├───────────────────────────────┐ ▼ ▼ [ 数据中心 AI 级系统 ] [ 消费级游戏显卡 ] (如重型整机机柜) ``` NVIDIA 是纯外包制造（Fabless）模式的开创者： - 从台积电（TSMC）的晶圆制造开始，经由外包封测厂商进行晶圆切割与测试； - 将 Die 与高带宽内存（HBM）、中介层（Interposer）、基板（Substrate）封装为芯片； - 再由代工伙伴制成板卡或 Turnkey 模组，最终分流至消费级游戏显卡或重达数吨的大型 AI 数据中心整机。 - 整个链路中处处存在物理瓶颈（晶圆、封测产能、HBM 显存、中介层与基板），叠加多层级复杂物料清单（BOM）、随时间变化的良品率（Time-based Yields）、物料互换性及跨据点调配，复杂程度远超传统传统供应链优化软件的处理能力。 --- ### [11:56](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=716s) NVIDIA AI Planner 的系统架构与核心能力 面对复杂供应链挑战，Anthony 带领团队与内部软件工程组织合作，先研发了自研运筹求解器 **cuOpt**（支持 LP 与 MIP，GPU 加速性能达市场上最快 CPU 求解器的 10 倍以上），随后在近两三年内构建了端到端的 **AI Planner**。 ``` ┌──────────────────────────────────────────────────────────┐ │ 输入通道 │ │ - 结构化数据 (NVIDIA NeMo Retriever + SQL/NL 检索) │ │ - 交互提示词 (对话式 Chat 界面) │ │ - 基础大模型接入 (Claude / ChatGPT / 开源 LLM) │ │ - MCP 外部工具连接 & 沙盒安全测试运行环境 │ └────────────────────────────┬─────────────────────────────┘ │ ▼ ┌──────────────────────────────────────────────────────────┐ │ AI Planner 核心大脑 (NVIDIA Nemotron 推理引擎) │ │ - 负责业务推理与决策规划，非单纯检索数据 │ └──────────────┬─────────────────────────────┬─────────────┘ │ │ ▼ ▼ ┌───────────────────────────┐ ┌──────────────────────────┐ │ 技能库 │ │ 持久化记忆与任务 Agent │ │ - 产品配额分配 (Allocation)│ │ - 优化 Agent (cuOpt 等) │ │ - 端到端约束供应规划 │ │ - 深度推理 Agent │ │ - cuOpt LP/MIP 求解器 │ │ - 代码生成 Agent │ └───────────────────────────┘ └──────────────────────────┘ ``` 该系统并非单纯的聊天机器人，而是一个完整的决策编排层： - **核心大脑**：采用 NVIDIA Nemotron 推理能力，负责分析排程决策。 - **输入集成**：支持 NVIDIA NeMo Retriever 驱动的自然语言/SQL 混合结构化查询，通过模型上下文协议（MCP）对接外部系统，并配备沙盒（Sandbox）确保代码与新方案在上线前安全测试。 - **技能库（Skills Library）**：AI 能力扩展的核心，涵盖端到端产能约束规划、物料分配，以及核心加速引擎 cuOpt。 - **任务分发**：结合持久化记忆（Persistent Context Memory），将复杂任务动态编排并分派给优化 Agent、深度 Agent 及代码 Agent 协同完成。 --- ### [16:23](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=983s) 交互工作流与端到端效益：从 15 小时到 10 分钟 在具体操作层面，AI Planner 同时保留了广受内部好评的现代化 React UI 与类似 ChatGPT 的自然语言交互界面。系统在底层自动执行四步闭环： 1. **意图解析与技能调用（Query Refinement & Selection）**：自动识别用户诉求，匹配对应技能与底层数据源，无需用户手动指定数据集或运筹工具。 2. **场景模拟与假设分析（Scenario & What-if Analysis）**：支持多方案并排对比（Side-by-side），验证不同决策路径。 3. **洞察生成与智能报告（Reports & Insights）**：将求解器算出的原始技术数据提炼为可读性强的业务摘要。 4. **共享仪表盘（Sharable Dashboards）**：打破跨团队信息壁垒，实现方案的长期跟踪与组织级协同。 - **显著成效**：过去运行一次全新的复杂供应排程假设分析（例如在高管 S&OP 会议中探讨“能否向台积电争取追加特定晶圆配额”），需要各计划团队协同耗时 15 小时甚至一整天；而在 cuOpt 与 AI Planner 编排下，完整模拟被压缩至 **10 到 15 分钟**。管理层能够在同一场会议内即时运行推演、评估结果并做出决策。 ### [19:25](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=1165s) 实时重排与效能跃升：黄仁勋对供应链敏捷性的评价 在讨论系统落地的实际成效时，演讲者强调该方案以近乎实时的响应速度极大地解放了计划员（Planner）的生产力。过去，计划团队需要耗费大量时间与精力跨越多个异构数据源收集数据，并进行繁琐的手动交叉分析；而现在，这套结合了 AI 的架构将数据整合与分析耗时大幅压缩。尽管企业整体的订单履行率与准时交付率（On-Time Delivery）不能完全单一归功于某一个软件系统，但这套方案已成为英伟达持续维持卓越供应链运营表现的核心支柱。 英伟达创始人兼 CEO 黄仁勋（Jensen Huang）也对该系统的落地效果给予了极高评价，指出：“对复杂且超大规模的供应链进行重新计划，现在仅需数秒钟即可完成，这赋予了我们团队应对瞬息万变供应链所需的极致速度与敏捷性。”（"replanning complex large scale supply chains now takes just seconds, giving our team the speed and agility to respond to the evershifting supply chain."） - **计划员生产力提升**：消除了跨多系统手动汇总数据和反复迭代计算的繁重流程。 - **业务表现支撑**：在复杂的供应链波动环境下，成为稳定订单履行与准时交付的关键基础设施。 - **重排计划耗时缩减**：大规模供应链的仿真重排计算从传统的小时/天级缩短至数秒之内。 --- ### [20:19](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=1219s) 总结与落地路径：将现场模拟融入 S&OP 会议，立足当下创造价值 在演讲接近尾声、进入超时 30 秒的收尾阶段时，Russell Halper 对整个用例的核心价值进行了高度概括。他指出，这个案例最引人入胜之处在于将基于运筹优化的求解器（如 NVIDIA cuOpt）与前沿 AI 能力进行了深度融合。这种结合带来的改变不仅是“算得更快”，更在于质的飞跃——因为计算速度达到了秒级，团队可以直接在产销协同会议（S&OP Meeting）的现场进行实时动态情景模拟（Scenario Simulation），彻底告别了以往在各部门之间来回拉扯、每次调整都要等待数天反馈的漫长迭代周期。 ``` 传统 S&OP 流程: [ 会议讨论 ] ──> [ 离线数据收集与建模 ] ──(数天延迟)──> [ 下次会议评审 ] Agentic AI + cuOpt 实时流程: [ S&OP 会议现场 ] <──(秒级实时情景模拟与决策验证)──> [ AI Agent + cuOpt 引擎 ] ``` Russell 坦言，尽管构建一套完整的端到端智能体供应链系统对普通企业来说可能显得过于庞大且雄心勃勃，但其中的诸多模块化组件完全具备高度的可行性，能够轻松切入并快速落地。他最后向观众呼吁，不必过度执着于揣测技术在数年后的终极形态，企业当前更应该思考的是如何审慎且巧妙地运用现已成熟的技术能力，切实解决当下的业务瓶颈，真正“推动改变现状的齿轮”。 - **实时交互式决策**：支持在 S&OP 会议现场即时调整参数并运行模拟，实现决策即验证。 - **渐进式实施策略**：无需一次性推翻现有架构，可分模块、分阶段落地优化与 AI Agent 能力。 - **务实创造业务价值**：立足于当下成熟可用的 AI 与优化工具，快速将技术洞察转化为落地的业务成效。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=4s) 相较于软件开发等爆发式落地的领域，AI 在供应链中的渗透率仍然严重不足，许多业内人士甚至不知道该如何迈出第一步。 — Russell Halper > [03:05](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=185s) 在拥有上千位演讲嘉宾的 Ai4 大会议程中，搜索“供应链”竟然只能检索到两场演讲，这印证了 AI 在该领域的渗透滞后。 — Russell Halper > [03:05](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=185s) 当引入智能体替代或辅助人工决策时，若 Agent 的表现只有人类的 90%，该如何量化衡量？如果决策速度大幅提升，是在帮助企业快速修正方向，还是在以极高速度做出更糟糕的决策？ — Russell Halper > [03:05](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=185s) 快消饮料的短链高频低货值流通，与半导体制造动辄巨额资本投入、长跨国周期、强产能约束的运作逻辑天差地别，不存在通用的一刀切方案。 — Russell Halper > [16:23](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=983s) 过去运行一次全新的复杂供应排程假设分析，需要各计划团队协同耗时 15 小时甚至一整天；而在 cuOpt 与 AI Planner 编排下，完整模拟被压缩至 10 到 15 分钟。 — Anthony Noguera > [19:25](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=1165s) replanning complex large scale supply chains now takes just seconds, giving our team the speed and agility to respond to the evershifting supply chain. — Jensen Huang > [19:25](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=1165s) 对复杂且超大规模的供应链进行重新计划，现在仅需数秒钟即可完成，这赋予了我们团队应对瞬息万变供应链所需的极致速度与敏捷性。 — Jensen Huang > [20:19](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=1219s) 因为计算速度达到了秒级，团队可以直接在产销协同会议（S&OP Meeting）的现场进行实时动态情景模拟（Scenario Simulation），彻底告别了以往在各部门之间来回拉扯、每次调整都要等待数天反馈的漫长迭代周期。 — Russell Halper > [20:19](https://www.youtube.com/watch?v=qKyKSQQOzBk&t=1219s) 不必过度执着于揣测技术在数年后的终极形态，企业当前更应该思考的是如何审慎且巧妙地运用现已成熟的技术能力，切实解决当下的业务瓶颈，真正“推动改变现状的齿轮”。 — Russell Halper ## Mentioned Russell Halper, Anthony Noguera, Jensen Huang, Insight Kitchen, NVIDIA, Ai4, Nestlé, End-to-End Analytics, Accenture, TSMC, SAP, cuOpt, AI Planner, NVIDIA NeMo Retriever, NVIDIA Nemotron, ChatGPT, Claude, MCP (Model Context Protocol), React ## Video description Russell Halper - Founder & Managing Director, Insight Kitchen LLC Anthony Noguera - Director - AI for Business Applications, NVIDIA Accessing supply chain data is no longer the challenge; turning that access into decisive, optimized action is. In this session, I will be joined by NVIDIA to demonstrate how AI is evolving from a passive reporting tool into an active partner in strategic decision-making. Together, we will walk through a real-world deployment to explore how AI interacting directly with your supply chain transforms the way organizations manage complexity. By enabling intelligent agents to bridge the gap between forecasts, capacity constraints, and disparate data sources, teams can diagnose bottlenecks and simulate solutions with unprecedented speed. Attendees will leave with a clear vision of achievable, next-generation capabilities they can build today—moving beyond simple visibility to create a supply chain that is responsive, integrated, and decision-centric. If you're looking to transform AI from insight into orchestrated action, this session will show you how. [Source](https://www.youtube.com/watch?v=qKyKSQQOzBk)

[0:04] Russell Halper: 相较于软件开发等爆发式落地的领域，AI 在供应链中的渗透率仍然严重不足，许多业内人士甚至不知道该如何迈出第一步。 [3:05] Russell Halper: 在拥有上千位演讲嘉宾的 Ai4 大会议程中，搜索“供应链”竟然只能检索到两场演讲，这印证了 AI 在该领域的渗透滞后。 [3:05] Russell Halper: 当引入智能体替代或辅助人工决策时，若 Agent 的表现只有人类的 90%，该如何量化衡量？如果决策速度大幅提升，是在帮助企业快速修正方向，还是在以极高速度做出更糟糕的决策？ [3:05] Russell Halper: 快消饮料的短链高频低货值流通，与半导体制造动辄巨额资本投入、长跨国周期、强产能约束的运作逻辑天差地别，不存在通用的一刀切方案。 [16:23] Anthony Noguera: 过去运行一次全新的复杂供应排程假设分析，需要各计划团队协同耗时 15 小时甚至一整天；而在 cuOpt 与 AI Planner 编排下，完整模拟被压缩至 10 到 15 分钟。 [19:25] Jensen Huang: replanning complex large scale supply chains now takes just seconds, giving our team the speed and agility to respond to the evershifting supply chain. [19:25] Jensen Huang: 对复杂且超大规模的供应链进行重新计划，现在仅需数秒钟即可完成，这赋予了我们团队应对瞬息万变供应链所需的极致速度与敏捷性。 [20:19] Russell Halper: 因为计算速度达到了秒级，团队可以直接在产销协同会议（S&OP Meeting）的现场进行实时动态情景模拟（Scenario Simulation），彻底告别了以往在各部门之间来回拉扯、每次调整都要等待数天反馈的漫长迭代周期。 [20:19] Russell Halper: 不必过度执着于揣测技术在数年后的终极形态，企业当前更应该思考的是如何审慎且巧妙地运用现已成熟的技术能力，切实解决当下的业务瓶颈，真正“推动改变现状的齿轮”。

[19:25] Jensen Huang: replanning complex large-scale supply chains now takes just seconds, giving our team the speed and agility to respond to the ever-shifting supply chain. [12:00] Anthony Noguera: It's not just retrieving data, it's actually reasoning through planning decisions.

[0:04] Uh thank you so much for coming.  Really appreciate you guys taking time, [0:07] especially the end of a long day to come and  hear about uh some some of the things that um [0:13] we've been working on. Uh just out of curiosity,  how many people here work in supply chain? [0:21] Okay. So, I'm going to I'm going to make a claim.  I don't think it's going to be that controversial, [0:26] but I'm going to claim that AI is underpenetrated  in supply chain relative to other things [0:33] uh other areas and you see places like you know  uh software development it being hugely impactful. [0:39] I think a lot of the people that we talk to in  the space are trying to figure out how to get [0:42] off the starting block. So we wanted to come  and talk about uh some of the things that uh [0:49] are out there some of the capabilities around how  do you use agents in the context of supply chain [0:53] planning obviously there's other parts of supply  chain execution physical logistics right we're [0:58] going to focus on this specific area for today and  talk about what we're seeing and and some of the [1:03] capabilities that Anthony and Vidia are developing  so uh uh my company's name inside kitchens on the [1:12] menu for today is uh introductions. We'll give  you a brief background of of Anthony and myself. [1:19] Uh we'll talk about uh our view on the current  state of AI and supply chain planning and [1:24] why is it hard? Where do we see the opportunities  uh Anthony is going to talk a little bit about uh [1:30] some of the things that are being developed  in Avidia and then we'll try to give you a few [1:34] takeaways uh and do our best within the 20 minutes  to to get through it all. So uh my background uh [1:43] I started off I actually started off  my career doing a PhD in chaos theory [1:46] mathematics and I took a class in the business  school on supply chain data science what we used [1:51] to call operations research and just fell in  love with supply chain. So I spent my entire [1:54] career at that intersection of supply chain  and data science and now supply chain and AI. [1:59] Uh was internal at Nestle for a while. Uh  was originally a junior consultant and then [2:04] a partner at a company named end toend analytics  who sold it to Accenture. I was a MD at Accenture. [2:10] I helped build the West Coast genai practice.  Uh they asked me to lead when this little thing [2:15] called chat GPT was released and left a few years  ago to start Insight Kitchen which is a uh focus [2:21] consulting firm on AI and supply chain of revenue  management. I have with me uh Anthony Aguero. [2:27] Anthony has been a director at uh Nvidia. He  leads a lot of the really interesting work [2:33] about how are they uh integrating O AI and some of  the really complex problems in their supply chain [2:41] uh including things like text optimization  and other leading edge capabilities that uh [2:47] we see uh almost nowhere else. Um and  you his thought leadership has driven [2:53] a tremendous amount of value in the company  over many many years. Uh Anony's got his MS [2:57] from industrial engineering from Berkeley and uh  Anthony thank you for coming and joining me today. [3:05] So um you know one one of the things actually  before we get here I looked at the agenda for [3:10] A4. If you go to the website says there's  over a thousand speakers at the conference. [3:14] If you type in supply chain into the into the  agenda uh last I checked two talks came up. [3:22] So again going back to the theme  of we're fairly underpenetrated [3:26] relative to the capabilities that are  out there. And so why is it hard? So [3:32] supply chains I like to emphasize I think this  won't be uh revolutionary for people here but [3:37] they underpin like every aspect of modern  life right we see what's happening with [3:42] oil right now and how that affects supply chain  what happens to us now what happened in covid [3:46] right uh and there's many many examples  of supply chain disruptions that affect [3:50] all of us daily right every physical  thing we interact with the computers and [3:54] data centers all these things have supply  chains associated with them and again um [4:00] the operations and supply chain right one of the  things that I find fun about them is most of the [4:05] operations are very quantifiable it's a datari and  measurable domain uh right if I'm trying to do the [4:12] analogy I made to somebody earlier today was if  I'm creating print advertising in a newspaper and [4:17] yes I know print newspapers are not a very popular  thing these days that's a lot harder to measure [4:22] the impact of right the decisions they're making  in supply chain are generally fairly quantifiable [4:28] Uh you could also argue one of the  origins of data science is supply chain [4:34] right you look at operations research and some  of the things that were developed forecasting [4:38] right those types of things it's not the only  origin but it's a really important contributor [4:43] and yet I mean we see this time and time again few  organizations we work with really have integrated [4:50] agentic capabilities into their  supply chain in any meaningful way [4:54] so one of the reasons so let me just make analogy  here. Uh data science, uh really impactful thing. [5:04] I think most supply chains will argue they're  using it in some way. Let's pick forecasting, [5:08] right? It's a pretty uh conceptually simple  thing to think about. Obviously, you can get [5:14] infinitely sophisticated with it, but for the sake  of argument, uh we're going to build a forecast. [5:19] That forecast is going to create a prediction of  a number of how much we're going to sell. And uh [5:25] my goal is for that number to be better than the  number a person would come up with on average, [5:30] right? So we can measure the forecast accuracy  and we can do that and we could quant like [5:33] quantifiably prove it's better than a person. What  does it mean when we're now leveraging agents, [5:38] right? And what does it mean if they're trying  to replace the work a person does? Well, what [5:42] if they're only 90% as good as a person? First of  all, measuring that 90% is hard, right? Then are [5:48] we able to make decisions more quickly with  them? So if that is that gap something that [5:53] is better because I can make decisions  more quickly so I can course correct or [5:56] is it bad because I'm now making  worse decisions faster than ever [6:00] uh or is there other ways that we can leverage  it right around supporting them augmenting [6:04] and getting to work was going to get to but  it's this thing I find fairly ironic in that [6:12] data scientist so well penetrated yet something  that you can argue a lot of data scientists [6:16] are moving into AI right most data scientists  reabeled as AI I uh is still early on the bench. [6:24] Uh you also have lots of different types of  supply chains, right? So I'm going to cover [6:29] this slide quickly, but just think about average  distribution versus semiconductor manufacturing. [6:34] They're both supply chains, right? One is  a fairly short supply chain where I'm doing [6:37] high volumes of low value touches and trying  to figure out how do I get them in the market, [6:42] right? Another one is huge capital investments  over long international supply chains where [6:47] I have very constrained capacity problems  and other challenges that come into play. [6:52] And so there's not really such a thing as a supply  chain. There are many types of supply chains and [6:56] each need different types of capabilities.  Again, making it harder to think about [7:00] how agents can be adopted  in a way that's widespread. [7:05] And then lastly, when I think about this,  there are lots of different types of challenges [7:10] related just to systems and data within supply  chains, right? You have islands of data, [7:15] islands of information that exist in different  systems structurally based around the way supply [7:21] chains operate. So if you think about demand  planning, inventory management, supply planning, [7:27] you know, we've seen sort of the pendulum swing  back and forth on do we want to build a connected [7:30] planning platform with one technology and  put it all in one place? Do we want to [7:34] make it bespoke? Is there some combination of best  and breed and what does that mean as organizations [7:40] are trying to think about how do they invest in  that roadmap for durable technologies and the [7:45] you know in the age of AI where things  seem to be changing faster than ever [7:49] and then of course beyond the planning level  systems you have execution level systems [7:54] typically that are provided by different  vendors right there's oftent times [7:58] large data plays on how do we get these all the  data all harmonized sorted in one organized place [8:04] we can leverage that even those after you  look at all that don't necessarily layer in [8:09] some of the context ontology and things that you  might need to effectively leverage AI at scale. [8:15] So there's lots of challenges. I don't think  there's uh hopefully it's not news to anybody [8:19] here. But uh with that said, I'm going to hand  it off to my colleague uh and have him talk about [8:25] some of the way things that Nvidia is building  to be able to uh tackle some of these challenges. [8:30] Thank you, Russell. So um yeah so as of today  I celebrated my 24th anniversary at Nvidia. [8:38] I I came in I was brought I came in from  SAP. So I I was brought into uh to bring up [8:44] uh um Nvidia's uh supply chain on SAP. So back  back long time ago we were mainly a a gaming cards [8:53] company right but like about a decade ago um we we  realized that the parallel processing architecture [9:01] that uh that's used for rendering graphics  uh rendering um rendering for games is a is [9:07] a great foundation for for AI. So now uh like the  company really has just transformed into now the [9:15] essentially the majority of all the  AI training and inferencing uh in the [9:20] world uh in data centers across the world  is is mostly using Nvidia graphics card. [9:27] So uh first let me actually go in there. [9:32] So that's what I was talking about. So now  I'm going to focus on the Nvidia supply chain. [9:39] And so this is this is a just like a very  oversimplified representation of our supply chain. [9:45] So we start off with uh with wafer manufacturing  uh we rely on we're a purely outsource uh company. [9:52] So when we started off uh we we  essentially were the very very first one to [9:56] to uh to invent that. Prior to that, Intel had its  own uh fabs, right? Had its own manufacturing uh [10:04] facilities versus but we we started  off uh with outsource manufacturing. [10:08] So, we rely on um our foundry partners such as  DSMC to do the wafer manufacturing to for us. [10:14] So then and then this uh this uh the die is  uh is now then um uh assembled uh it's it's [10:22] cut and it's tested uh again by other foundry  uh other um contract manufacturing partners [10:29] and then the the d assembled into chips and this  these chips are like uh essentially the main [10:36] components of the chips are are the memories such  as HBM high high bandwidth memory the interposers [10:42] and substrates that are used for packaging and  uh turnkey modules uh that are manufactured by [10:47] our contract manufacturing partners and uh these  these chips are now built into boards. Now some [10:53] of the chips we sell by themselves and and then  and then uh some of them are built into boards [10:59] and then these boards are now uh kind of like [11:03] um split into either going to data  center AI or uh or uh or consumer gaming. [11:11] So it's worth noting like like all throughout  the whole uh process we at any given point [11:16] in time we we we run into constraints right so  whether it's wafers uh or uh assembly capacity [11:25] not you heard about memory constraints it's  affecting um the other companies as well why [11:29] your iPads are more expensive I think and then  um and then but like interposers and substrates [11:35] can can also become like a big  constraint for us so this is really [11:39] just a very it's oversimplified I view the real  network as many many more tiers but it captures [11:45] the path of our supply chain from raw silicon to  uh to the GPUs powering the AI infrastructure. [11:56] All right. So I mean in terms of the history  like uh we went through the whole u the whole [12:01] journey of uh implementing maybe the uh you know  legacy solutions not just SAP some other legacy [12:08] legacy u software companies as well like uh to to  manage our our planning um for our supply chain [12:17] and uh but uh in the in the the last couple of  years I actually uh pivoted my career into like I [12:24] love the SAP world business applications world and  uh started working with our software engineering [12:30] organizations to first build build uh our own  proprietary we call it co-opt LP and uh MIP [12:40] solvers and also uh and working with them to to  implement them within Nvidia first the solvers [12:48] and then also the the whole AI layer around it and  uh in the span of the last uh two or three years [12:54] we built this AI I planner. Um, so basically  this is the this is the architecture. So at [13:00] the at the center is the uh is is the AI planner  powered by our own NVIDIA Neotron reasoning. [13:06] So think of it as the brains of the system. It's  not just retrieving data, it's actually reasoning [13:11] through planning decisions. So feing on the  left we have these uh four four input channels. [13:18] There's the structure data through our Nvidia  Nemo retriever with SQL and natural language [13:23] query support. uh planners can also type the text  prompts chat GPT style directly and we actually [13:31] have integrated these foundational models. So  we did have to recreate our own our own LLMs uh [13:38] uh in in our AI planner. We we integrated the  clouds and the and the the chat GPTs as well. [13:45] So, and then we've connected these external  tools via MCP and then there's this sandbox [13:49] environment so that whenever the new plans or  code uh are generated, they can be tested safely [13:56] before they introduced into production. So,  underneath all of that is a what we call a [14:02] skills library. So, that that's really eventually  how you grow your uh your AI by by creating skills [14:09] and developing these skills. and uh we're really  scratching the surface. So we've we focused on [14:17] on these skills uh around some of the more  important um supply chain use cases such as [14:22] product allocation end to end planning for  for maximizing our supply given constraints. [14:28] Um and then part of the skills is this co-op uh  solver that I mentioned to you. It's a proprietary [14:34] solution that uh I mean by itself is really world  class now in LP like uh performing 10 times or [14:42] more uh than than your fastest uh CPUbased solver  in the market. So um [clears throat] so on the [14:49] right is uh the planner draws on memory for  persistent context and then delegates the work [14:56] out to the specialized we call task agents. So  there's the optimizer, the deep agent um and then [15:03] the the coding agent. So basically it's not just  uh doing things by itself but it's orchestrating. [15:09] So why does this matter like uh our supply chain  isn't simple. uh so you see on the right side [15:15] so the be besides uh besides the the flow that  I went through earlier there there all these [15:21] considerations like the multi-level bills  of materials the many many layers of deep [15:26] uh we don't have a lot of parts but our parts are  are super complex we're building these uh big data [15:32] centers like uh like weighing like you know a  lot of I it's tons of I can't remember exactly [15:39] how how much it weighs it's like massive massive  data centers So uh we have uh we have to deal with [15:45] uh timebased uh yields. So some of the some of  the some of the challenges that your traditional [15:52] uh traditional supply chain optimization solutions  today like uh really struggle to handle uh are all [15:58] we were able to handle the interchangeability  and then location and transferability. [16:06] If if it helps I'll just say Anthony  is dramatically understating the [16:09] complexity. So, I think it it's been a 20-minute  talk here and still barely scratched the surface [16:16] of uh the challenges that are there. Mhm.  All right. Uh so, uh now let's look at the [16:23] uh what's inside. So, this is these are just I  know it's probably hard to read unfortunately, [16:28] but uh these are just screenshots of uh how  the what the what the AI planner looks like. [16:33] So you can see on the left side that uh that  the planner uh has the option we still we still [16:37] provide these traditional react based UI which  arguably which people at NVIDIA love like compared [16:45] to the to the UIs that they're used to with their  with the with the legacy legacy software systems. [16:51] uh we and then we also give them the  ability the the option to uh to to [16:55] uh interact with AI planner chat GPT style by  uh by uh entering a text based uh uh a question [17:04] right and then um so then then there are  four main things happening under the hood [17:08] there's the query refinement and selection so  this the system interprets the user's intent [17:14] and selects the right skill or model to apply  and pulls in the appropriate data source [17:19] automatically so that users or need  to know which tool or data set to pick [17:24] and then there is the uh scenario uh and what if  analysis. So once the um queries understood the [17:31] users can model different scenarios and compare  outcomes side by side which is core to planning [17:36] and optimization use cases and then the third  the planner generates uh reports and insights. [17:43] So turning the analysis into a digestible summary  rather than leaving users to interpret the raw [17:49] output. And finally everything can flow into  dashboards uh that can be shared uh so that [17:55] these insights don't get siloed and they can be  distributed across teams and tracked over time. [18:00] So the through line here is that AI planner is  not just a chatbot. It's an orchestration layer [18:06] that takes a plain language task  and turns it into refined queries, [18:10] scenario modeling, rep reporting and  sharable dashboards all in one flow. [18:16] So I wanted to share here some of the uh some  of the documented benefits that we have uh we [18:21] have uh realized uh through throughout this  whole program. So um like uh let me just first [18:29] kind of like stress that like uh implementing  implementing our own uh solver uh plus creating [18:37] these more um more efficient modeling capabilities  like have really reduced the the the cycle time [18:46] like what what would take uh let's say 15 hours  or almost a day to run through new scenarios [18:52] can run in like 10 to 15 minutes. So scenarios  can be can be can happen within an executive [18:59] SNOP meeting for example right what used to be  you [snorts] know like an executive says uh I want [19:04] to uh I want to run the scenario where I can talk  to TSMC and increase our wafer for this product [19:10] like uh it would take like a whole day for all  the planning teams to to generate the output [19:16] but with AI planner there's this capability to  just run it through within the meeting itself [19:21] and for the for the executives to be able  to make the decision s in in in real time, [19:25] almost in real time. And uh so [clears throat]  that has certainly increased the planner [19:31] productivity because uh like uh what used to to  have to take the the planners to to gather data [19:38] and do to run analysis using multiple different  sources uh uh is uh is reduced and finally it [19:46] like I we cannot obviously just fully attribute to  a system solution the improvement of the the let's [19:53] say the the the fulfillment of the orders or the  the on-time delivery Okay. But it's a it's a major [19:59] contributor for us to be able to to sustain that.  And we were able to get this quote from Jensen, [20:05] our CEO and founding CEO that uh replanning  complex large scale supply chains now takes just [20:12] seconds, giving our team the speed and agility  to respond to the evershifting supply chain. [20:19] So I think we're I think we're 30 seconds  past our time. So I'll just say quickly [20:25] it's it's to me a really interesting  use case. How do you take [20:28] uh optimization based solutions in this case  Q opt right be able to integrate them with AI [20:34] be able to ask run the scenarios not more  quickly sorry more quickly not only more quickly [20:39] but also because you can run them more quickly  you can run them more you run them live in the [20:42] context actually of your SNOP meeting rather  with those the delays and iteration and even if [20:49] the overall scope of this is relatively  ambitious for a normal organization there [20:53] are definitely pieces of this that are are quite  applicable and can be easily uh implemented. So um [21:02] you know with that said I think it's it's  more about how do you thoughtfully use the [21:06] capabilities that exist today to drive value  not so much about what do you have what do you [21:10] think and where it's going to be in a few years  but there are things that we can think about [21:14] how do we actually help organizations  today be able to move that needle. So [21:18] I said thank you all very very  much. Uh, I know we're at time.