HCLTech Launches AI Factory Stack to Cut Enterprise Compute Costs

HCLTech推出AI Factory技术栈以降低企业算力成本 HCLTech认为企业未能成功实现AI规模化是因为缺乏对基础设施的统一控制，并称其全新AI Factory技术栈及Token级FinOps可将GPU利用率提升最高达40%，同时降低总体拥有成本。 Rampal Singh详细阐述了企业AI规模化所需的五大基础支柱：经济性、主权性、责任性、可持续性与分布式架构。为帮助机构弥合基础设施与智能之间的鸿沟，HCLTech提供了AI Factory产品组合——这是一套涵盖从数据中心设计到专用小语言模型（SLM）及专有AI FinOps的工程化、以成果为导向的技术栈。借助AI Force和AI Foundry等统一平台，企业能够优化GPU利用率、加速投产进程，并将总体拥有成本降低最高达40%。 企业在规模化应用 AI 时的核心挑战在于如何兼具企业级安全、治理与切实的投资回报率（ROI）；实现这一目标需要基于由成本与明确回报、主权、责任伦理、可持续性以及分布式架构构成的五大基石，并由董事会及设立的 AI 卓越中心（CoE）自上而下驱动转型。 (0:04) HCLTech 推出了端到端工程化架构 AI Factory，覆盖从被动基础设施、液冷改造到模型即服务（MaaS）和垂直行业场景的完整服务层，旨在消除企业从 POC 概念验证到生产落地的脱节问题。 (7:13) 来自全球 2000 强企业的落地数据显示，AI Factory 可将 GPU 利用率提升 35% 至 40%，POC 转化投产速度加快 40% 至 60%，并通过 FinOps 与集中式管理降低 30% 至 40% 的总体拥有成本（TCO）。 (10:20) 为解决大模型 token 消耗及 AI 算力成本不透明的痛点，HCLTech 打造了专用 AI FinOps 平台，结合细粒度用量监控与从业者主导的智能模型路由优化，承诺将企业 AI 支出降低 30% 至 40%。 (13:34) 依托 AI Force、AI Foundry 及 AI Engineering 等生态模块，HCLTech 还推出了可在企业防火墙内完全隔离运行的 AI in a Box 主权解决方案，支持以小步快跑的模式扩展物理 AI、生成式 AI 及 Agent 智能体工作流。 (14:30)

Rampal Singh of HCLTech argues that enterprises struggle to scale artificial intelligence due to underlying production and infrastructure constraints rather than model limitations. Achieving sustainable return on investment requires an integrated hybrid foundation that pairs sovereign, cost-governed infrastructure with disciplined operational oversight. Foundational Pillars for Enterprise-Scale AI (0:04) Singh contends that while the enterprise value of AI is broadly acknowledged, organizations rarely achieve scale because their foundational infrastructure remains fragmented. To address this, he identifies five core pillars: pre-defined affordability metrics, operational and jurisdictional sovereignty, adherence to regulatory and ethical standards, environmental sustainability through technologies like liquid cooling and green power, and a distributed architecture spanning public clouds, edge environments, and on-premise infrastructure. Because AI simultaneously influences front-office operations, back-office workflows, and workforce dynamics, Singh asserts that adoption cannot be treated as a standard IT project. Scaling instead demands direct C-suite and board sponsorship alongside an AI Center of Excellence (CoE) tasked with setting guardrails, facilitating reskilling, and transitioning ITIL processes toward outcome-based service models. Industrializing Production via the AI Factory (7:13) To counter the widespread failure of proofs of concept reaching deployment, HCLTech positions its AI Factory as an end-to-end engineered stack designed to solve production bottlenecks. The architecture encompasses passive data center design and high-density liquid cooling retrofits, physical operations managed via digital twins and edge robotics, and platform services offering Model-as-a-Service, inference pipelines, and specialized vertical use cases across major hardware OEMs and cloud platforms. HCLTech reports that implementing this standardized, vendor-agnostic infrastructure across Global 2000 client environments yields operational improvements: GPU utilization increases by 35% to 40%, the transition from proof of concept to production accelerates by 40% to 60%, platform engineering effort drops by 30% to 45%, and the overall total cost of ownership declines by 30% to 40%. Controlling Expenses with Specialized AI FinOps (13:34) Standard cloud financial management tools fail to account for the unique consumption metrics of AI workloads, such as variable token utilization across disparate business units. HCLTech argues that this lack of visibility obscures cost justification and limits measurable returns. In response, the provider implements a two-pronged AI FinOps model that couples an executive-to-user token tracking dashboard with active practitioner optimization, including dynamic model routing, capacity rightsizing, and specialized procurement. HCLTech claims this system delivers complete consumption transparency and targets a 30% to 40% reduction in overall enterprise AI expenditure. Modular Platforms and Sovereign Packaging (13:34) Beyond physical and platform infrastructure, HCLTech's ecosystem incorporates targeted operational suites. These include AI Force, which reports more than 40% productivity gains across the software development lifecycle and IT operations, and AI Foundry, which manages data lineage, modernization, and agentic workflows. To accommodate stringent regulatory and data residency mandates, HCLTech packages these capabilities into 'AI in a Box.' This turn-key, fully sovereign deployment runs entirely behind an enterprise's firewall, integrating physical, generative, and agentic AI use cases to allow organizations to pilot specialized implementations at a controlled scale before expanding. HCLTech presents an infrastructure-first framework to convert experimental AI into enterprise-grade operations; however, realizing its reported efficiency gains relies heavily on successful C-suite alignment, complex physical facility retrofits, and the effectiveness of specialized FinOps controls in managing real-world consumption.

HCLTech 指出，当前全球企业面临的核心困境已非论证人工智能的技术潜能，而是如何在确保合规、安全与投资回报（ROI）的前提下跨越概念验证并实现规模化投产。演讲者主张，企业必须依托贯通基础设施、平台与垂直应用的全栈混合架构，并结合专属的 AI FinOps 成本管控机制来化解生产落地障碍。 规模化落地的核心挑战与五大支柱 (0:04) Rampal Singh 提出，人工智能的效能完全受制于其底层架构的支撑能力。尽管各行业在数据中心、云基础设施以及原型概念验证（POC）上投入了巨额资本，但目前仅有极少数企业成功实现了规模化落地。演讲者认为，行业探讨的焦点已从 AI 是否具备商业价值，转移到能否在具备企业级安全、主权合规、流程治理并保证切实现金回报的条件下实现规模化交付。为此，Singh 归纳了支撑智能化规模落地的五大基础支柱：必须预设财务回报目标的成本可承受性、涵盖数据与技术栈的管辖主权、严格遵循合规伦理、采用液冷及绿色能源等手段的可持续性，以及横跨公有云、边缘节点与本地数据中心的分布式混合架构。 此外，演讲者认为企业客户关注的是基于业务成果的服务模式，而非底层微观硬件型号的短期炒作。这需要企业重构传统的 ITIL 流程，并在私有边界内部署针对特定场景的小型语言模型（SLM）以配合前沿大模型运作。在组织架构层面，Singh 强调 AI 项目必须由董事会与最高管理层直接主导，并通过设立 AI 卓越中心（AI CoE）来推进组织变革与技能重塑，从而应对 AI 对企业前端界面、后台系统、员工日常流程及核心业务模式带来的全方位冲击。 AI Factory 全栈架构与生产效能验证 (7:13) 针对“绝大多数企业并非面临模型算法难题，而是面临生产工程难题”的行业现状，HCLTech 推出了名为 AI Factory 的端到端工程化产品组合。该架构分为四层服务体系：最底层涵盖传统数据中心被动基础设施改造与高密度机架液冷技术升级；往上提供依托数字化孪生与边缘机器人自主巡检的数据中心运维支持；第三层提供裸金属算力调度、模型即服务（MaaS）、推理服务及模型目录；顶层则对接物理 AI、智能体（Agentic AI）与垂直行业认知流程自动化。该技术栈支持硬件中立整合，全面接入 NVIDIA、AMD、Intel、HPE 等硬件厂商，以及微软 Azure、谷歌云、AWS 和主权新型云（NCP）。 根据 HCLTech 援引的全球2000强（G2000）企业实际部署数据，该方案声称能为企业实现显著的效能优化：现有基础设施的 GPU 算力利用率提升 35% 至 40%；AI 用例从概念验证走向生产环境的速度加快 40% 至 60%；平台工程与日常运维工作量降低 30% 至 45%；借助集中式管理与 FinOps 实践使总体拥有成本（TCO）下降 30% 至 40%；同时将单次模拟计算的成本压缩 20% 至 30%。 细粒度 AI FinOps 与成本可视化治理 (13:34) 衡量 AI 投资回报率的主要障碍之一，在于常规云成本管理工具无法准确计量跨组织部门的模型调用及 Token 消耗等特定指标。为此，HCLTech 开发了专用的 AI FinOps 平台，并推行所谓的“双管齐下（2P）”治理策略：第一维度是打造一级可见性平台，提供覆盖从管理层宏观监控看板到终端用户级别 Token 消耗的端到端细粒度追踪；第二维度是由实践专家主导的主动成本优化，利用智能模型路由调度、算力规格右移适配（rightsizing）以及专项采购策略控制算力开支。 根据演讲者的承诺，该组合方案旨在为企业实现 100% 的资源消耗可见性，并帮助削减 30% 至 40% 的企业 AI 相关支出。这种颗粒度的成本治理机制旨在消除“黑盒”式调用带来的预算超支风险，使企业投资与具体产出建立透明的关联。 模块化产品生态与端到端交付矩阵 (13:34) 在 AI Factory 基础设施之上，HCLTech 进一步展示了其覆盖企业全生命周期的产品生态矩阵。核心服务平台 HCLTech AI Force 针对软件开发生命周期（SDLC）及 IT 运维提供模块化支持，据演讲称能带来超过 40% 的研发生产力与上市速度增益，并具备面向 SAP 与 Oracle 环境的专属模块；AI Foundry 专注数据生命周期管理（包括发现、血缘、现代化与变现）并提供智能体工作流构建环境；AI Engineering 侧重物理 AI 与先进机器人系统的集成研发；AI Advisory 则提供高管战略咨询与组织变革支持。 针对安全合规要求极高的企业环境，HCLTech 宣布推出完全运行于客户防火墙内的预集成主权方案“AI in a Box”。该方案预置了生成式 AI、物理 AI 与垂直领域成熟用例，支持企业以较低门槛起步并随业务需求弹性扩展。整个生态力求通过从底层硬件改造至顶层智能应用的闭环交付，依托基于业务成果的商业协议协助企业规避概念验证阶段的算力闲置与试错损耗。 HCLTech 认为跨越 AI 价值鸿沟的关键在于建立从机房物理设施到顶层智能体应用的统一混合工程栈与成本透明度；但本场演讲主要基于该服务商自身的技术框架与客户数据进行论述，对于异构基础设施环境下的跨厂商深度适配难度，以及传统复杂系统在进行小型语言模型与自动化改造时的长期运维成本，仍有待行业进一步的实践验证。

## Background 该视频由 Ai4 于 2026 年 8 月 18 日发布在 YouTube 平台上，内容为 HCLTech 混合云服务与 AI Factory 高级副总裁兼全球业务负责人 Rampal Singh 的主题演讲。根据简介，面对近半数大型 AI 项目面临失败或停留在实验阶段的现状，演讲深入阐述了如何通过工程化全栈蓝图打通从底层基础设施到上层智能的断层，帮助企业实现兼具安全性、主权性与可衡量 ROI 的规模化 AI 落地。 ## TL;DR 演讲指出企业当前面临的不是“AI 问题”而是“生产级投产难题”，大多数原型无法转化为商业回报的核心在于缺乏统一的智能混合基础底座。为此，HCLTech 提出了以经济可承受性、主权可控、负责任伦理、绿色可持续和分布式架构为五大支柱的 AI Factory 工业级解决方案。方案配套自研的 AI FinOps 平台与“双 P 策略”，实现 Token 粒度的成本穿透并将企业 TCO 降低 30% 至 40%。结合 AI Force、AI Foundry、主权一体化 AI in a Box 以及物理 AI 矩阵，该体系成功帮助大型企业大幅提升 GPU 利用率并加速投产周期，打通投资到收益的确定性闭环。 ## Atmosphere & style 整场演讲洋溢着务实、严谨且高度聚焦商业价值的工业级工程基调，彻底剥离了市场对大模型的盲目炒作与概念包装。Rampal Singh 以资深技术领导者的从容与洞察，直面企业 CXO 最关心的成本黑盒与投产死穴，条理清晰地展开全栈解法。配合扎实的实测运营数据与系统化架构图，现场氛围理性而笃定，带给观众一种从虚幻技术狂热回归坚实工程底座的踏实感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=4s) 企业 AI 的核心瓶颈已从“概念验证（PoC）”转向“生产投产”，AI 的能力上限取决于底层基础设施底座的工程化支持。 - [02:04](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=124s) 实现规模化智能必须依托五大核心支柱：经济可承受性（Affordable）、主权可控性（Sovereign）、负责任治理（Responsible）、绿色可持续（Sustainable）和分布式原生架构（Distributed）。 - [05:16](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=316s) 企业级 AI 必须通过设立 AI 卓越中心（AI COE）实施自上而下的统一治理，并深度推行组织变革管理与员工技能重塑。 - [07:40](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=460s) HCLTech AI Factory 以单一责任主体模式覆盖 AI 数据中心设计与构建、智能化运维以及认知平台与模型服务三大板块。 - [10:01](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=601s) 在 G2000 企业的实测中，AI Factory 助力 GPU 利用率提升 35%-40%，PoC 到生产速度加快 40%-60%，并通过 FinOps 降低 30%-40% 的 TCO。 - [13:34](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=814s) 传统云 FinOps 难以适应 AI 的高频 Token 消耗，自研专属 AI FinOps 平台结合“双 P 策略（平台+实践专家）”可降低 30%-40% 的 AI 综合支出。 - [15:22](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=922s) HCLTech 构建了包含 AI Force（流程重构与 SDLC）、AI Foundry（数据资产与 Agentic AI）及全周期咨询的全栈产品矩阵。 - [15:22](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=922s) AI in a Box 方案为企业提供部署在本地防火墙内的主权一体化方案，支持低门槛启动并弹性扩展。 - [17:23](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=1043s) 物理 AI（Physical AI）与工业垂直 IP 深度预装于工厂方案中，覆盖智能安防、机器人作业与资产运维等具体业务成果交付。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=4s) 基础决定上限：从原型验证走向规模化智能 企业级 AI 的演进正在经历深刻的分水岭。Rampal Singh 强调了一个核心论断：“AI 的强大程度取决于其背后的基础设施底座”（"AI is only as powerful as the foundation behind it."）。过去数年间，企业在数据中心、云计算应用和数据治理上投入了海量资本，近期又全力押注 AI，但绝大多数企业未能从中获得显著的商业回报。市场上充斥着大量的概念验证（PoC）和实验原型，真正能够跨越鸿沟、在全企业范围内实现“规模化智能”（Intelligence at Scale）的客户寥寥无几。 当下的市场诉求已经发生转变。企业不再怀疑 AI 是否能带来价值——大家普遍认可 AI 在重塑用户体验、提升数字信任、优化端到端效率以及拉动顶线（Top line）与底线（Bottom line）增长上的潜力。真正的核心痛点在于：企业能否在兼顾企业级安全性、合规性、严格治理与清晰 ROI 的前提下，大规模交付智能应用？要解开这一难题，必须构建一层“智能混合基础架构底座”（Intelligent Hybrid Foundations Layer），将底层的 AI 基础设施、数据平台、AI 模型及上层业务应用整合为统一的工程化体系。 ```mermaid flowchart LR accTitle: Intelligent Hybrid Foundation Model subgraph S1["基础层"] A["物理基础设施 / 数据中心"] B["数据治理与算力网络"] end subgraph S2["平台与模型层"] C["AI 平台与模型目录"] D["安全/合规/主权治理"] end subgraph S3["应用与价值层"] E["业务应用与智能体"] F["可衡量的 ROI 产出"] end A --> C B --> C C --> E D --> E E --> F ``` --- ### [02:04](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=124s) 规模化智能的五大基石支柱 要在企业环境中落地规模化智能，必须依靠五大核心支柱来支撑整个运营架构： 1. **经济可承受性（Affordable）**：AI 投资必须具备清晰可量化的 ROI 定义。企业业务结构高度复杂，每一笔大额技术支出都必须有可预测的投资回报作为依据。 2. **主权可控性（Sovereign）**：针对数据驻留、AI 模型栈、底层运维以及司法管辖权的合规顾虑，AI 堆栈必须在数据层与操作层建立严格的安全护栏与主权边界。 3. **负责任与伦理治理（Responsible）**：所有 AI 业务场景必须符合道德伦理标准，并严格满足行业监管与合规要求。 4. **绿色与可持续性（Sustainable）**：GPU 集群和高密度 AI 数据中心正在消耗惊人的电力和水资源。必须引入 FinOps 实践、绿色可再生能源及先进液冷技术来确保环境可持续性。 5. **分布式原生架构（Distributed）**：AI 部署天生具有分布式特征，既运行在公有云上，也广泛分布于本地私有数据中心、新一代云服务商（Neocloud Providers / NCPs）以及制造业现场和工控边缘端。 企业客户最终需要的是基于业务成果交付的模式（Outcome-based model），而非深陷于 GPU 选型、网络参数等底层技术细节中。与此同时，通用前沿大模型虽在持续领跑，但企业内部正在加速涌现针对特定垂直场景定制的小语言模型（SLMs），在企业私有边界内高效运转。 --- ### [05:16](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=316s) 顶层治理与变革管理：建立 AI 卓越中心（AI COE） AI 的影响范围与以往的技术浪潮截然不同。云计算转型虽然深刻，但往往局限于 IT 和特定的业务部门，而 AI 具有全域渗透性（Pervasive），直接触及企业的前台、后台、员工生产力以及客户交互的每一个触点。面对这种无所不及的技术渗透，AI 的推进必须由董事会和高管团队自上而下发起并实施统一治理。 借鉴云计算时代的演进路径，企业必须设立专门的“AI 卓越中心”（AI COE）。AI COE 承担着制定技术标准、划定安全合规护栏、规划业务路线图以及监控 ROI 的职责。伴随技术落地而来的，是一场深度的组织变革管理（Change Management）：不仅是对员工进行技能重塑（Reskilling），更关键的是扭转组织心智。如果仅仅引入 AI 工具却依旧沿用传统的官僚流程与工作方式，AI 将根本无法产生预期的效能。 --- ### [07:40](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=460s) HCLTech AI Factory：三层端到端服务体系 针对上述工程与运营挑战，HCLTech 推出了全栈式的 AI Factory 解决方案，旨在以“单一责任主体模式”（Single Ownership Model）帮助企业跨越从底层硬件到上层认知的鸿沟。AI Factory 包含三大核心服务板块： * **AI 数据中心设计与构建（AI Data Center Design & Build）**：此前主要服务于超大规模云厂商（Hyperscalers）、新型云提供商（NCPs）和托管机房（Colo）。随着企业内部算力需求爆发，该服务帮助企业改造与翻新现有数据中心，引入高密度机架与液冷系统。 * **AI 数据中心支持与运维（AI Data Center Support & Management）**：AI 算力设施的运维逻辑与传统数据中心截然不同。该服务结合了机器人巡检与数字孪生技术，实现高密度集群的自动化与智能化运维。 * **认知基础设施与平台服务（Cognitive Infrastructure & Services）**：提供从裸金属管理、模型即服务（MaaS）、推理即服务、模型目录，一直到垂直行业用例开发的全流程服务。 该体系实现了位置与架构无关性，支持公有云、本地机房、托管中心以及边缘计算节点的全场景平滑适配。 --- ### [10:01](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=601s) 工业级工程实践与 G2000 企业实测成果 现场展示的视频明确指出：“大多数企业面临的并不是 AI 问题，而是生产投产难题”（"Most enterprises do not have an AI problem. They have a production problem."）。绝大部分行业方案并未针对生产级落地进行工程化打磨，而 HCLTech AI Factory 依托三十年关键基础设施运维经验，构建了从架空地板到生产级 AI 工作负载的完整支撑。 AI Factory 具备三大核心差异化特质： 1. **立足于工程严谨性而非抽象概念**：在部署前完成蓝图架构的成本与性能优化。 2. **广泛深厚的生态集成**：硬件层深度对接 NVIDIA、AMD、Intel、HPE 等；云底座整合 Microsoft、Google、AWS；同时支持主流大模型与平台生态。 3. **原生内置生产级运维与负责任 AI**：自带 GPU 智能编排、全栈可观测性、FinOps 机制与 7x24 小时托管运维。 对于已经采用 AI Factory 方案的全球 2000 强（G2000）大型企业，实际运行取得了显著的数据突破： - **现有基础设施 GPU 利用率提升**：35% - 40% - **从概念验证（PoC）到企业级投产速度加快**：40% - 60% - **平台工程与运维人力工作量降低**：30% - 45% - **通过集中化基础设施与 FinOps 使 TCO 降低**：30% - 40% - **单次模拟计算成本降低**：20% - 30% --- ### [13:34](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=814s) 突破黑盒：AI 专属 FinOps 平台与双 P 优化法 针对企业最为头疼的 ROI 难题，HCLTech 调研市场后发现，现有成熟的云 FinOps 平台无法无缝迁移到复杂的 AI 工作负载场景。AI 调用的 Token 消耗量激增，但企业管理层往往无法看清算力具体被谁消耗、流向何处。为此，HCLTech 自研了专属的 AI FinOps 平台。 该平台提供端到端的成本透视能力，能够生成从 CXO 宏观视图到个人用户级的极细粒度 Token 消耗与成本分析。在落地实践中，HCLTech 采取“双 P 策略”（2P Approach）： - **优秀的产品平台（Platform）**：内置智能提示词优化、动态模型路由（Model Routing）与实时计量。 - **实践者落地方法（Practitioner）**：由工程专家介入业务流进行深度调优。 通过这套组合拳，企业可获得 100% 的 AI 算力使用透明度，并承诺将 AI 综合支出直接降低 30% 至 40%。 --- ### [15:22](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=922s) 全栈产品阵列：AI Force、AI Foundry 与 AI in a Box 围绕 AI Factory 底座，HCLTech 构建了一套完整的全栈赋能矩阵： ```mermaid flowchart TD accTitle: HCLTech Full Stack AI Portfolio subgraph APPS["上层平台与服务"] AF["AI Force (SDLC / IT 运维重构)"] AY["AI Foundry (数据血缘与 Agentic AI)"] AE["AI Engineering & Physical AI"] AD["AI Advisory 咨询服务"] end subgraph BOX["交付模式"] AIB["AI in a Box (主权全栈一体化)"] end subgraph BASE["底座支撑"] FC["AI Factory (算力/数据中心/混合云)"] end AF --> AIB AY --> AIB AE --> AIB AD --> AIB AIB --> FC ``` - **AI Force**：业务流程重构旗舰平台，内含针对软件开发生命周期（SDLC）的专用模块（使全周期研发效率与上市速度提升 40% 以上），并深度优化 IT 运维、应用运维以及 SAP 和 Oracle 等企业级企业系统的日常运作。 - **AI Foundry**：专注于企业数据资产的全生命周期管理（涵盖数据发现、血缘分析、现代化改造与商业化变现），并提供企业级安全合规环境下的智能体（Agentic AI）构建平台。 - **AI Engineering**：聚焦于物理 AI（Physical AI）与具身机器人等高精尖领域。 - **AI Advisory Services**：提供全生命周期咨询服务，护航企业 AI 转型。 - **AI in a Box（开箱即用 AI 方案）**：最新发布的主权 AI 一体机方案。该方案直接部署在企业本地防火墙内部，集成了全栈模型、开发环境与安全治理护栏，企业可低门槛启动（Start Small）并随业务扩展弹性扩容（Grow Big）。 --- ### [17:23](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=1043s) 物理 AI、垂直 IP 资产与价值落地总结 在硬件与工业智能结合领域，HCLTech 打造了丰富的物理 AI（Physical AI）应用矩阵，直接预装并打包于 AI Factory 方案中以业务成果模式交付。具体应用涵盖：智能安防监控、工业安全规范检测、现场机器人作业、认知流程自动化以及智能设备资产与运营管理。 此外，HCLTech 将深厚的行业知识沉淀为可复用、可规模化的垂直行业 IP 解决方案，并全面升级为智能化 AI 套件。演讲最后对全方位能力进行了四点核心归纳： - **全栈端到端服务目录**：打通被动基础设施、主动算力、基础模型到上层智能体与物理 AI 的完整链路。 - **低门槛弹性演进机制**：支持从小规模场景切入，随使用量平滑扩展，从第一天起保障明确的 ROI。 - **广泛的战略伙伴生态**：兼具全球顶级软硬件巨头与垂直赛道创新先锋的生态协同。 - **工业级自主 IP 与加速器**：提供大量经过实战检验的自研资产，加速企业规模化智能的落地进程。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=4s) AI is only as powerful as the foundation behind it. — Rampal Singh > [00:04](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=4s) 企业能否在兼顾企业级安全性、合规性、严格治理与清晰 ROI 的前提下，大规模交付智能应用？ — Rampal Singh > [02:04](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=124s) 企业客户最终需要的是基于业务成果交付的模式（Outcome-based model），而非深陷于 GPU 选型、网络参数等底层技术细节中。 — Rampal Singh > [05:16](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=316s) AI 具有全域渗透性（Pervasive），直接触及企业的前台、后台、员工生产力以及客户交互的每一个触点。 — Rampal Singh > [05:16](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=316s) 如果仅仅引入 AI 工具却依旧沿用传统的官僚流程与工作方式，AI 将根本无法产生预期的效能。 — Rampal Singh > [07:40](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=460s) 以“单一责任主体模式”（Single Ownership Model）帮助企业跨越从底层硬件到上层认知的鸿沟。 — Rampal Singh > [10:01](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=601s) Most enterprises do not have an AI problem. They have a production problem. — Rampal Singh > [13:34](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=814s) 现有成熟的云 FinOps 平台无法无缝迁移到复杂的 AI 工作负载场景。AI 调用的 Token 消耗量激增，但企业管理层往往无法看清算力具体被谁消耗、流向何处。 — Rampal Singh > [15:22](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=922s) 企业可低门槛启动（Start Small）并随业务扩展弹性扩容（Grow Big）。 — Rampal Singh > [17:23](https://www.youtube.com/watch?v=DgaxSyjD7n4&t=1043s) 打通被动基础设施、主动算力、基础模型到上层智能体与物理 AI 的完整链路。 — Rampal Singh ## Mentioned Rampal Singh, HCLTech, Ai4, AI Factory, AI Force, AI Foundry, AI in a Box, AI Engineering, AI Advisory Services, AI COE, AI FinOps, NVIDIA, AMD, Intel, HPE, Microsoft, Google, AWS, SAP, Oracle ## Video description Rampal Singh - SVP & Global Business Head, Hybrid Cloud Services and AI Factory, HCLTech Enterprise AI ambition is at an all-time high, yet less than half of AI experiments make it to production and 43% of major AI initiatives are expected to fail, according to HCLTech’s AI Impact Imperatives, 2026, an enterprise AI market report. In this session, Rampal Singh lays out an end-to-end blueprint that connects infrastructure, platforms and intelligence into a single engineered stack, so enterprises can scale AI that consistently delivers ROI, responsibly. Audience takeaways 1\. A leadership view of the operating model that turns AI experiments into measurable business outcomes 2\. Why the infrastructure to intelligence gap needs to be managed for achieving enterprise AI ROI 3\. How to architect a sovereign stack that still delivers enterprise-grade AI at scale 4\. How this blueprint translates into predictable cost, performance and business outcomes 5\. How AI in a Box and SLM-as-a-service transform the operating model 6\. Enterprise examples of how the Factory approach translates AI investments into measurable ROI [Source](https://www.youtube.com/watch?v=DgaxSyjD7n4)

[0:04] Rampal Singh: AI is only as powerful as the foundation behind it. [0:04] Rampal Singh: 企业能否在兼顾企业级安全性、合规性、严格治理与清晰 ROI 的前提下，大规模交付智能应用？ [2:04] Rampal Singh: 企业客户最终需要的是基于业务成果交付的模式（Outcome-based model），而非深陷于 GPU 选型、网络参数等底层技术细节中。 [5:16] Rampal Singh: AI 具有全域渗透性（Pervasive），直接触及企业的前台、后台、员工生产力以及客户交互的每一个触点。 [5:16] Rampal Singh: 如果仅仅引入 AI 工具却依旧沿用传统的官僚流程与工作方式，AI 将根本无法产生预期的效能。 [7:40] Rampal Singh: 以“单一责任主体模式”（Single Ownership Model）帮助企业跨越从底层硬件到上层认知的鸿沟。 [10:01] Rampal Singh: Most enterprises do not have an AI problem. They have a production problem. [13:34] Rampal Singh: 现有成熟的云 FinOps 平台无法无缝迁移到复杂的 AI 工作负载场景。AI 调用的 Token 消耗量激增，但企业管理层往往无法看清算力具体被谁消耗、流向何处。 [15:22] Rampal Singh: 企业可低门槛启动（Start Small）并随业务扩展弹性扩容（Grow Big）。 [17:23] Rampal Singh: 打通被动基础设施、主动算力、基础模型到上层智能体与物理 AI 的完整链路。

[0:04] Rampal Singh: AI is only as powerful as the foundation behind it. [7:13] Presenter: most enterprises do not have an AI problem; they have a production problem.

[0:04] Good afternoon everyone and uh  thank you for being over here today. [0:10] Let me start today with a simple  but a very powerful statement. [0:16] AI is only as powerful as the foundation  behind it. Over the years, customers have [0:24] invested significantly into data centers,  cloud applications, data and recently into AI [0:34] but they failed to miserable significant gain  out of the AI obviously like there have been [0:41] the prototypes PC's but there are only limited  set of the customers who were able to make [0:47] meaningful progress or able to implement AI at  scale. or I should say intelligent at scale. [0:55] Uh customers have understood like customers  are expecting significant gain out of the [1:01] AI. They're expecting like user experience could  be elevated. All the trust points, digital trust [1:07] points could be elevated, improvement in the  bottom line, top line, new avenues of revenue, [1:12] efficiencies, productivity. So in nutshell  question is no longer does AI deliver value. [1:20] I think everyone understands in the market  AI could deliver the value. Question is [1:26] can AI deliver value at scale or intelligent at  scale with enterprisegrade security, compliance, [1:34] governance, ROI etc. And we strongly believe that  is not possible without having an intelligent [1:43] hybrid foundations layer which connects  AI applications data infrastructure [1:50] into single unified model to deliver intelligence  at scale and that is our narrative about how do we [1:57] really implement intelligence at scale  and we're leading the market now with that [2:04] what we strongly believe is you need Five pillars  to implement intelligence at scale. Number one is [2:13] AI has to be affordable. So there has to be  clear definitions of ROI against the spend [2:20] or investment into AI. Enterprise are know very  complex and definitely they look at ROI before [2:27] making large investments and hence getting ROI  out of AI becomes very very compelling reasons [2:35] and hence AI has to be affordable. Number two,  AI has to be sovereign in nature. So there are [2:41] the concerns which has been which have been  raised by the customers in terms of AI being [2:46] sovereign which means like your data layer your  AI layer tax stack operations jurisdictions etc [2:54] have to be sovereign in nature without  having those security compliance guard rails [3:01] it's very difficult to scale AI or in implement  intelligence at scale number three and very [3:08] critical is AI has to be respons responsible. So  AI use cases has to be ethical. They need to meet [3:16] regulatory governance compliances requirements.  Number four, AI has to be sustainable. [3:23] All these investment into you know data centers,  no GPUs, they are guzzling power, water no off [3:29] there is against investments into AI and AI data  centers. Which means we really need to look at the [3:36] ways we make AI know sustainable through phops  no practices having like no re reusable power [3:43] green power liquid cooling all those advanced  concepts including heavy data centers on the Mars [3:49] right and number five by default AI would be not  distributed which means you would see significant [3:57] implementations of AI on the public cloud as  they have been happening over last no 18 months [4:02] but you would also see significant  implement is happening in the data centers [4:07] with NCPs, Neocloud providers, but  also significant AI would be sitting [4:12] on the edge locations in terms of manufacturing  sites, no factory sites etc. Right? And all that [4:19] have to be delivered in a outcome based model to  the customer. Customer shouldn't be looking into [4:24] nuts and bolts and nitty and integrity. Right?  Now there could be overenthusiasms in terms of [4:29] knowing about no GPUs model this one that one but  eventually that will fade away and customer would [4:34] be only worried about one thing which is can you  deliver AI as an outcome to the customers right to [4:42] make that happen you need to reimagine the  services you need to reimagine every notice point [4:48] all the ITIL processes or processes we really need  to relook at and see how do we really re-engineer [4:54] them into the era of AI a native enterprise and  last but not the least I think while there would [5:01] be frontier model keep leading the market but  there is an emergence of having purpose-built [5:07] specific use cases built SL models which would be  implemented within the boundaries of enterprise [5:16] enterprise are not inherently complex right  and very bureaucratic as well at the same time [5:23] to implement something like AI which little  you know touches anything and everything. I [5:28] think this is the first thing which is really  pervasive in nature testing there is nothing [5:32] which is being untouched by AI. Even when you look  at the cloud and you compare it with cloud cloud [5:37] was also impacting some parts of an enterprise but  AI is something which is really impacting it know [5:43] business side user sides employee side you  know front end back end. So there there is [5:49] literally no one untouched by AI which means  there has to be mandate which is not top down. [5:55] So AI implementation at scale has to be led by  board and the leadership of the company and they [6:02] need to govern it. Then only there is a way out  to implement AI at scale. Like in case of cloud we [6:08] had cloud COE. Cloud CE was responsible to set up  the standards no guard rails implementations bring [6:15] stakeholders at the same level. In the same way we  need something like AICOE center of excellence to [6:21] implement AI at scale which would also take care  of again the standards and the guard rails and the [6:27] road map and ROI etc. uh which would be you know  governed by the board and the you know leadership [6:34] there's a huge element of the change  management when it comes to AI adoptions [6:39] right without change management it's not  feasible to implement AI or intelligent at scale [6:45] which not only not just include the reskilling  of the people which is very critical but they [6:51] at the same time to make it successful how do we  really navigate the mindset shift and enterprise [6:57] we apply AI but we keep working with the old  ways of working AI won't yield the no desired [7:04] results and we always would have the questions  around AI as part of know SL techch and large [7:13] global companies IT services companies we  look at on these problem statements and [7:18] we built a very comprehensive set of portfolios  around these services our ambition is to be the [7:24] best AI solutions provider in the market  and in order To achieve that ambitions, [7:30] we have built a very comprehensive end-to-end  know set of no services which include [7:35] infrastructure but also the verticals and the  models and the use cases etc. As part of the [7:41] a factory the number one service which we have  built is a AI data center no design and build. [7:47] So so far we have been servicing hyperscalers and  ncps and coloss for AI data center and design and [7:54] build. But as we start implementing AI into the  data centers, it which means we also need to [8:00] retrofit the data centers a bit not really at the  scale of the NCPs and colo providers but we might [8:06] need to have certain racks in the data center  with liquid cooling capabilities and all that. [8:11] That's where I think this service becomes very  critical very instrumental for the enterprise [8:16] to retrofit their data center. Once you build  this data center, design this data center [8:21] then you also need to manage this  highly sophisticated data centers. [8:25] So managing a data center is inherently different  than managing no traditional data center. [8:30] That's where know service number two becomes very  critical. We have also implemented the use cases [8:36] which are robotics and know digital twin driven  as part of the DC portfolio manage services. [8:42] And once you have managed it then third  thing which is the cognitive infra in terms [8:47] of starting from bare metal services model as  a service inference as a service model catalog [8:53] and vertical use cases that's where service  number three comes into the picture. So if [8:57] you combine all three services fundamentally what  we are saying is we can really provide end to end [9:04] comprehensive set of the services we can become we  can provide single ownership model to the customer [9:10] to implement AI or intelligence at scale. If you  really look at this picture starting from AI data [9:17] centers or we can implement it across it could be  on the hyperscalers it could be in data centers [9:23] as locations with NCPs. So we are locations  agnostic or model agnostic in that sense. [9:29] Then we have the capabilities with respect to AI  data centers which is the passive infra and moving [9:36] further in terms of AI infra platform services  AI data platform model services and vertical [9:42] use cases on one side on the other side starting  from consulting advisory manage services for AI [9:48] security compliance. So we have very comprehensive  set of or portfolio of services for the customers [9:54] to implement intelligence at scale. And with that  I will give you some glimpse in terms of how a [10:01] factoino looks like. Most enterprises do not have  an AI problem. They have a production problem. [10:11] Getting AI to run affordably and responsibly  at scale with clear returns is where the real [10:17] work is. And most of the industry is [music] not  built for that. Meet HCL Techch AI factory, an [10:24] engineeringled end-to-end offering that covers the  full stack from data center design and build to [10:31] infrastructure operations to the cognitive layer  where AI actually runs. Sovereign, deployment [10:37] ready and vendor agnostic. Built on three decades  of [music] running missionritical infrastructure [10:43] for global enterprises. structured around three  services. AI data center [music] design and build [10:50] validated reference architectures for highdensity  AI workloads engineered [music] for power, [10:55] cooling and performance from the ground  up. Second, AI data center support [11:01] physical infrastructure operations, rack [music]  integration, sustainability and automation. [11:07] And third, AI infrastructure and services. AI  environments [music] across private, public, [11:13] edge, sovereign, and neo cloud, including  inference, model services, and fin. Together, [11:21] they cover everything between the raised floor  and a production AI workload. And we are taking [11:27] AI factory a step further. Most AI infrastructure  [music] is monitored by people. Ours is [11:33] increasingly run by AI itself. through edge AI and  robotics embedded within the factory environment, [11:39] autonomously inspecting, managing, and optimizing  the [music] very infrastructure that AI runs on. [11:46] Three things make HCL Tech AI factory different.  First, it is built on engineering rigor, [11:52] not abstraction. [music] Blueprint-ledd  architectures that are optimized for cost [11:57] and performance before [music] deployment, not  after. Second, the ecosystem breadth is unmatched [12:03] from Nvidia, AMD, Intel, [music] HPE, and  many others on hardware to Microsoft, Google, [12:10] AWS, and others on cloud to AI model and platform  [music] providers across the stack. And third, [12:18] production-grade operations and responsible AI are  built in [music] GPU orchestration, observability, [12:25] PIOPS, and 247 managed services [music]  designed for secured and sovereign AI, not [12:31] adapted from something else. For G2000 enterprises  already running HCL [music] Tech's AI factory, [12:38] the numbers are in. GPU utilization [music] up 35  to 40% more from infrastructure already in place. [12:46] AI use cases [music] in production 40 to 60%  faster from proof of concept to enterprise ready [12:54] platform engineering and operations effort [music]  down 30 to 45%. 30 to 40% lower TCO through PIOPS [13:01] and [music] centralized infrastructure 20 to  30% reduced cost per simulation enterprisegrade [13:08] security and sovereignty compliance built in  [music] as standard. The next phase of enterprise [13:14] AI belongs to those who can run it. HCLTech AI  factory. AI built for scale. Engineered to run. [13:22] [music] [13:28] [music] [13:34] So one of the key questions around AI is ROI. And  we have scanned the market and we found out while [13:42] there are very matured phenops platform in the  market for the cloud right now there's no single [13:49] platform who is equally good when it comes to AI  and hence we built our own platform from scratch [13:57] when it comes to AI in our workloads ROI is know  very critical in case of AI and I think Phobs is [14:04] going to play a very critical role for the same  thing Right now there is a challenge in terms of [14:09] getting visibility from the AI usage. We while  there is no AI usage exploding in the enterprise [14:18] there is no visibility in terms of who is doing  what in terms of token usage and the consumptions [14:24] that's where know AI platform becomes very  instrumental. It can provide very different kind [14:29] of visibility which could be for CXO level,  view level even to the like individual no [14:35] user level to that no granularity we can provide  the visibility. In case of phenops we have the two [14:41] key approach one is having the absolutely first  class no platform but at the same times once [14:47] you have the visibility how do you really deliver  on that no visibility how do you really go about [14:53] optimizing on AI spend that's where know practic  comes into the practitioner approach comes into [14:59] the picture in terms of uh be it know propering  having the right model model routing capabilities [15:06] all that has been baked in as part of the platform  and you to practicing led approach. Uh with this [15:12] 2P approach, we can provide 100% visibility and  we also commit to optimize AI spend by 30 to 40%. [15:22] This is our full stack AI offerings while AI  factory remains one of critical element but [15:27] on top of AI factory we have AI force which  is our flagship no platform for service [15:33] transformation or service reimaginations.  It has multiple modules. One module is SDLC [15:39] where we can deliver more than 40% uh uh  productivity and no speed to market for the full [15:47] SDLC no life cycle. In the same way when it comes  to IT operations and applications operations, [15:53] we can deliver significant efficiency and  productivity over there. Same thing goes for the [15:57] SAP and the Oracle. AI foundry is all about how  do we really take care of the data starting from [16:03] discovery lineage modernizations monetizations the  full life cycle of the data and also the providing [16:10] the very mature aendic platform to build agents  with enterpriseg grade security and governance [16:15] AI engineering again is a very mature portfolio  for us we have made huge investment in terms of [16:20] physical AI and robotics no use cases uh in  that AI factory underpins everything and last [16:27] but not the least a advisory services to handle  customers to take them through the AI journey. [16:35] This is our flagship offering in terms of AI in  a box. This we have newly launched. This provide [16:41] sovereign AI solutions which could be  implemented in the within the boundaries [16:45] of the enterprise within the firewall  of the enterprise. Full stack model [16:49] including the use cases. This also provide  the operating environment to build agents [16:54] within the security and governance. This has all  these use cases with respect to the physical AI, [16:59] Zen AI, agentic AI, robotics etc. And  this also has the vertical use cases. [17:05] We can start with small and grow with the usage  as no AI grows AI usage grows in the enterprise. [17:10] So this becomes a very good approach in terms of  delivering not just the ROI but also implementing [17:16] AI with enterpriseg grade no security  compliance guard rails and responsible AI. [17:23] In the same way we have built a number of the  physical AI use cases which have been bundled [17:27] along with the AI factory which could  be delivered in an outcome based models. [17:32] Uh those use cases include like no surveillance  know safety related use cases, robotics use cases, [17:38] cognitive processes use cases,  smart assets, no smart operations. [17:42] There are myriad number of these use cases uh in  terms of the physical AI know we have built in [17:47] and bundled with AI factory which could  be delivered in outcome based model. [17:52] In the same way we have built number of the use  cases which are rep uh scalable and repeatable in [17:57] nature which are IPled uh vertical solutions those  no cut across no multiple verticals any vertical [18:03] you could think of and we have been we have built  number of the industries those solutions and then [18:09] we have transformed those solutions into the AI  solutions again these solutions have been bundled [18:14] along with a factory and could be delivered in  outcome based models right so in summary I think [18:20] number one we have extensive active service  catalog end to end service catalog where we can [18:25] fulfill the need of enterprise for end to end  no AI use cases which includes passive infra [18:30] active infra you know platform use cases model  use cases and all the vertical use cases could be [18:36] physical AI agent AI prescriptive AI robotics etc  number two we have an approach where we can start [18:44] small and know grow big as know a usage grows  which means we take care of ROI as part of the [18:50] model from day Number three, we have the huge set  of the partnership, large partnership as well as [18:55] very niche partnership in this uh space which  we leverage to deliver value to the customer. [19:01] And number four, we have built  number of IPS and excellators [19:05] to jumpst start the AI journey for the customers  and to implement AI or intelligence at scale for [19:12] the customers. With that know, thank you very  much for attending it. uh you can you know come [19:17] to booth number 837 to have further conversation  on this. Thank you for being over here today.