HPE, Nvidia Deploy On-Premises Sandboxes to Curb Rogue AI Agent Risks

HPE与NVIDIA部署本地沙箱以遏制失控AI智能体风险 面对82%的企业存在未授权自主智能体的情况，在主权基础设施上运行零信任执行沙箱可在削减50%运营成本的同时，防止未经批准的数据外泄。 随着自主AI智能体开始执行代码和调用工具，企业面临着Token成本失控、未授权智能体泛滥以及数据安全违规等严峻挑战。NVIDIA与HPE联合推出了一套零信任架构，利用OpenShell与NeMo Claw沙箱对所有外部技能及企业工具实施严格的白名单审批策略。作为该运行时安全层的后盾，HPE Private Cloud AI与AI Factory Composer提供了主权本地基础设施、精细化Token经济学跟踪，以及基于OpsRamp的全栈可观测性。本次会议最后进行了务实的问答环节，探讨了通过硬件合理选型应对GPU供应交付周期问题，并介绍了灵活的开发者工具集成以及HPE创下的18项MLPerf世界基准测试纪录。 向自主智能体AI的转变为企业带来了严重的安全风险，82%的组织内部存在未授权智能体，65%的组织报告了数据泄露和业务中断等安全事件。 (0:04) 为交付真正的业务价值，智能体需要受约束的零信任执行环境，在允许工具调用与代码生成的同时，避免出现失控的反馈循环或Token成本暴增风险。 (0:04) NVIDIA OpenShell和NeMo Claw将Hermes与Open Claw等执行套件嵌入零信任企业沙箱中，在默认情况下拒绝所有外部访问。 (4:21) NVIDIA通过提供经过验证的Markdown技能来解决智能体指令文件中的供应链漏洞，从而防范来自未经审查的公共技能市场的恶意利用。 (4:21) HPE AI Factory Composer提供了图形化治理与可观测性层，可用于发现影子智能体、跟踪实时工具使用情况，并评估每个智能体的Token使用效率。 (8:20)

As enterprise adoption shifts from generative AI to autonomous agentic systems, organizations face heightened risks from rogue agent behaviors, data breaches, and uncontrolled token expenses. To address these operational vulnerabilities, Hewlett Packard Enterprise and NVIDIA advocate combining zero-trust execution sandboxes with observable, on-premises private cloud infrastructure. Operational Disruptions and Rogue Agent Security Risks (0:04) The transition to agentic AI introduces severe operational and security challenges, notably runaway token costs and unauthorized agent activity. Research cited in the presentation indicates that 82% of organizations already harbor unknown or rogue agents within their networks, while 65% have experienced security incidents stemming from agent deployments. Reported consequences include sensitive enterprise data exposure (61%), workflow and operational outages (43%), and direct financial losses (35%). Because agents require autonomy to call tools, execute commands, and run generated code to deliver business value, isolated execution environments cannot simply lock them away from all systems. Instead, the continuous feedback loops between execution harnesses, external tools, and large language models require bounded zero-trust architectures and end-to-end visibility to prevent automated enterprise breaches. Zero-Trust Sandboxing with OpenShell and Verified Skills (4:21) To mitigate runtime vulnerabilities, NVIDIA is porting its OpenShell and NeMo Claw environments from developer desktop setups to enterprise server infrastructure. Rather than replacing existing execution harnesses, the platform embeds existing frameworks like Hermes and Open Claw into isolated sandboxes. Agents run with zero external privileges by default, requiring administrative whitelisting and explicit approvals for any network access, tool calls, or data retrieval. This framework also addresses software supply chain vulnerabilities tied to agent skills. Because autonomous agents utilize markdown instruction files to direct their actions, downloading unvetted skills from public repositories introduces malicious access vectors. NVIDIA addresses this risk by offering verified, pre-screened skills designed to execute safely within OpenShell sandboxes. Hardware Lead Times and Flexible Developer Workflows (16:32) Amid high market demand and supply chain pressures for cutting-edge NVIDIA platforms like Blackwell B200 and B300 chips, enterprise hardware procurement often involves significant backlogs. HPE addresses these lead times by structuring hardware offerings across distinct tiers. While multi-rack sovereign clusters experience long wait times, organizations can expedite deployments by right-sizing models to fit professional workstation cards, such as ProLiant servers equipped with RTX Pro GPUs. To prevent infrastructure from restricting engineering productivity, HPE Private Cloud AI integrates a Data Lakehouse and global namespace that preserve standard developer workflows. Additionally, an integrated AI Gateway gives administrators dynamic routing control, allowing organizations to process confidential tasks locally on sovereign models while forwarding standard queries to commercial cloud APIs such as Google Gemini. Enterprise Observability and Infrastructure Performance Differentiation (8:20) To manage deployments scaling from ProLiant servers to sovereign Cray supercomputers, HPE provides the AI Factory Composer. This platform replaces command-line interfaces with a graphical management console that uncovers shadow agents, tracks tool execution flows, and monitors per-agent token consumption rates. HPE claims that hosting agent workloads locally on Private Cloud AI provides up to a 50% cost savings compared to public cloud hosting while guaranteeing data sovereignty. Contrasting its architecture with competitors such as Dell, HPE points to its dedicated enterprise datacenter focus, supercomputing background, and deep integration with OpsRamp for full-stack telemetry across the agent lifecycle. HPE also highlights 18 world records on standardized MLPerf benchmarks achieved on identical NVIDIA silicon, attributing the gains to proprietary thermal engineering, interconnect design, and hardware tuning. Securing agentic enterprise systems requires balancing zero-trust sandboxing and operational observability against the real-world constraints of GPU hardware availability and developer workflow flexibility. While on-premises private clouds mitigate token inflation and regulatory exposure, enterprises must carefully right-size their model compute requirements to navigate industry-wide supply delays.

Ai4发布的技术研讨内容指出，随着企业从传统生成式AI转向具备自主行动能力的智能体（Agentic AI），失控的Token成本与安全漏洞已构成严峻挑战；HPE与NVIDIA据此提出了结合沙箱隔离、全栈可观测性与本地AI工厂的软硬件联合治理方案。 智能体落地的安全隐患与失控风险 (0:04) 从检索增强生成（RAG）迈向自主智能体的演进，使系统获得了调用外部工具、下发系统指令和运行生成代码的能力，但这也打破了传统的被动交互模式。数据显示，企业内网中未经授权的“流氓智能体”现象已相当普遍：82%的受访组织网络中存在未知智能体，65%的组织已报告过与智能体部署相关的安全事件。在已发生的事件中，敏感数据违规外泄占61%，业务流程中断占43%，直接由智能体失误引发的经济损失占35%。 NVIDIA沙箱技术与零信任执行边界 (4:21) 针对智能体执行环境的漏洞，NVIDIA将OpenShell与NeMo Claw技术引入企业级服务器，直接与Hermes、Open Claw等现有执行底座（Harness）集成。在OpenShell安全沙箱内，智能体默认遵循零信任原则，不具备任何外部访问权限；所有网络调用、工具触发以及技能（Skill）集成动作，必须经过IT管理员显式配置白名单并逐项审批后方可执行。此外，为防范公开市场下载未经审核的Markdown任务指令所带来的供应链投毒风险，方案主张采用端到端验证的技能集以保障运行安全。 HPE全栈基础设施与可观测性监控 (8:20) 在硬件与运行时管理层面，HPE提供从配备RTX Pro GPU的ProLiant服务器到隔离运行的Cray超算集群等全套基础设施，并通过管理平台中的AI Factory Composer实现智能体治理。该系统提供图形化界面，支持全企业范围的活跃智能体发现，以彻底清除影子智能体，同时实时追踪工具调用、技能执行与数据流向，并按智能体维度监控Token消耗速率。HPE方面宣称，采用本地HPE Private Cloud AI部署智能体工作负载，相较公共云端可节省高达50%的成本，并能确保核心主权数据不流出企业边界。 算力交付周期与开发者工作流兼容 (16:32) 针对业界普遍关注的GPU供应紧张与超算平台漫长交付周期问题，研讨指出超大规模云厂商对Blackwell及下一代Rubin架构芯片的争夺加剧了排队压力。HPE提出分级选型策略：对不需要顶级B200或B300计算密度的任务，可采用交付周期显著缩短且成本较低的RTX Pro GPU服务器。在开发体验与混合架构方面，HPE强调Private Cloud AI不会强绑定专有工具链，系统内建的数据湖仓（Data Lakehouse）、全局命名空间及AI网关，允许工程师在保留原生工作流的前提下，由管理员策略动态控制本地主权推理与外部商用API（如Google Gemini）之间的请求分流。 软硬件调优表现与企业级定位对比 (19:34) 在回应与Dell等同类企业级方案的差异时，主讲团队强调了系统可观测性与工程优化积累。HPE方案深度整合了OpsRamp，能够覆盖智能体全生命周期提供全栈遥测与监控；而在底层工程调优方面，演讲者表示HPE在运行相同NVIDIA芯片的标准MLPerf基准测试中保有18项世界纪录，体现了其在互连与散热工程上的积累。同时，区别于兼顾消费级PC业务的竞争对手，HPE强调自身源自超算血统、专精于数据中心服务器架构的企业级业务重心。 本场研讨展示了HPE与NVIDIA在本地智能体安全与基础设施治理上的整合路径，但企业在实际落地时，仍需在顶级GPU交付周期、各级工作负载的显存匹配度以及混合云路由策略之间进行权衡。

## Background 该内容由 Ai4 频道于 2026 年 8 月 19 日发布在 YouTube 平台上。视频记录了 NVIDIA 产品营销专家 Steve Stein 与 HPE 解决方案架构师 Mark Seither 的联合技术演讲，主要介绍了结合 NVIDIA AI Agent 工具包的沙箱策略与 HPE Private Cloud AI 及可观测性体系，构建安全、可信且可审计的企业级智能体运营（Agentic Ops）架构。 ## TL;DR 演讲首先指出企业在拥抱智能体（Agentic AI）时代时面临失控的代币成本与“野生智能体”带来的安全合规危机。为此，NVIDIA 推出了基于零信任架构的 OpenShell 与 NeMo Claw 沙箱隔离框架，严格限制外部调用并实施技能安全验证。HPE 则提供了从 RTX Pro 服务器到主权超算级的全栈 AI 工厂基座，并结合 HPE AI Factory Composer 实现全局可观测性与权限审计，相比公有云可降低高达 50% 的代币成本。在最后的 Q&A 环节中，专家们深入解答了 GPU 供应链交付应对策略、混合云 AI 网关路由自由度，以及 HPE 相比竞争对手在超算工程积淀与可观测性上的独特优势。 ## Atmosphere & style 整场演讲风格严谨务实、直击行业痛点，主讲人以 Black Hat 期间的安全警示数据切入，营造出一种技术变革与安全危机并存的紧迫感。随后的技术架构阐述层层递进、逻辑清晰，而在面对观众关于硬件漫长交付期与友商竞争等尖锐提问时，嘉宾表现出极高的坦诚度与工程自信，使得整场分享充满了前沿技术落地与企业级实操相交织的专业氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=vjy3spGIJB8&t=4s) 智能体（Agentic AI）正推动企业迈入实用 AI 时代，但同时也带来了代币经济学（Token 成本失控）与企业级安全风险两大核心痛点。 - [00:04](https://www.youtube.com/watch?v=vjy3spGIJB8&t=4s) 安全调研显示，82% 的企业存在未受纳管的野生智能体，65% 曾遭遇相关安全事故，造成严重的数据泄露、业务中断及直接财务损失。 - [02:33](https://www.youtube.com/watch?v=vjy3spGIJB8&t=153s) NVIDIA 推出基于零信任原则的 OpenShell 与 NeMo Claw 沙箱隔离工具，确保智能体在具备自治执行能力的同时，处于默认拒绝与显式白名单管辖之下。 - [02:33](https://www.youtube.com/watch?v=vjy3spGIJB8&t=153s) 针对智能体依赖的外部技能包，NVIDIA 引入端到端安全验证机制，杜绝开发者无意间引入携带恶意代码的指令或脚本。 - [08:20](https://www.youtube.com/watch?v=vjy3spGIJB8&t=500s) HPE 与 NVIDIA 合作提供全栈企业级 AI 基础设施，涵盖 HPE ProLiant RTX Pro 系列、开箱即用的 HPE Private Cloud AI (PC AI) 以及物理隔离的主权 AI 方案。 - [10:46](https://www.youtube.com/watch?v=vjy3spGIJB8&t=646s) HPE AI Factory Composer 补齐了全局可观测性层，能够全面追踪智能体调用链、权限边界、活跃度及实时 Token 消耗，并大幅降低运维使用门槛。 - [10:46](https://www.youtube.com/watch?v=vjy3spGIJB8&t=646s) 相较于按 Token 计费的公有云 API，将模型迁移至本地私有化部署可为企业削减高达 50% 的长期成本，并彻底解决核心数据外泄风险。 - [16:32](https://www.youtube.com/watch?v=vjy3spGIJB8&t=992s) 面对全球高端 GPU（如 B200/B300）漫长的交付周期，企业可采取差异化算力选型，选用供货更快、成本更低的 RTX Pro 专业卡作为高性价比替代方案。 - [18:10](https://www.youtube.com/watch?v=vjy3spGIJB8&t=1090s) HPE Private Cloud AI 内置 AI Gateway 并整合数据湖仓与全局命名空间，支持企业灵活配置混合云路由，在不破坏开发者原有开发习惯的同时严格把控敏感数据边界。 - [19:34](https://www.youtube.com/watch?v=vjy3spGIJB8&t=1174s) 相较于 Dell 等竞品，HPE 依托其数十年来在顶级超级计算领域的工程调优经验与 OpsRamp 全栈可观测性软件能力，在 MLPerf 基准测试中保持着 18 项世界纪录。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=vjy3spGIJB8&t=4s) 智能体时代的到来与两大核心痛点：失控的代币成本与安全隐患 来自 NVIDIA 的产品营销专家 Steve Stein 登台开场。他承接了当天上午的主题演讲，强调当今世界已正式跨入“Agentic AI”（智能体 AI）与实用 AI 的时代。企业必须学会接纳这一趋势，并着手构建 AI 工厂来迎接全新的技术挑战。然而，在充分释放智能体潜力的同时，企业正面临两大极为棘手的核心难题：其一是代币成本的失控（被称为“代币经济学” Tokconomics），智能体在执行复杂推理与自主循环时会消耗天文数字级别的 Token，成本远超早期的生成式 AI 和 RAG 系统；其二则是亟待解决的企业级安全危机。 演讲现场不远处正举办着 Black Hat 全球黑客大会，安全圈最关注的焦点正是 AI 智能体越狱、滥用权限甚至攻击企业系统等问题。Stein 直截了当地指出，这绝非遥不可及的理论推演，而是当下正在发生的真实危机： - **82% 的受访企业**内部已经存在未被纳管、处于“野生”状态的未知智能体（Shadow Agents）。 - **65% 的企业**报告遭遇过与智能体相关的安全事故。 - 在发生安全事件的企业中，**61% 面临数据泄露与暴露风险**，**43% 遭遇了业务与运营中断**。 - **35% 的企业因智能体安全事故蒙受了直接的经济损失**。 智能体的初衷是为企业大幅提升生产力，绝不应该反过来损害业务与财务安全。企业必须在赋予智能体必要行动权限的同时，彻底杜绝它们对内部系统造成破坏。 ```mermaid flowchart TD accTitle: 智能体安全事件影响分布 subgraph Enterprise["企业内部现状"] A["企业网络环境"] --> B["82% 存在未知智能体运行"] B --> C["65% 遭遇安全事故"] end subgraph Impact["具体危害表现"] C --> D["61% 数据暴露与泄露"] C --> E["43% 运营与业务中断"] C --> F["35% 直接财务损失"] end ``` --- ### [02:33](https://www.youtube.com/watch?v=vjy3spGIJB8&t=153s) 智能体执行框架与沙箱隔离：NVIDIA NeMo Claw 与 OpenShell 要让智能体真正发挥价值，就不能将其单纯“关在盒子里不让干活”；智能体必须具备高度自治权、能够演进，并支持执行代码与调用外部工具。智能体框架（Harness）充当了生成式 AI 的“执行手臂”，负责把大语言模型生成的 Token 转化为实际的代码指令并执行，再将结果安全回传给 LLM 分析。为了确保这一执行循环绝对可控且安全，NVIDIA 推出了基于零信任原则构建的底层工具——**OpenShell** 与 **NeMo Claw**。 以往 NeMo Claw 和 OpenShell 主要运行在个人桌面设备或开发者本地环境（如 DGX Station 等工作站）中，如今 NVIDIA 正与 HPE 携手将其推向企业级服务器生态。NeMo Claw 能够直接与市面主流的智能体框架（如 Hermes、OpenClaw）原生集成，而不是另搞一套专属私有框架。在架构中，智能体实例被严格限定在隔离沙箱内运行，默认处于“零访问权限”的拒绝状态；任何外部工具调用、API 请求或技能加载，都必须由企业 IT 部门显式配置白名单并经过实时策略审批，一旦发生越权行为将即时向管理员告警。 - **默认零信任（Deny by Default）**：沙箱内部智能体默认无法访问任何外部网络、文件系统或接口。 - **显式白名单与告警拦截**：对所有外部访问请求逐一审计，确保 IT 部门拥有 100% 的控制权。 - **技能市场安全验证**：针对智能体依赖的外部技能包（通常为指导其完成特定任务的 Markdown 指令文件），NVIDIA 展开端到端的安全验证，防止开发者从外部社区下载并引入携带恶意代码的技能。 ```mermaid flowchart LR accTitle: NVIDIA NeMo Claw 与 OpenShell 安全沙箱架构 subgraph Sandbox["OpenShell 安全沙箱环境 (NeMo Claw)"] Harness["执行框架 (Hermes / OpenClaw)"] Agent["主智能体 (Primary Agent)"] Harness --> Agent end subgraph Security["策略控制与网关"] Whitelist{"白名单 & 权限审批策略"} Audit["IT 审计与异常告警"] end subgraph External["外部资源"] VerifiedSkills["NVIDIA 认证的安全技能库"] Tools["企业内部工具 / API"] end Agent -->|"请求外部资源"| Whitelist Whitelist -->|"未授权"| Audit Whitelist -->|"已授权"| External ``` --- ### [08:20](https://www.youtube.com/watch?v=vjy3spGIJB8&t=500s) HPE 算力基座与全栈 AI 工厂解决方案 在安全沙箱的基础上，企业还需要一套稳健的本地化基础设施来承载生产级工作负载。Stein 强调，HPE 在整个方案中扮演着“承担底层运维重担”的关键角色，让企业客户无需为复杂的软硬件集成而烦恼。HPE 拥有从轻量工作站到顶级超算中心的完整产品矩阵，在业界公认的 TOP500 超算榜单中常年位列前茅（如 Cray 超算技术），具备支撑大规模 AI 工厂的高性能交付经验。 双方联合打造的企业级 AI 基础设施涵盖了多种落地形态：针对泛用型 AI 任务，提供基于 NVIDIA Pro 系列 GPU（如 RTX Pro 系列）的 HPE ProLiant 服务器；针对追求快速投产的企业，提供预先集成且经过性能调优的开箱即用型（Turnkey）AI 工厂与 **HPE Private Cloud AI (PC AI)**，内置完整的全托管软件栈及 Kubernetes 编排层；针对政府与高合规行业，则支持物理隔离（Air-Gapped）的端到端主权 AI 部署方案。 - **HPE ProLiant RTX Pro 系列**：高能效比算力基座，适应多样化企业级业务负载。 - **HPE Private Cloud AI (PC AI)**：集成 Kubernetes 与全托管软件栈，实现快速开箱部署。 - **主权与物理隔离方案**：为高安全合规场景提供严密的本地物理边界与数据驻留保护。 - **Cray 超级计算技术支撑**：具备支撑大规模多智能体并发与极限推理的集群扩展能力。 --- ### [10:46](https://www.youtube.com/watch?v=vjy3spGIJB8&t=646s) 全局可观测性、50% 成本削减与安全企业落地 针对多智能体协同带来的不可预测性，HPE 提供了专属管理套件——**HPE AI Factory Composer**，为企业补齐了至关重要的**智能体全局可观测性层（Observability Layer）**。很多企业之所以面临安全事故，正是因为对内部“野生”智能体的行为一无所知。借助该层能力，管理员不仅能清楚洞察每个智能体正在访问哪些工具与技能，还能实时统计每个智能体的 Token 消耗速率、追踪交互链路，并评估不同智能体的实际产出效率与 ROI。此外，HPE 还提供了直观的可视化图形交互界面（GUI），改善了传统 NeMo Claw 纯命令行界面的使用门槛。 在经济效益方面，Stein 指出本地私有化部署能够从根本上解决公有云 API 带来的代币账单雪崩问题。不少企业仅用几个月就烧光了全年的云端 Token 预算；而在本地运行开源或专有大模型，相比公有云 API 可实现 **高达 50% 的成本节约**。不仅如此，本地化运行还能确保企业核心数据绝对不出域，在合规性、主权可控、数据安全与可观测性之间取得完美平衡。 - **HPE AI Factory Composer**：实现对智能体调用链、权限边界、活跃度及生产力的全程可视化审计。 - **50% 长期成本降低**：彻底告别公有云高昂的按 Token 计费模式，大幅压降企业级 TCO。 - **数据完全自主可控**：推理数据与业务流水线完全保留在本地私有环境中，杜绝外泄隐患。 --- ### [15:36](https://www.youtube.com/watch?v=vjy3spGIJB8&t=936s) 总结与现场问答互动：本地部署与硬件交付周期的现实考量 Stein 总结道，NVIDIA 与 HPE 联合打造的这套安全智能体架构，通过“底层安全沙箱 + 经过验证的安全技能 + HPE 专属可观测性层 + 本地私有算力基座”，为企业构建安全可靠的 Agentic 生产环境提供了标准范式。演讲结束后，Stein 邀请台下的 HPE 专家一同参与现场 Q&A 环节。 随后，现场一位参会者提出了首个极具现实意义的尖锐问题： > 当前行业报告中普遍反映 AI 硬件存在严重的供应短缺与漫长交付等待期。刚刚在其他分会场也听到类似讨论，很多人因此更倾向于选择公有云基础设施而非本地部署。企业在选择本地私有化架构时，究竟该如何应对漫长的硬件就绪周期？ ### [16:32](https://www.youtube.com/watch?v=vjy3spGIJB8&t=992s) GPU 供应链交付周期与异构算力选型策略 现场观众首先针对硬件交付周期提出了关键疑问：当前行业内往往需要等待近一年左右的时间才能拿到顶级算力，面对这种普遍的硬件供应链积压，NVIDIA 与 HPE 双方如何看待并解决这一痛点。NVIDIA 代表坦言，这确实直接反映了当前全球对 NVIDIA GPU 的巨大需求与供应链瓶颈，而且这种短缺不仅存在于私有云领域，即便是能够率先拿到下一代 Rubin 架构系统的超大规模云厂商（Hyperscalers），同样面临着交付压力。然而，HPE 在应对交付周期方面构建了分层分级的“AI Factory”（AI 工厂）交付策略，主要划分为开箱即用（Turnkey）、规模化（At Scale）以及主权级（Sovereign）等不同层级。 尽管超大规模与主权级基础设施因体量庞大确实需要相对较长的交付期，但 HPE 为开发者系统和 HPE Private Cloud AI（PCAI）做了专门优化，其设计初衷就是为了能以极快的速度完成交付部署，帮助客户快速切入 AI 开发阶段。此外，演讲者建议企业在面临算力短缺时，应结合实际业务负载重新评估 GPU 选型。当下积压最严重、等待时间最长的是诸如 Blackwell 架构的 B200、B300 等顶尖性能 GPU；但如果企业的模型体量相对较小，采用 RTX Pro 专业级 GPU 则是极具性价比的替代方案。RTX Pro 显卡的供货周期明显更短，硬件成本也大幅降低，只要工作负载能够装入其显存限制内，就能有效避开漫长的供应链等待期。 - **AI Factory 交付梯队**：分为开箱即用（Turnkey）、规模化级（At Scale）及主权级（Sovereign），后两者由于硬件规模庞大交付周期较长。 - **快速交付方案**：针对开发环境的 HPE 专用开发者系统与 HPE Private Cloud AI 设计有极短的交付链路，可加速业务落地。 - **算力选型权衡**： - **B200 / B300 GPU**：极致算力但面临全行业最长的排单等待。 - **RTX Pro GPU**：交付周期更短且采购成本显著降低，主要制约在于模型参数规模与显存容量。 --- ### [18:10](https://www.youtube.com/watch?v=vjy3spGIJB8&t=1090s) 本地部署的开发者自由度与混合云 AI 网关路由 第二位观众提出了一个关乎工程团队日常体验的核心问题：当企业将模型部署在本地私有服务器上时，研发工程师在使用上会受到哪些限制？工程师是否必须被死死绑定在 HPE 专属软件中，还是依然可以使用他们目前习惯的日常开发工具链？在保证安全可控的同时，安全边界究竟划在哪里？ ```mermaid flowchart TD accTitle: HPE 混合云与本地 AI 网关路由架构 subgraph Clients["开发者与应用端"] DEV["工程师团队<br/>(通用工具链 / 原生 API)"] end subgraph HPE_Platform["HPE Private Cloud AI 平台"] GW["内置 AI Gateway<br/>(统一策略治理 / 动态路由)"] DLH["Data Lakehouse<br/>全局命名空间"] end subgraph Targets["多端执行环境"] LOCAL["本地基础设施<br/>(敏感私有模型推理/训练)"] CLOUD["公有云端模型<br/>(弹性计算资源)"] GEMINI["第三方商业模型 API<br/>(Google Gemini 等)"] end DEV -->|"发送请求"| GW GW <-->|"数据管道交互"| DLH GW -->|"本地私有路由"| LOCAL GW -->|"公有云路由"| CLOUD GW -->|"商业 API 路由"| GEMINI ``` HPE 专家回应指出，这完全取决于企业自身如何配置该平台。HPE 的私有平台具有极高的可扩展性与灵活性，其底层原生集成了数据湖仓（Data Lakehouse）、全套数据集成机制以及全局命名空间（Global Namespace）。这意味着平台不仅不会将开发人员限制在单一封闭环境内，反而赋予了企业统一分流的自由度： 1. 开发者可以通过内置的 AI Gateway（AI 网关）实现混合路由策略。 2. 企业可以自主定义哪些非敏感或通用负载转发至公有云处理，哪些任务调用 Google Gemini 等第三方前沿模型，而将涉及核心资产的敏感负载严格保留在本地私有环境中执行。 3. 无论底层的集群规模有多大，内置网关都能提供高度一致的接入与安全审计规范，既不打乱工程师既有的开发习惯，又全面收敛了企业级安全风险。 --- ### [19:34](https://www.youtube.com/watch?v=vjy3spGIJB8&t=1174s) HPE 与 Dell AI Factory 的核心差异与超算优势 最后一位观众提到，在刚刚参观展厅时看到 Dell 的展台同样展示了可以下载模型的“Dell AI Factory”，因此非常好奇 HPE 的 AI Factory 方案与 Dell 相比到底有何本质区别。 NVIDIA 与 HPE 双方代表指出，两者的差异非常明显，主要集中在软件可观测性与底层企业级超算底蕴两个维度： 首先是软件与可观测性层面。HPE 方案中深度整合了端到端的可观测性层以及企业级 IT 运营管理平台 OpsRamp。这使得企业在管理复杂 AI 工作流和智能体运行时，具备深入到底层硬件与上层模型的统一全栈洞察力，这是竞争对手所不具备的软实力。 其次是硬件工程与超算性能。虽然 HPE 与 Dell 都能采购到 NVIDIA 完全相同的芯片与底层硬件，但 HPE 在软硬件调优和系统级架构设计上优势明显，目前在 MLPerf 官方基准测试排行榜上保持着多达 18 项世界纪录。演讲者特别强调了企业定位上的根本分工：“HPE 名字末尾的‘E’代表着企业级（Enterprise）”。Dell 的产品线广泛覆盖了桌面级终端和 PC 系统，而 HPE 则是数十年如一日专注于企业级服务器和全球最顶尖的超级计算机（Supercomputing）。凭借打造全球顶级超算系统的深厚工程积累，HPE 在构建高可靠、高密度、受行业信任的超大规模企业级基础设施方面具有不可替代的天然优势。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=vjy3spGIJB8&t=4s) 82% 的受访企业内部已经存在未被纳管、处于“野生”状态的未知智能体（Shadow Agents）。 — Steve Stein > [00:04](https://www.youtube.com/watch?v=vjy3spGIJB8&t=4s) 65% 的企业报告遭遇过与智能体相关的安全事故。 — Steve Stein > [00:04](https://www.youtube.com/watch?v=vjy3spGIJB8&t=4s) 智能体的初衷是为企业大幅提升生产力，绝不应该反过来损害业务与财务安全。企业必须在赋予智能体必要行动权限的同时，彻底杜绝它们对内部系统造成破坏。 — Steve Stein > [02:33](https://www.youtube.com/watch?v=vjy3spGIJB8&t=153s) 默认零信任（Deny by Default）：沙箱内部智能体默认无法访问任何外部网络、文件系统或接口。 — Steve Stein > [02:33](https://www.youtube.com/watch?v=vjy3spGIJB8&t=153s) 显式白名单与告警拦截：对所有外部访问请求逐一审计，确保 IT 部门拥有 100% 的控制权。 — Steve Stein > [10:46](https://www.youtube.com/watch?v=vjy3spGIJB8&t=646s) 在本地运行开源或专有大模型，相比公有云 API 可实现高达 50% 的成本节约。 — Steve Stein > [15:36](https://www.youtube.com/watch?v=vjy3spGIJB8&t=936s) 当前行业报告中普遍反映 AI 硬件存在严重的供应短缺与漫长交付等待期。刚刚在其他分会场也听到类似讨论，很多人因此更倾向于选择公有云基础设施而非本地部署。企业在选择本地私有化架构时，究竟该如何应对漫长的硬件就绪周期？ — 现场参会者 > [18:10](https://www.youtube.com/watch?v=vjy3spGIJB8&t=1090s) 当企业将模型部署在本地私有服务器上时，研发工程师在使用上会受到哪些限制？工程师是否必须被死死绑定在 HPE 专属软件中，还是依然可以使用他们目前习惯的日常开发工具链？在保证安全可控的同时，安全边界究竟划在哪里？ — 现场参会者 > [19:34](https://www.youtube.com/watch?v=vjy3spGIJB8&t=1174s) HPE 名字末尾的‘E’代表着企业级（Enterprise）。 — NVIDIA 与 HPE 代表 ## Mentioned Steve Stein, Mark Seither, NVIDIA, HPE, Dell, OpenShell, NeMo Claw, Hermes, OpenClaw, HPE Private Cloud AI, HPE ProLiant, HPE AI Factory Composer, OpsRamp, Cray, Blackwell B200, B300, Rubin, RTX Pro, Google Gemini, MLPerf, Black Hat ## Video description Teve Stein - Product Marketing, NVIDIA Mark Seither - Solution Architect - APS Solution Sales, HPE Enterprise agents are becoming operators: executing commands, calling tools, spawning sub-agents, and running across hybrid environments. In this technical session, HPE and NVIDIA will share an emerging architecture for the Secure Agentic Enterprise, combining NVIDIA AI Agent Toolkit for sandboxed, policy-enforced execution with HPE Private Cloud AI and observability to deploy, inspect, interrupt, and audit agents on sovereign infrastructure. We’ll cover the Agentic Ops control plane: agent sandboxes, policy gateways, human and agent identity, token controls, observability layers, and human stop-lines for privileged actions. Attendees will leave with a practical design pattern for agents that are useful enough to act, bounded enough to trust, and observable enough to operate in production [Source](https://www.youtube.com/watch?v=vjy3spGIJB8)

[0:04] Steve Stein: 82% 的受访企业内部已经存在未被纳管、处于“野生”状态的未知智能体（Shadow Agents）。 [0:04] Steve Stein: 65% 的企业报告遭遇过与智能体相关的安全事故。 [0:04] Steve Stein: 智能体的初衷是为企业大幅提升生产力，绝不应该反过来损害业务与财务安全。企业必须在赋予智能体必要行动权限的同时，彻底杜绝它们对内部系统造成破坏。 [2:33] Steve Stein: 默认零信任（Deny by Default）：沙箱内部智能体默认无法访问任何外部网络、文件系统或接口。 [2:33] Steve Stein: 显式白名单与告警拦截：对所有外部访问请求逐一审计，确保 IT 部门拥有 100% 的控制权。 [10:46] Steve Stein: 在本地运行开源或专有大模型，相比公有云 API 可实现高达 50% 的成本节约。 [15:36] 现场参会者: 当前行业报告中普遍反映 AI 硬件存在严重的供应短缺与漫长交付等待期。刚刚在其他分会场也听到类似讨论，很多人因此更倾向于选择公有云基础设施而非本地部署。企业在选择本地私有化架构时，究竟该如何应对漫长的硬件就绪周期？ [18:10] 现场参会者: 当企业将模型部署在本地私有服务器上时，研发工程师在使用上会受到哪些限制？工程师是否必须被死死绑定在 HPE 专属软件中，还是依然可以使用他们目前习惯的日常开发工具链？在保证安全可控的同时，安全边界究竟划在哪里？ [19:34] NVIDIA 与 HPE 代表: HPE 名字末尾的‘E’代表着企业级（Enterprise）。

[0:04] 82% of organizations already have unknown or rogue agents operating within their networks [0:04] 65% have actively reported security-related incidents stemming from agentic AI deployments. [0:04] For agents to deliver actual business value, they cannot simply be locked away in an isolated box where they do nothing; they must have the autonomy to call tools, execute commands, and run generated code. [4:21] Inside an OpenShell sandbox, an agent is instantiated with zero external privileges by default. [8:20] Deploying agentic systems on-premises through HPE Private Cloud AI provides up to a 50% cost savings compared to public cloud providers, while ensuring sensitive company data never escapes sovereign enterprise boundaries. [16:32] top-tier Blackwell B200 and B300 GPUs carry the longest lead times, but professional workstation cards like RTX Pro-based GPU servers offer significantly shorter turnaround times and lower capital costs [19:34] the E at the end stands for enterprise [19:34] HPE holds 18 world records on the standardized MLPerf benchmarks while running on the exact same underlying NVIDIA silicon as competitors.

[0:04] Thank you everyone. Uh my name is Steve Stein.  I work over at NVIDIA and today I want to talk [0:08] about how to make agents safe. So if some of you  were around for the talk this morning, I spent a [0:13] lot of time talking about how, you know, the age  of agentic AI is here. The age of useful AI is [0:19] here. We need to learn how to embrace it. We need  to start building AI factories to help solve some [0:24] of these challenges. And more specifically, we can  start thinking about building onrem AI factories [0:29] to solve some of the major challenges that we're  seeing with Agentic AI to really take advantage [0:33] of the power and frankly usefulness of Agentic  AI. So, I'm going to carry on from the talk this [0:38] morning. These were the two topics I really tried  to go over in terms of this issues that we're [0:42] seeing with Agentic AI. One side of the equation  here is about the runaway token cost. Some people [0:46] call this tokconomics. you know, agents consume a  huge amount of tokens, therefore a huge amount of [0:52] cost compared to what happened in past years with  generative AI and rag and everything else that [0:57] came into that. The other problem with Aentic AI  that we're going to spend a lot more time with you [1:03] today is to talk about the security challenges  and specifically we're going to go through I [1:09] want I want to show some data on what kind of  security incidents we're already seeing. You know, [1:14] you know the these things are already top of  mind for most folks. We actually have our black [1:17] hat conference just down the road. I'm sure a lot  of them are thinking about all these different AI [1:21] agent escapes, uh, AI agents causing security  issues, you know, hacking different companies, [1:27] you know, that this is top of mind for a lot of  folks. So, how do we create safe agents inside [1:33] your infrastructure that won't wreck your system?  That's a major problem, a big thing that people [1:37] are thinking about. And so, let's talk a little  bit about some of the uh challenges that other [1:42] companies are seeing. So 82% of organizations  already have unknown agents running inside [1:49] their walls and 65% have reported security related  incidents when it comes to aensic AI. So this is [1:57] not a theoretical problem. This is not something  we're trying to project into the future and think [2:02] this might be happening. These are real companies,  real organizations experiencing agent, you know, [2:07] experiencing issues with agents today. When you  kind of break apart that 65% number, you know, [2:12] 61% have data exposure issues, 43% of operational  disruption. But the thing that really stood out [2:17] to me, 35% have experienced financial losses  because of agents. That's a big deal. You know, [2:24] agents are supposed to be helping you. They're  supposed to be adding productivity to you and [2:28] your organization and to your company. They're  not supposed to be detracting from it. So, how [2:33] can we actually build safe systems? Another thing  to really think about with agents as well is, [2:38] you know, for them to be useful, they really  do do need to have autonomy. They need to be [2:43] safe. They need to have zero trust. They need to  evolve. And so, we can't just keep them in a box [2:48] and tell them not to do anything. They need to  be doing productive things. In order to do that, [2:52] they need access to tools. They need to be  executing commands. They need to be executing [2:57] uh code. That's what harnesses do at their core  is they're the execution arm of generative AI. [3:04] you have gener you're generating tokens from  your you know token generation systems going [3:08] over to the harnesses and therefore executing the  code. They need to be executing useful things and [3:13] and returning frankly valuable results back to  your LLM systems for them to analyze and come [3:19] up with new prompts. But that loop needs to be  secure. That loop needs to be bounded. And so we [3:25] need to be able to create agents that are safe.  We also need to know what's happening with the [3:28] agents inside your organization. You can't be one  of those organizations, the 82% that have unknown [3:33] agents running inside your enterprise. You should  have observability over all of them. And so, how [3:39] do we get there? So, I'm going to talk a little  bit about some solutions we have from Nvidia, [3:43] some solutions we have from HPE that are going to  allow you to spawn safe agents in the first place, [3:49] and then two, HPE is going to give you a lot of  tools with your onrem infrastructure to be able [3:55] to see what's happening with your agents. This is  going to allow you to one of course know what's [4:00] happening with your agents but then two what tools  they're accessing what skills they're accessing [4:05] but also so you can actually protect your entire  data flow as you're building agents as you're [4:09] using them. You don't want that information  going different directions and escaping your [4:14] organization. We want to keep that inside as much  as possible and observing what's happening with [4:18] the agents is one of the tools that's going to  allow you to do that. That's what HP is going [4:21] to be able to provide to you for you. Uh before  I get to that, I want to talk a little bit about [4:25] some of the Nvidia solutions that we have on  the software side for this. So OpenShell and [4:29] Nemo Claw, if you haven't heard those terms  before, uh they're starting to be, you know, [4:34] more filtered out into the server ecosystem.  Nemo claw and open shell have largely existed [4:41] say on desktop systems, but we're introducing this  with HPE into servers into organizations to allow [4:49] you to spawn safe agents. So how does it work? So  or probably more specifically, what is it? So we [4:55] work with some of the major harness providers out  there like Hermes and Open Claw. And we have Nemo [5:03] Claw implementations for each of these harnesses  where you're able to spawn safe agents inside of [5:10] protected sandboxes that have access to nothing.  And I mentioned before, if you have an agent [5:16] that has access to nothing, that's not useful. But  what our software does, what our tools do is every [5:22] single external access that the agent is asking  for has to be explicitly whitelisted by you, [5:29] your IT department, and it flags it to you when  it's trying to access something outside of that [5:34] sandbox. So you have full control over what it's  accessing and what it doesn't because everything [5:40] is denied by default. That's what Nemo claw is  doing. And we work again directly with these [5:46] harness providers. So you can actually use the  actual harness. This isn't some custom harness [5:51] that Nvidia is releasing to the market. You're  working with the actual harnesses here. You know, [5:55] the the Hermes, the Open Claws to be able to get  access to, you know, real harnesses. So you can [6:00] do real agentic AI on top of this, but it's just  protected. We have safe agents inside of these [6:06] boundaries. And so when you look at like from an  architecture diagram point of view, this is what's [6:11] happening inside of the OpenShell sandbox is where  you create your agent. This is where your Hermes [6:18] instance is running. This is where Hermes is  creating your primary agents inside of that. And [6:23] again, if it wants to access anything outside of  that, any kind of tool or skill, everything has to [6:29] be explicitly whitelisted approved before you can  get access to it. And so this is what we're trying [6:34] to come up with is the way to spawn safe agents in  the first place so you don't have runaway agents [6:41] wrecking havoc inside your organization. It's also  going to give you a great leg up when you want to [6:45] observe what these agents are happening because  you know what they have access to and what they [6:51] don't. And on top of that, you know, harnesses  and agents need skills. And what skills are, [6:58] they're they're essentially markdown files that  contain a lot of text that help the harness the [7:04] agent figure out how to do specialized tasks. And  for a lot of these harnesses, you get these skills [7:10] from external marketplaces. And it can be iffy  sometimes if you know if when you're downloading [7:16] a skill, you don't know if it contains malicious  code, if it's going to do something that you don't [7:19] want it to do. Uh but you need these skills for  these agents and these harnesses to really be [7:24] effective. And what we're doing in Nvidia is we're  actually verifying skills end to end that you can [7:31] deploy inside your organization that have been  verified. They're safe. So not only do you have [7:36] safe agents being spawned, but you also have safe  tools, safe skills that we're able to release to [7:42] you. So you can actually keep your again all your  agents safe as they're accessing outside systems [7:46] and they're getting access to skills to make them  more productive. And the the Nemo claw open shell [7:53] community is growing. We're seeing more and  more downloads from GitHub, more GitHub clones, [7:58] and it's just accelerating. I personally have  deployed Nemo Claw on my personal machine. Uh [8:03] it's not that hard. If I can, trust me, if I can  learn how to do it, you can learn how to do it. [8:06] It's not that it's not that complicated. and plus  and so we're seeing a lot a big uptake and you [8:11] know the amount of you know get home clones and  downloads we're seeing with Nemo claw is something [8:14] the more people are interacting with. So that's  the Nvidia part of the solution here where we're [8:20] trying to make agents safe uh in the first place  while they're being spawned. Uh for HPE, a huge [8:27] portfolio of solutions for you to make systems  secure as they're as they're deployed on prem. [8:32] And I always like to say, you know, HPE bears  the operational burden of making this work inside [8:37] your organization. That's what they're going to do  for you. And they have a huge suite of solutions. [8:41] We're partnering with them very heavily to bring  solutions to market. And everything that you see [8:46] on the screen, everything that I'm talking about  today is very hard behind the scenes. But trust [8:50] me, when you see it from your side and you're  you're you know getting HP to kind of help you [8:54] deploy these systems, it'll be very easy and  you're going to get access to these systems, [8:58] you're getting access to the to the right tools  to be able to deploy it up and down up and down [9:02] the stack here. So everything from, you know,  enterprisegrade AI with Proline RTX Probased [9:08] GPUs that are designed to run a huge range of  workloads and be very efficient or if you need AI [9:13] factory solutions that either get to you quickly,  we call that turnkey AI factories, you know, or [9:18] the private cloud AI solutions that allow you to  get access to to to these systems incredibly fast. [9:24] They're fully managed software stacks on top of  that with a Kubernetes layer on top of there. Um, [9:29] so you can get you you you can get these the these  uh these systems in much faster or if you need to [9:34] scale HPE AI factories at scale. And one thing if  you don't know this, HPE builds some of the most [9:42] some of the biggest most complicated systems in  the world. If you just look up the top 500 list, [9:49] HP is always at the top. HP knows how to how  to how to deliver performance. They know how to [9:54] how to deliver systems that work. They know how to  deliver safe systems for your organization. Again, [9:59] they're partnering very heavily with NVIDIA  to make sure all of these systems are tuned. [10:02] They're ready to be deployed. They're a wonderful  partner and I think they're very well very well [10:08] positioned to be able to take advantage of this  of this agentic moment, not just with tokconomics, [10:12] but specifically on the security side and making  sure that these systems actually work inside your [10:17] organization. And then if you need a boundary  around some of these AI factories, you know, [10:21] sovereign solutions for this, they have a whole  suite of solutions. air gap solutions to make [10:26] sure that these systems are protected uh from  uh from the get-go and also I mentioned before [10:32] they they make some of the largest most powerful  systems in the world uh with their supercomputing [10:37] craze solutions. So HPE, wonderful partner. We  love them. I love them. Uh they and they can get [10:41] you the systems that you need for this. And on top  of being able to provide the hardware, the systems [10:46] when you need it, they have a wonderful software  stack that adds value over and above anything [10:50] that Nvidia can provide. And specifically, if you  look at that layer, the HPE AI factory composer, [10:56] that layer specifically in the management  platform, they're going to provide you a level [11:00] of observability of your agents. I mentioned  before in a lot of organizations have you know [11:04] 82% of them have unknown agents running inside  their organization and 65% of them are causing [11:09] security related incidents. Wouldn't it be better  if you can just observe what's happening with all [11:13] these agents and not just know you know what  they're up to, what tools they're accessing, [11:17] maybe even how many tokens they're generating  per agent, which agents are more productive, [11:22] which agents are less productive. That's what  this observability layer is going to be able [11:26] to give you is this, you know, ability to kind of  observe what's happening across your organization [11:30] is all these agents are being spawned to create,  you know, to essentially kind of create value for [11:35] your organization. And you know, Nvidia Nemoclaw,  OpenShell, it's really kind of built on, you know, [11:42] first principles. It's zero trust by default. As  I mentioned before, uh we're really again trying [11:46] to create safe agents, solve this security  challenge that exists inside of the agentic [11:51] AI world so we can get productive agents being  deployed inside of organizations and HPE is one [11:59] of the first OEMs to bring OpenShell and Nemoclaw  into the server environment. I mentioned before, [12:07] you know, from March up until now, all of the Nemo  cloud deployments have happened on desktopbased, [12:13] you know, computing systems like a DGX Spark  or DGX station or name your other, you know, [12:18] OEM desktop provider. That's how they've mostly  been deployed and it's for developers working [12:22] inside of a system or inside of a system that's on  a desk. We're trying to bring this over to again [12:29] the server world. We're also, you know, HP does a  wonderful job of making this very usable as well. [12:35] Now, I'm sure we got a lot of developers in here  that love working with command line interfaces [12:38] all the time, but sometimes it's a little bit  better to work with a more interactive guey, [12:42] something that's a little bit more intuitive to  see. And so, that's what HP is going to give you [12:46] is this kind of the ability to kind of interact  with it a little bit more intuitively versus [12:49] nothing but command line interfaces, which is kind  of what Nemoclaw is using right now. But to get to [12:55] this agentic, it's a a secure agentic enterprise.  We're we're working very closely with HPE to bring [13:02] our capabilities into their systems, specifically  Nemo Claw, so you can spawn safe agents. Then you [13:07] add on top of all the great capabilities that HP  has to create safe systems, observable systems, [13:12] and that's what they're going to be able to be  be able to provide with you. And just some more [13:15] facts about how great HPE is. You know, they're  named as a leader in uh specifically, you know, [13:21] for for infrastructure for AI. Um, everything's  co-engineered, co-developed with us. And when [13:26] you look at the cost savings, this is something  I'll get into later. I'm actually coming back and [13:32] talking tomorrow about the tokconomics side of  this equation. But the main reason why you want [13:38] to deploy on prem is cost. Runaway token costs  are bank not really bankrupting, but you know, [13:44] a lot of organizations out there, if you just  read the news, you know, a lot of them have burned [13:48] through their entire yearly budgets of tokens  within months. And the best solution is to build, [13:54] you know, at least the the solution to that  problem is to build those on-prem systems, [13:58] run local AI, run local LLMs, and that's how you  can get access to AI systems at relatively low [14:03] cost. Specifically, if you look at how the cost  of, you know, HPE is compared to the public cloud, [14:09] we're looking at 50% cost savings. This isn't just  like a 10 20% 15% cost savings. This is massive [14:14] cost savings over the long run. And you get all  the benefits of on-rem infrastructure. You get [14:19] all the benefits of owning your data, owning your  pipeline, um, owning your data. It's not going [14:24] off-rem other places, and you get all the security  benefits of running all your agents on prem, [14:28] the observability. There's a lot of reasons  why it makes sense to run your agents on prem, [14:33] specifically with HPE. And they have the PC AI  agent platform. Again, that's private cloud AI. [14:39] That's their platform to get you systems quick.  And they have a number of different harnesses [14:45] and you know there are these abilities to kind of  kind of be be built into that governance layers [14:50] agent runtime layers access to frontier models  if you need that. Uh so lots of reasons and then [14:57] I always like to you know close with this slide  here that really kind of brings it home for why [15:02] HP really can be your trusted source for the  aentic enterprise but specifically how they [15:06] solve the specific challenges for aentic AI. So  namely kind of again by eliminating the runaway [15:12] token costs security by being able to spawn  the safe agents and providing those air gap [15:18] solutions around your infrastructure providing  that observability layer which I don't believe [15:22] anyone else really can at this point and again  it's all co-engineered co-developed with Nvidia. [15:26] So you know you're getting access to the best  most token efficient systems by working directly [15:31] with Nvidia to be able to provide those systems  that work for you. So that's all I have for you [15:36] today. I am open to questions. Uh we got some  experts here from HBE to help answer questions. [15:41] If there are any questions in the audience, feels  you feel free to stand up, raise your hand. If you [15:46] don't want to ask question, that's fine. I'll be  outside after the talk and uh and and so will some [15:51] of our HP friends as well to be able to answer any  questions you have. But if there's any questions, [15:56] go ahead and raise your hand, ask it now.  Otherwise, we can give you some time back. [16:03] Does anyone have any questions? I've got the  microphone here. Oh, we've got someone over here. [16:07] Oh, we are seeing in reports that um hardware  availability is not that much you know like [16:16] setting up the onremise system requires it's a  long waiting time to get all the infrastructure [16:21] set up like you I was coming from the another  session and there was a conversation going [16:25] around the same thing that preferably to go for a  cloud setup for infrastructure than the on-remise [16:32] because it's it's kind of a waiting of a year or  so is that what's What's your opinion about it [16:38] from Nvidia and HP perspective? That's a great  question and it's it kind of speaks, you know, [16:43] speaks to those like supply chain issues and  also just frankly how there's a huge demand [16:46] for Nvidia GPUs out there. This isn't unique to  it's it's not even unique to the cloud players. [16:52] The hyperscalers are getting access to some of  your Rubin systems that are coming to market [16:55] right now. So this is a something that's top of  mind. But I will say you know with HPE they have [17:00] different kind of tiers of offerings for lack  of a better word. Kind of walk through like [17:04] the AI factory tiers. They got that turnkey,  they got the at scale, they got the sovereign, [17:09] the at scale and sovereign side. Yeah, that  may take a while just because of, you know, [17:13] the scale that you're talking about, but great  thing about HP, they have access to developer [17:17] systems. They're they're delivered pretty  quickly. Same thing with the PC AI. It's kind of [17:23] designed to be delivered much more quickly from a  development point of view. So, you can get access [17:27] to these systems faster. That is a larger issue  though. I don't want to discount that. You know, [17:31] for some of these larger systems, the backlog  is pretty large. I would also encourage to maybe [17:35] think about the workloads a little bit because  we have very powerful GPUs. Our B200, B300 GPUs, [17:42] they're very powerful. Those are the ones that  are on the longest back orders right now. RTX Pro [17:48] servers, RTX Probased GPU cards have a little  bit lower lead times. They're also relatively [17:56] cheap compared to those B2300s. But the main  limitation there is model size. So, if you can [18:01] fit your models into those GPUs, that's another  great option that I'd look at. Great question. [18:10] Hi. Uh so one of the big things I I'm trying  to wrap my head around with the on-prem [18:18] uh like local models is you know if you have  you know all your engineers interacting with [18:24] an on-prem server for that like what are what  are like the limitations on the engineers like [18:32] are they only interacting through the HPE  software or are they still available to use [18:39] you know the same types of tools they're using  today or like like where like where where does [18:44] the security limit you I guess is what I keep  having a hard time wrapping probably that's a [18:51] I'll go ahead question I can take that question  so it's really your choice on how you configure [18:56] the platform the platform is fully extensible  and flexible in that it has a data lake house [19:02] it has data integrations and it has the ability  to have your global namespace all of this to put [19:07] together what it means is you can send some of  your loads to public cloud you can even choose [19:11] hey I want this to go to cloud this one to go  to Gemini and something else and rest of this [19:16] I want to run on prem and you can use the  AI gateway there's a built-in AI gateway [19:20] in private cloud AI and larger systems you can  configure it similarly so you have the options [19:34] hi uh actually I just came from the exhibit  hall so I Um I went to Dell booth. They also [19:42] have this AI factory where you can download the  model. I'm just wondering I'm curious about how [19:48] it's different from HP AI factory. Very different.  So um what I went through with that observability [19:55] layer HPE you know there's also a product  called ops ramp. Um that's the main difference [20:02] in my opinion. Also HPE is very well known for  building the most powerful systems in the world. [20:08] HPE owns 18 world records on the ML  Perf website and it's access to the [20:13] same hardware that Dell has. So, you know,  they're doing something right there. Um, [20:17] and and they're also known for providing very  trusted systems into the market as well. So, uh, [20:22] there are a lot of natural advantages I believe,  you know, biased around HPE versus Dell. And, [20:29] uh, we we we believe that that that there are  a lot of great advantages there. The only the [20:33] only main difference that you keep on your mind is  you know Dell does provide desktopbased systems as [20:38] well. HPE the E at the end stands for enterprise  and therefore servers is what HPE focuses on and [20:45] specializes on. They have a lot of expertise  in that and again most powerful systems in the [20:49] world. You can't really discount that, you know,  incredible supercomputing expertise inside of HP.