Enterprise AI Leaders Pivot to Layered Infrastructure Stack Strategy

企业AI领袖转向分层基础设施架构策略 企业应用需要采取混合策略，在各个不同的架构层级评估“自研还是购买”，而非做出单一的二元抉择。 与会嘉宾深入剖析了企业级 AI 基础设施复杂的经济机制与运营现实，审视了从直接租用 API 到完全自有的裸金属技术栈等全谱系的 Token 消耗模式。他们强调了隐性基础设施成本——例如补贴下的 Token 定价、专业化实体劳动力和电力供应限制——并反对将“自研还是购买”视为单一的二元选择。相反，普遍共识倾向于一种灵活的“自研与购买并存”模式，即企业对具体的架构层逐一评估，并在具备适应性且不绑定特定厂商的平台上进行构建。 Meta 从零开始构建自身的基础模型，以完全掌握自身命运，而非依赖第三方提供商。 (20:08) AI Token 消耗涵盖三种主要模式：直接租用 API（纯 OpEx 模式）、用于自主定制控制的推理合作，以及自有的裸金属技术栈（侧重 CapEx 模式）。 (20:08) 像制造工厂等需要 24/7 全天候运行的关键业务，需要本地边缘模型以避免停工风险，而常规用例则可依赖云基础设施。 (20:08) 受风险投资和科技巨头高额补贴的影响，当前商业 Token 价格被人为压低。 (23:22) 自建 AI 基础设施需要完全承担无补贴的运营成本，包括硬件采购和巨大的能源消耗。 (23:22)

Rather than approaching AI infrastructure as a binary choice between building or buying, organizations must evaluate individual architectural layers to implement a hybrid strategy that balances strategic independence, physical deployment realities, and capital allocation. Token Consumption Models and Strategic Independence (20:08) Devang Modi notes that tech giants like Meta spend billions to recruit elite talent and build proprietary foundation models to retain control over their destiny and avoid future lock-in. Ryan Horjus outlines three distinct operational models across the token consumption spectrum: renting retail tokens through direct APIs (pure OpEx), partnering with specialized inference providers for better infrastructure control and lower rates, and owning bare-metal hardware and data center contracts (CapEx-heavy with minimal ongoing OpEx). Modi observes that operational requirements dictate where workloads belong. High-stakes 24/7 environments, such as manufacturing plants, require locally hosted edge models to prevent catastrophic factory halts, whereas standard enterprise applications can comfortably rely on rented cloud environments. While startup burn rates often compel pure OpEx models, mature enterprises frequently shift toward ownership to mitigate escalating cloud rental expenses. Subsidized Token Pricing and True Ownership Costs (23:22) David Chun points out that current commercial token pricing is artificially deflated by heavy subsidies from venture capital firms, major technology corporations, and government grants. When an enterprise elects to build and host infrastructure internally, it must directly absorb the un-subsidized reality of capital expenditures, most notably specialized hardware procurement and massive electricity consumption. Despite these long-term cost differences, renting cloud capacity or purchasing third-party vendor solutions enables executives to book expenses under OpEx. This accounting distinction remains attractive for financial planning, even as companies must weigh the convenience of vendor-managed services against the eventual risk of price adjustments once market subsidies diminish. Physical Deployment Bottlenecks and Model-Agnostic Frameworks Physical deployment remains a primary constraint for scaling AI infrastructure. Expanding capacity relies directly on traditional trades—such as running pipes, laying cables, placing transformers, and installing power whips. Due to the high financial stakes and tight delivery timelines, enterprises are paying substantial wage premiums to secure this specialized skilled labor force. To hedge against market volatility, the panel emphasizes the necessity of maintaining model-agnostic infrastructure frameworks. Because leadership cannot predict which underlying large language model will ultimately dominate, an agnostic foundation ensures that organizations retain data portability and complete ownership of proprietary assets, allowing them to switch model providers without dismantling their core systems. Architectural Layering and Targeted Capital Allocation Panelists caution against making an organization-wide, all-or-nothing commitment to either building or buying. Devang Modi advises breaking down the technology stack and addressing the build-versus-buy decision at each individual architectural layer. Supporting this hybrid approach, Mike Burgio recommends purchasing adaptable platforms that teams can build upon without becoming locked into rigid vendor ecosystems, ensuring cost predictability amid rapid industry shifts. Capital allocation must also correspond to clear business outcomes. David Chun urges enterprise leaders to focus their AI deployments on their most difficult, previously unsolvable operational challenges to secure measurable return on investment, while Ryan Horjus emphasizes determining which specific tier of the infrastructure stack produces the highest return before deploying capital. The panel concludes that successful enterprise adoption rests on a nuanced 'build and buy' framework rather than a single procurement decision. However, organizations must navigate practical physical constraints, including trade labor shortages and rising electrical overhead, while structuring their infrastructure to prevent vendor lock-in as market models and subsidized token economics shift.

与会专家指出，企业在AI基础设施决策中不应将“自主构建”（Build）与“外部采购”（Buy）视为非此即彼的二元对立，而应转向按技术层级评估的“构建加采购”（Build and Buy）混合策略。各方在选择技术路径时，需权衡战略自主权、底层物理基建的硬性约束，以及风险投资对云端代币价格补贴所掩盖的真实总拥有成本。 战略自主权与代币消费层级 (20:08) 针对企业为何自建大模型，Devang Modi引述Meta内部团队的说法指出，Meta投入数十亿美元招募顶尖团队自研基础模型，核心动机是不受外部供应商的技术牵制并完全掌控自身命运。Ryan Horjus据此梳理了企业消费AI代币与规划成本结构的三种典型层级：第一种是直接调用API模式（纯运营支出OpEx），即直接采购OpenAI或Anthropic（Claude）等供应商的零售代币，该方式启动迅速，但规模化运营后持续成本高昂；第二种是与专门的推理服务商合作，在获取部分基础设施控制权的同时压低代币采购单价；第三种是自建裸金属技术栈（资本支出CapEx主导），企业直接持有数据中心合约、电力与物理硬件，自托管软件以最小化运营成本。 Devang Modi进一步补充，不同的业务连续性需求决定了部署形态的分野。例如对于需要全天候24小时不间断运转的制造工厂，边缘模型必须直接部署在本地以彻底杜绝停产风险；相比之下，普通企业级用例则可灵活采用租用云端基础设施的方案。此外，初创企业受制于资金消耗速率通常倾向纯OpEx模式，而成熟企业为避免遭遇天价云账单，更倾向于自购或掌控物理资产。 代币补贴假象与显隐性成本 (23:22) David Chun揭示了当前AI代币市场中被掩盖的经济规律。他指出，由于风险投资机构、科技巨头及政府补贴的大量注资，当前市面上的代币价格被严重人为压低。一旦企业决定完全在内部自主构建AI基础设施，就必须直面未经补贴的真实成本，这包括昂贵的硬件采购开销以及极其高昂的电力能源消耗。 尽管自建方案面临巨大的显性与隐性成本，但通过采购第三方厂商方案或租赁云端代币，管理层能够将相关支出归类为运营支出（OpEx）。这种模式高度契合财务主管（CFO）对于控制现金流波动与优化财务报表的诉求。企业在测算总拥有成本（TCO）时，若忽视了外部代币生态中隐藏的巨额补贴，可能导致对自建或外购的经济效益产生严重误判。 物理基建制约与技术中立框架 围绕AI基础设施扩展的物理约束，讨论强调硬件系统的实际落地高度依赖传统土木与机电工程，包括管道敷设、电缆铺设、电源配线安装以及变压器布置等底层工作。AI技术的飞跃并未取代实体工种，反而催生了对传统熟练技工前所未有的旺盛需求。面对巨大的资金投入与紧迫的交付工期，各类组织不惜支付巨额溢价以确保专业劳动力按期交付，这表明AI的进步无法免除企业在劳动力物流与工期管理方面的基本功。 针对快速变化的技术市场，与会专家一致认为建立技术中立框架（infrastructure-agnostic framework）至关重要。不论未来哪个大语言模型（LLM）在市场竞争中胜出，中立架构都能保障企业随时切换底层技术以规避供应商锁定。目前传统基础设施供应商也深刻意识到企业对数据主权与资产可移植性的核心关切，企业在引入开源、闭源或专有方案时，必须始终将保留对自有资产与数据的绝对控制权作为首要前提。 架构分层评估与混合落地策略 在总结企业落地的核心战略时，与会嘉宾达成共识，强调必须跳出非黑即白的思维定势，转向精细化的“构建加采购”（Build and Buy）哲学。Devang Modi建议，决策者切勿对整个技术栈做出一揽子的二元选择，而应将整体架构切分为独立的层级，逐层推敲自建与外购的匹配度，从而拼装出结构平衡的综合解决方案。Mike Burgio同样提倡弹性的混合策略，主张企业应优先采购能够在其上进行二次构建的平台方案以避免技术锁定，在敲定具体工具前先确立可预测的成本结构与灵活的技术底座。 从资本与业务价值出发，Ryan Horjus强调企业必须明确自身在资本配置层面的战略定位，看清技术堆栈中的哪一层能够带来最大投资回报，再决定是独立攻关还是与高阶供应商及咨询团队合作。David Chun则告诫企业领导者，应将精力和资源投向具备清晰ROI的重大应用场景，特别是锁定那些过往无法攻克的复杂企业难题，以此释放AI基础设施落地的最大价值。 本场研讨表明，AI基础设施建设不存在一劳永逸的最佳方案，企业必须在自主控制权、成本弹性以及严苛的物理制约之间持续权衡。尽管专家一致推崇按技术层级进行混合部署的务实路线，但考虑到大模型迭代日新月异且外部代币的人为补贴机制难言长期稳固，企业如何在保持基础设施敏捷性的同时兼顾资本投入产出比，依然是亟待在实践中摸索的关键课题。

## Background 该内容由 YouTube 频道 Ai4 于 2026 年 8 月 21 日发布。视频记录了一场关于企业 AI 基础设施决策的圆桌研讨会，来自哥伦比亚大学、OutSystems、Hydra Host、博士伦（Bausch & Lomb）及德国 NGU 的行业领袖与学者共同探讨了企业在 AI 时代面临的自建与采购（Build vs. Buy）战略权衡、供应商锁定风险、数据主权与成本控制等核心议题。 ## TL;DR 圆桌嘉宾一致指出，AI 时代的自建与采购已不再是非黑即白的二选一，而是演化为“分层决策”与“自建且采购（Build AND Buy）”的混合模式。在底层算力与前沿基础模型层，企业宜通过采购与租赁快速换取效率；而在智能体编排、业务上下文与私有数据等核心壁垒层，则必须牢牢自建并掌控。面对快速演进的模型与高昂的隐形基础设施成本，企业需要借助中立且具高可移植性的解耦架构来规避供应商锁定，并在合同中设立明确的退出机制，以平衡速度、合规与长期财务风险。 ## Atmosphere & style 整场研讨会专业务实且节奏明快，嘉宾们摒弃了空洞的行业炒作，直击企业在部署 AI 时面临的真实痛点与财务算盘。主持人的学术视角与来自大药企、基础设施提供商、低代码平台的嘉宾形成了多维度的思维碰撞，既有探讨电力容量与线缆铺设等“硬核物理现实”的接地气对话，也有对供应商锁定和代码所有权的坦诚交锋与幽默互动，整体氛围既严谨深刻又充满启发性。 ## Key takeaways - [05:09](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=309s) 企业在 AI 时代的决策模式已从简单的“自建 vs. 采购”转变为分层组合的“自建且采购”（Build AND Buy），必须按具体架构层级评估而非整栈一刀切。 - [05:09](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=309s) 通用算力与前沿基础模型层应果断采用采购或租赁以换取落地速度，而编排层、企业记忆、上下文及专有业务数据属于核心护城河，必须由企业深度掌控自建。 - [14:55](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=895s) 随着原型开发门槛大幅降低，“灵活性 vs. 速度”的争论核心已转向“风险管理”：自建需从零搭建合规护栏，而采购成熟平台可直接获得企业级安全与治理防护网。 - [10:04](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=604s) 为规避供应商锁定风险，企业应推行“价格翻倍测试”并在合同中设置 60-90 天无痛退出条款，同时优先选择支持底层模型与组件自由解耦插拔的混合架构平台。 - [20:51](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1251s) 公有云 Token 报价多包含风投及巨头补贴并非真实完全成本，企业在评估自建 TCO 时需全面核算机房电力消耗、物理硬件折旧以及蓝领工程技师等高昂隐形成本。 - [20:51](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1251s) 由于医疗（受 PHI 约束）与制造业等行业对数据出境和生产连续性有极高要求，企业必须在厂区部署本地化的 Edge 模型以确保业务永续。 - [25:08](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1508s) 采购部门的考核红线已从传统的合同 SLA 转向架构级的可移植性、数据主权及多维风险管控（包括 PHI/IP 泄露与 Token 账单爆炸等）。 - [30:29](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1829s) 无节制将多源数据与合成内容塞入模型会导致上下文窗口耗尽并引发决策退化，面对频繁变动的闭源模型，适时转投可控的开源开放权重（Open-weight）自建方案能沉淀宝贵经验。 - [36:34](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=2194s) 现代 AI 工具大幅压缩了遗留系统迁移的代码理解成本，且当前 LLM 已能胜任 70%-80% 的常规知识工作，企业可将成熟工作流固化于本地私有环境以降低外部依赖。 - [38:40](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=2320s) 算力底层正回归高功耗与物理施工等传统基建现实，企业需借助开源工具链构建中立于上层模型与硬件 OEM 的解耦底座，以应对漫长芯片交期与电力短缺限制。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=4s) 开场介绍与圆桌嘉宾背景 会议由德国纽廷根-盖斯林根大学（NGU）应用人工智能教授 Andrea Spielvogel 主持。Andrea 介绍自己的研究领域主要聚焦于 AI 用户创新（User Innovation）以及伴随而来的治理、风险与合规（GRC, Governance, Risk, and Compliance）。在业界工作数十年后，他于去年转入学术界，致力于更具使命感的 AI 项目，目前正与图宾根大学医院（University Hospital Tübingen）合作，研究如何利用 AI 辅助脑癌患者进行共同决策（shared decision-making），并改善医患关系。Andrea 提到他近期刚走访哈佛大学（在开放与用户创新会议上发表关于 AI 合规与治理盲点的研究）及加利福尼亚大学伯克利分校（参加 Agentic AI 峰会，探讨智能体基础设施及伦理与可持续发展）。 随后，三位圆桌嘉宾依次进行了自我介绍： * **Ryan Horjus**（Hydra Host 部署副总裁）：负责 AI 基础设施部署，工作处于供给与需求、即自建与购买中资本支出（CapEx）与运营支出（OpEx）的交汇点，协助各类企业评估是自持基础设施还是选择租用。 * **Mike Burgio**（OutSystems 解决方案架构高级经理）：带领 AI 解决方案顾问团队。OutSystems 是一个智能体系统平台（Agentic systems platform），为企业提供在全技术栈中构建、部署、管理和治理企业级应用与智能体的能力。 * **Devang Modi**（博士伦 Bausch & Lomb 首席 AI 官 CAIO）：专注于企业的应用 AI 落地。在自建与购买的决策中，他通常站在“买方（Buy-side）”的视角，并在实践中依据技术演进频繁调整策略。 --- ### [05:09](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=309s) 现实落地案例：以 AI 为中心重塑业务流程与分层决策 主持人引导嘉宾分享近期的实际 AI 部署与基础设施项目。Devang Modi 结合博士伦的实践指出，企业现阶段思考 AI 的方式已经发生了根本性转变：过去大家只是在现有工作流中局部加入 AI 进行功能辅助（Augment），而现在则是**将 AI 置于核心，彻底重塑端到端的工作流**。 Devang 以企业内部正在推进的端到端“寻源到付款”（Source-to-Pay / Procure-to-Pay）流程重塑为例，说明了使用智能体（Agents）重构业务的全过程。在这一过程中，自建与购买不再是 10-15 年前“自建 CRM 还是购买 Salesforce”那样非黑即白的简单选择，而是演变为了**按技术架构分层决定自建还是采购**： ```mermaid flowchart TD accTitle: AI技术架构分层自建与采购决策 accDescription: 展示从底层基础设施、前沿大模型到编排层、数据上下文层的自建与购买权衡 subgraph CommodityLayer ["通用标准层（建议采购/租用）"] A["底层算力与硬件基础设施 (GPU/Data Center)"] B["前沿基础模型层 (Frontier LLMs)"] end subgraph EnterpriseLayer ["核心差异化层（建议自建/深度掌控）"] C["智能体编排与工作流 (Orchestration Layer)"] D["企业记忆与上下文 (Memory & Context)"] E["专有业务数据与护城河 (Enterprise Data & Moat)"] end A --> B B --> C C --> D D --> E ``` * **基础模型与算力层**：绝大多数非底层研发企业都不应该自己去训练基础大语言模型（LLM），直接采购或租用算力与模型才是合理的选择。 * **编排、记忆与数据层**：编排层（Orchestration）、记忆（Memory）、上下文（Context）以及专有数据，是企业真正的竞争优势与护城河所在，必须牢牢掌控在自己手中。 * **自建的边界考量**：虽然企业具备研发自建的技术能力，但受限于资源、研发时间周期以及最终的投资回报率（ROI），只有那些能为企业建立深厚护城河的端到端核心业务系统（如 Source-to-Pay 核心流）才值得投入资源自建。 --- ### [07:32](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=452s) 供应商锁定疑虑与算力基础设施的解耦 Andrea 谈及自己在担任首席数字官（CDO）时最担心的痛点——**供应商锁定（Vendor Lock-in）**，即企业购买基础设施后是否会陷入与特定供应商深度绑定的困境。 针对算力与基础设施层面的灵活性，Ryan Horjus 指出，AI 浪潮带来了向数据中心基础设施的某种回归。传统的普通云工作负载迁移非常方便，但 GPU 基础设施不仅极其昂贵，管理模式也大不相同： * **算力与模型解耦**：企业既可以选择自持硬件，也可以选择向云厂商租赁，这取决于企业的资产头寸与资源消耗率。当下的 AI 工作负载具有很强的平台无关性（Agnostic），因为核心在于运行模型，企业可以通过驱动更新等机制在同一套硬件基础设施上快速切换不同的工作流与模型。 * **控制权的核心在数据**：运行在基础设施上的企业数据资产，才是决定控制权归属以及权衡自建与购买的核心考量点。 --- ### [10:04](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=604s) 混合架构应对锁定风险与商业可持续性评估 Mike Burgio 从供应商的角度坦率表示，客户对供应商锁定的担忧完全合理，企业必须在选型时严密考量可移植性（Portability）与灵活性。 * **“价格翻倍”测试**：Mike 引用行业观点发问——“如果你的 AI 供应商明天把价格翻倍，你能在 90 天内把应用无缝迁移到另一套系统吗？”大多数企业目前都无法做到。 * **混合解耦平台**：现代平台架构（如 OutSystems）通常提供混合架构方案，底层支持对接不同的基础大模型或异构数据源，允许企业在保留核心业务资产的同时具备自由切换组件的灵活性。 * **评估供应商的财务与生存风险**：Devang 从实际采购视角补充指出，应对供应商锁定不仅要看技术架构，还要深入考察 AI 初创公司本身的财务背景与生存能力。大量缺乏资本支持的新兴 AI 企业面临被并购或破产的风险，一旦停止研发投入，采用其方案的企业就会蒙受巨大损失。 * **底层放开，核心收紧**：在基础设施与前沿基础模型层，企业应果断签约采购以换取效率；而在企业数据与记忆层，则必须保持最高的自主度与灵活性，绝不能妥协。 ### [12:48](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=768s) 技术栈定位与避免供应商锁定 在基础设施的规划中，企业必须清楚自己所处的技术栈层级。如果企业的数据被死死锁定在某一家模型供应商（如 Claude/Anthropic）那里，未来想迁移到其他模型就会变得极其困难。因此，弄清楚数据主权和技术栈的分层是企业在选型时首要考虑的问题。 Hydra Host 的 Ryan Horjus 指出，基础设施的自建与采购完全取决于企业在技术栈中所处的位置以及内部所拥有的技术能力： - 如果企业处在底层，可能会选择采购裸金属服务器（Bare Metal）； - 如果企业缺乏设计推理层（Inference Layer）或 Token 处理层的工程人才，就会直接购买现成的 Token 服务； - 越来越多的企业开始采用“裸金属上的 Kubernetes”（Kubernetes on Bare Metal）方案，通过直接剔除虚拟化层，不仅降低了性能损耗，还有效消除了底层云厂商的锁定风险。 ```mermaid flowchart TD accTitle: AI 技术栈分层决策流程 subgraph S1["底层基础设施"] A["裸金属服务器 (Bare Metal)"] B["裸金属 Kubernetes (消除虚拟化锁定)"] end subgraph S2["模型与推理层"] C["自建推理与 Token 流水线"] D["直接购买 Token (API 调用)"] end subgraph S3["应用与治理层"] E["企业级统一治理与安全防护网"] F["各业务部门自主构建 AI Agent"] end A --> C B --> C C --> E D --> E E --> F ``` --- ### [13:57](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=837s) 哥伦比亚大学的责任 AI 理念 来自哥伦比亚大学（Columbia University）的 David Chun 随后加入讨论并介绍了哥大的研究侧重点。他强调，哥伦比亚大学非常注重**责任 AI（Responsible AI）**与透明度： - 他们不会直接采购现成的“开箱即用”商业软件，尽管也会投资和使用如 OpenAI、Anthropic 等前沿大模型，但在工具和自动化流程的引入上极其审慎。 - 哥大的核心任务是严格审查 AI 系统的责任机制与透明度，优先选择具备可解释性（Explainable AI）、能清晰展现决策逻辑并且具备完整可审计性的 AI 引擎。 --- ### [14:55](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=895s) 灵活性与速度之争转向“风险管理” 针对“自建的高灵活性”与“采购平台的快速上线”之间的博弈，OutSystems 的 Mike Burgio 提出了一个全新视角：如今“灵活性 vs 速度”的争论已经不再是单纯的部署周期问题，因为现阶段几乎任何人都能极快地搭建出原型和早期应用。真正的核心矛盾已经转移到了**风险管理（Risk Management）**上。 如果企业选择完全自建，就必须从零自建所有的安全合规体系与防护栏（Guardrails）；而如果采购一个现成的基础平台，平台本身自带企业级治理能力，企业在平台之上进行定制化构建，既获得了快速上线的速度，又能享受到覆盖整个企业技术栈的统一安全防护。 - **原型 vs 生产的鸿沟**：博士伦（Bausch & Lomb）的 Devang Modi 提到，作为开发者，自己写代码搭 Agent 非常轻松，但当企业面临 13,000 名员工同时在构建各种 Agent 时，如何建立端到端的治理体系、如何确保投资回报率（ROI）、如何在关键节点设置人工介入（Human-in-the-loop）、以及如何保护知识产权（IP）和患者健康隐私数据（PHI），才是真正困难的挑战。 - **企业的本质目标**：制造型或医疗企业不可能去无限招聘专职开发者；企业的核心使命是把产品交付给患者。因此，底层的安全、连接与治理平台必须通过“采购”快速筑牢，在上层再赋予业务团队足够的自主权去构建。 --- ### [18:13](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1093s) 混合策略与自主掌控权：Meta 的前车之鉴 针对自建与采购的历史演变，David Chun 剖析了现代企业面临的技术困境。过去，企业凭借优秀学府培养的顶尖工程师和内部私有数据，可以完全在内部自建；但如今 AI 已经全球化，性能超越顶尖闭源模型的新开源权重模型（Open-weight models）几乎每天都在涌现，企业必须采用“能够随时调整转向（Pivot）的混合策略”。 随着消费级 AI（如 iPhone 时代的消费级应用、桌面端的 Gemini 等）的爆发，终端用户的体验倒逼企业 IT 追赶。然而，大型复杂企业受到法务与合规制度的严格制约，自建不仅昂贵，而且测试流程冗长。David 特别分享了他与 Meta 团队的交流案例： - 当被问及“作为一家广告巨头，为何要耗费数十亿美元组建顶级团队自建基础大模型”时，Meta 团队回答：“你绝不会希望在遥远的未来被困在别人的模型里。” - 自建的核心诉求在于**掌控自己的命运（Control over your destiny）**。因此，企业的最佳平衡策略是将自身最擅长、最核心的资产（如数据与核心壁垒）留在内部自建与保护，而对非核心的通用能力坚决进行采购。 --- ### [20:51](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1251s) AI 基础设施的隐形成本与三种消费模式 Ryan Horjus 与 Devang Modi 详细拆解了企业消费 Token 及建设 AI 算力时容易忽视的隐形财务与运营成本。在算力采购策略上，企业通常会经历三个阶段的权衡： | 模式层级 | 基础设施形态 | 成本结构 | 适用场景与权衡 | |:--- |:--- |:--- |:--- | | **1. 零售 API 模式** | 直接购买 OpenAI / Claude API | 纯 OpEx（按量付费） | 适用于起步探索，但大规模调用下账单极易失控 | | **2. 推理服务商模式** | 与专业推理托管商合作 | 相对便宜的 Token 费用 | 对基础设施有部分控制权，依然按 Token 结算 | | **3. 基础设施自建模式** | 签约数据中心、承包电力、自购裸金属设备 | 大额 CapEx + 少量机房租金 OpEx | 适合业务规模大、需自托管平台/数据流水线、或利润敏感型企业 | - **制造业的边缘需求**：Devang Modi 补充道，制造业存在必须 24/7 不间断运行的特殊场景，为了防止断网或外部故障导致产线停摆，必须在厂区部署本地化的 Edge 模型。这类涉及生产连续性的场景必须自建并部署在内部，而常规通用场景则直接租赁即可。 - **被补贴的 Token 幻象**：David Chun 提醒大家，目前市场上很多公有云的 Token 单价其实受到了风投机构、科技巨头甚至政府的大规模补贴，并不是真实的基础设施完全成本。一旦企业决定完全在内部自建，就必须直接面对包括高昂电费在内的所有隐形成本。 ### [023:46] 基础设施采购的范式转变与总拥有成本（TCO）权衡 在探讨构建（Build）与购买（Buy）的决策时，企业首先面临的是资本性支出（CapEx）与运营性支出（OpEx）之间的微妙平衡。购买商业方案为企业带来了一个显著的财务优势：能够以 OpEx 的形式分摊成本。对于技术高管而言，这远比每隔五年向董事会和首席财务官（CFO）“摇尾乞怜”申请一笔巨额预算去购买硬件要容易得多。然而，AI 基础设施的复杂性导致总体拥有成本（TCO）极难精准预测，整个行业目前几乎都处于摸索阶段。 主持人将话题从纯粹的 AI 运营拉回基础设施层面，指出当前企业的采购流程（Procurement）已经发生了根本性重塑。过去在评估供应商时，采购部门最关心的是服务等级协议（SLA）等常规指标；而如今的“新常态”则更加聚焦于如何避免供应商锁定（Vendor Lock-in）以及如何动态管理基础设施成本。面对不断迭代的模型和剧烈波动的价格，企业需要全新的评估框架。 ```mermaid flowchart TD accTitle: AI基础设施采购考量变迁 subgraph Traditional["传统采购模式"] A["固定资本支出 CapEx<br/>5年一次硬件大采购"] B["重点考核指标:<br/>SLA 服务等级协议"] end subgraph Modern["AI 时代新型采购"] C["灵活运营支出 OpEx<br/>按需与按Token付费"] D["管理层级可移植性<br/>防止供应商锁定"] E["数据主权与合规审核<br/>PHI/IP/GXP 风险控制"] end Traditional --> Modern ``` --- ### [025:08] 可移植性架构与数据主权：采购部门的新“硬性红线” Mike Burgio 指出，现在的采购部门提出了一系列全新的“非谈判条件”（non-negotiables）。随着底层大模型日新月异、推理成本持续变化，采购人员最核心的诉求变成了：“随着成本与模型的演进，我们迁移系统的难度到底有多大？架构是否具有充分的可移植性（Portability）？”为了不被单一供应商牢牢绑定，企业必须精心挑选合适的中间管理层，构建可灵活替换底层模型的解耦架构。 此外，**数据主权（Data Sovereignty）**已成为采购与合规审查中的重中之重。Mike 分享了一个生动的案例：前一天晚上他与一位来自澳大利亚医疗健康领域的同行交流，对方正试图完全在内部闭环构建一切系统，因为当地法规严禁任何受保护健康信息（PHI）流出本地基础设施。但这种做法也带来了代价——他们无法充分利用顶尖通用大模型（LLM）的强大能力。因此，最理想的解决方案依然是“混合架构”（Hybrid Approach），在保障合规的同时保留按需切换不同模型的能力，毕竟 AI 模型从来都不是“一刀切”（One-size-fits-all）的。 - **采购关注焦点的迁移**：从过去的合同 SLA 转向架构层面的管理层可移植性与解耦能力。 - **数据合规与大模型能力的矛盾**：医疗健康等高合规要求行业（受 PHI 约束）在追求数据不出域的同时，面临无法使用外部超大规模模型的权衡。 - **混合部署的价值**：建立能够随着成本与业务需求变化、随时插拔和替换模型的管理平台。 --- ### [027:36] 跨职能风险矩阵与底层硬件供应链的博弈 Devang Modi 进一步拓展了采购的定义，强调如今做出一项“购买”决策绝非单纯的商务行为，而是必须直面庞大的企业跨职能风险矩阵： 1. **IT 安全与数据主权风险**：数据是否会被滥用或违规出境； 2. **知识产权（IP）风险**：AI 生成内容及输入数据是否涉及侵权或产权流失； 3. **财务控制与 Token 成本风险**：如何实时监控并遏制 Token 调用的潜在账单爆炸； 4. **合规与监管风险**：例如生命科学领域的 GxP 规范，监管机构可能根本不认可某些第三方黑盒方案。 面对充满变数的初创生态，Devang 分享了他们应对供应商锁定的实战策略：如果对某家供应商不够放心，他们会在合同中明确要求设立“60 到 90 天的无痛退出期（Out Clause）”；只有当供应商证明了自身实力与稳定性后，才会签署 2 到 3 年的长期锁定期限。 ```mermaid flowchart LR accTitle: 企业AI采购的多维风险评估矩阵 subgraph EnterpriseRisks["企业采购综合风险矩阵"] direction TB R1["数据主权 & 隐私风险<br/>PHI 出境限制"] R2["知识产权 IP 风险<br/>训练数据与产出归属"] R3["财务与 Token 成本控制<br/>防止调用量失控"] R4["合规与监管风险<br/>如医药行业 GxP 标准"] end subgraph ContractStrategy["合同谈判与退出机制"] direction TB C1["试水期: 60-90 天退出条款"] C2["成熟期: 2-3 年长期锁定合同"] end EnterpriseRisks --> ContractStrategy ``` 从底层硬件视角来看，Ryan Horjus 指出基础设施采购的核心永远是“成本与交期（Lead Time）”。目前业界正承受着某些头部算力芯片厂商（即被称为“绿厂”的 NVIDIA）极其漫长的供货周期。为此，Hydra Host 正在打造开源的中立管理平台，旨在让企业能够无视服务器 OEM 厂商的差异直接掌控裸金属（Bare Metal）算力。当特定交换机需要等待 20 周而替代品仅需 4 天时，这种中立性能够让企业迅速换用不同供应商的硬件，极大地缩短算力基础设施上线的周期。 --- ### [030:29] 项目中途推倒重来：上下文耗尽与“买转建”的认知沉淀 当演示环境（Demo）中光鲜亮丽的供应商平台接入真实的业务系统后，推倒重来的情况屡见不鲜。David Chun 分享了一起典型的“买转建”中途转向案例。由于当前 AI Agent 的采购往往由业务部门（BU）直接发起而非技术高管主导，新泽西一所拥有 40,000 名学生的大型大学部署了一套直接调用前沿商业闭源模型的学生服务 Agent，覆盖了 15,000 名学生。第一年系统运行平稳，团队通过 RAG 进行微调；但到了第二年，系统开始暴露出严重的认知崩溃： > “由于我们不加甄别地把所有数据都塞给模型训练，加上 AI 自身生成的合成内容，它彻底耗尽了上下文窗口（Context Window）。它开始遗忘最初精心设定的关键内容，甚至在某种程度上发生了退化……它缺乏判断哪些信息重要、哪些不重要的执行决策能力。” 加之闭源模型厂商频繁升级底层版本（如每周推送 4.2、7.9 等新版本），导致输出结果难以预测。最终，校方决定“冻结”模型行为，转而采用相对稳定、成本更低的开源开放权重（Open-weight）模型并自建基础设施。 Devang 对此深有同感。他透露大约两年前他们采购了一套顶尖的供应商方案，但在实施一年半后，由于开源生态和内部工具的飞速发展，连市场营销部门自己都能低成本搞定该功能，于是团队果断叫停了项目。Devang 强调，这种“半途而废”绝不可惜： > “这并不是什么令人遗憾的事，你应该大胆去做。你从整个过程中汲取了大量知识与经验……在选择供应商时，你需要看重的是能否借其研发能力加速你的交付和市场验证，未来你随时可以根据形势进行切换。” --- ### [033:53] 供应商视角下的“退出机制”：坦诚沟通与代码所有权 面对客户在合作中途可能出现的“买错了”或战略转向，Mike Burgio 开玩笑称“从未听任何客户说过买我们平台是错误的决定”，引发了全场笑声。随后他认真阐述了成熟供应商在应对此类顾虑时的专业态度。 Mike 表示，OutSystems 在与客户建立合作的初期（“就像在第一次或第二次约会时”），就会主动把最敏感的退出机制摆上台面。尽管这对于销售方而言并不容易启齿，但他们会明确告知客户：如果在平台上构建了 Agent 应用，平台具备直接导出完整底层代码的能力。“这就像在说：‘我知道我们才刚开始约会，但如果我们将来离婚了，你的所有东西都可以带走。’”这种保障客户不被锁定、随时保留成果资产的机制，不仅减轻了企业在采购初期的决策心理负担，也成为了现代企业级 AI 平台建立长期信任的关键基石。 ### [35:12](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=2112s) 项目范围界定困境与底层基础设施的可移植性 许多企业项目之所以走向失败或进展不顺，根本原因往往在于前期的范围界定（scoping）存在缺陷，团队未能充分理解项目的全貌以及确保成功所需的关键能力，导致最终不得不中途叫停，转而自行进行定制化构建。在基础设施层面，情况则呈现出高度的灵活性。对于像 Hydra Host 这样的底层基础设施提供商而言，无论客户选择自建还是购买，本质上都只是一个资本决策问题。不少初创公司在测算基础设施的使用周期与财务账目后，会突然意识到自购硬件或许更为划算。 开源工具在此类场景中展现了极大的价值。企业一旦承诺采用开源工具链，便能获得极强的可移植性（portability），无论选择哪条路径都能平滑迁移。尽管底层服务器在不同数据中心之间的物理搬迁仍然极具挑战——因为各个机房的合同归属不同，且电力容量（兆瓦，Megawatts）极度稀缺、各方都在争抢——但在明确“究竟由谁拥有基础设施资产”的维度上，开源和中立的架构能为企业保留完全的自主选择权。 ```mermaid flowchart TD accTitle: 基础设施决策维度与开源可移植性 subgraph Capital["资本与架构决策"] A["评估基础设施生命周期与 ROI"] --> B{"资本路径选择"} B -->|"自建 / 自购硬件"| C["资产归属企业"] B -->|"采购云 / 托管服务"| D["服务订阅模式"] end subgraph Portability["开源工具链赋能"] E["开源工具与中立架构"] -.->|"实现高可移植性"| C E -.->|"无缝对接与切换"| D end subgraph Physical["物理层制约因素"] F["数据中心合同归属"] G["稀缺的电力容量 (Megawatts)"] F --- G end C -.-> Physical D -.-> Physical ``` --- ### [36:34](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=2194s) AI 如何重塑系统迁移并颠覆知识工作模式 探讨从传统 IT 时代到 AI 时代基础设施切换的难易程度时，情况发生了显著分化。从软件与应用迁移的角度来看，AI 工具实际上大幅降低了迁移门槛。如今借助 AI 工具，企业迁移像 COBOL、Lotus Notes 这样的古老遗留系统，所花费的时间仅仅是十到二十年前的一小部分。AI 带来了前所未有的代码理解力与洞察力，使企业能够在需要调转技术方向时迅速执行，极大地压缩了替换老旧系统的时间成本。 与技术供应商视角不同，企业应用端则观察到了更深层次的技术锁定与生产力重构趋势。当前大语言模型（LLM）的发展水平已经能够胜任 70% 到 80% 的日常知识工作者任务（例如基础调研与文档处理等，高精尖的物理学等前沿探索除外）。这意味着企业可以开始对特定层面的工作流与技术栈进行“固化与锁定”（locking stuff）。企业甚至不再强制依赖公有云，在合规允许的前提下完全可以将这些高频、成熟的模型与任务直接运行在本地私有环境（on-prem），从而在摆脱对外部云厂商过度依赖的同时，牢牢掌控核心数据与生产力资产。 --- ### [38:40](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=2320s) 底层物理现实的回归与技术消费者的黄金时代 在基础设施的最底层，技术演进反而呈现出一种“重返传统工种（return to old school）”的奇特景象。无论上层的 LLM 模型多么先进，底层设施的落地依然高度依赖管道工、线缆工、电力技师等实体技术工种，必须依靠现场工人铺设管线、安装配电线束（power whips）和放置变压器。由于算力建设资金海量涌入，这些蓝领技术工种正面临极其严重的供不应求，企业必须支付高昂的溢价来争夺稀缺劳动力以加速交付。因此，底层基础设施建设的核心诉求是构建一套完全与上层模型无关的“中立框架”（agnostic framework），无论行业未来跑什么模型，底座都能稳定承载。 对于技术买方和消费者而言，当下正处于创新的黄金时代： - **开源与闭源竞争白热化**：开源社区与专有商业模型每天都在以极高速度迭代，为抢占市场先机展开激烈竞争。 - **供应商理念升级**：现代基础设施供应商更加理解企业对“数据主权（data portability and ownership）”的核心诉求。 - **多方协作生态**：企业拥有极其丰富的合作伙伴生态，能够协助自身在复杂的技术路线中找到最佳落地路径。 --- ### [40:49](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=2449s) 专家核心建议：从“非此即彼”转向“构建与采购并存” 针对企业在 AI 基础设施决策中的核心考量，各位专家给出了各自最关键的一条建议： - **David Chun**：必须从一个扎实的业务用例（use case）切入。精准定义预期的商业价值，并优先去攻克那些“当前业务无法解决的最棘手痛点”，以此作为切入 AI 的支点。 - **Ryan Horjus**：从资本结构与财务预期出发做决策。评估自身最适合在技术栈的哪一层生存以榨取最大价值，并积极寻求高水平的咨询与供应商伙伴协作。 - **Mike Burgio**：构建良好的混合方案（hybrid solution）。采购一个具备高扩展性且不会造成厂商锁定的基础平台，在平台之上开展自建。鉴于当前 AI 技术格局“按日迭代”，必须在当下打牢架构地基，锁定可预测的成本结构。 - **Devang Modi**：决策颗粒度必须下沉至“层级（layer level）”，而非对整套技术栈一刀切。评估每个具体技术分层究竟该买还是该建，将其视为拼图中的独立模块。 **总结**：当前企业的 AI 战略绝非简单的“自建 vs. 采购”（Build vs. Buy），而是深度的“自建与采购并存”（Build AND Buy）。企业必须在明晰业务需求、做好功课的前提下，在不同技术层级上灵活组合采购与自建策略，以应对瞬息万变的 AI 时代。 ## Notable quotes > [05:09](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=309s) 将 AI 置于核心，彻底重塑端到端的工作流 — Devang Modi > [10:04](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=604s) 如果你的 AI 供应商明天把价格翻倍，你能在 90 天内把应用无缝迁移到另一套系统吗？ — Mike Burgio > [18:13](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1093s) 你绝不会希望在遥远的未来被困在别人的模型里。 — David Chun > [18:13](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1093s) 掌控自己的命运（Control over your destiny） — David Chun > [25:08](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1508s) 随着成本与模型的演进，我们迁移系统的难度到底有多大？架构是否具有充分的可移植性？ — Mike Burgio > [30:29](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1829s) 由于我们不加甄别地把所有数据都塞给模型训练，加上 AI 自身生成的合成内容，它彻底耗尽了上下文窗口（Context Window）。它开始遗忘最初精心设定的关键内容，甚至在某种程度上发生了退化……它缺乏判断哪些信息重要、哪些不重要的执行决策能力。 — David Chun > [30:29](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=1829s) 这并不是什么令人遗憾的事，你应该大胆去做。你从整个过程中汲取了大量知识与经验……在选择供应商时，你需要看重的是能否借其研发能力加速你的交付和市场验证，未来你随时可以根据形势进行切换。 — Devang Modi > [33:53](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=2033s) 这就像在说：‘我知道我们才刚开始约会，但如果我们将来离婚了，你的所有东西都可以带走。’ — Mike Burgio > [40:49](https://www.youtube.com/watch?v=Z3lTpVMhAOM&t=2449s) 自建与采购并存（Build AND Buy） — Devang Modi ## Mentioned Andrea Spielvogel, Ryan Horjus, Mike Burgio, Devang Modi, David Chun, Nürtingen-Geislingen University (NGU), University Hospital Tübingen, Hydra Host, OutSystems, Bausch & Lomb, Columbia University, Harvard University, UC Berkeley, OpenAI, Anthropic (Claude), Meta, NVIDIA, Kubernetes, Salesforce, COBOL, Lotus Notes, Gemini ## Video description David Chun - Researcher, Columbia University Mike Burgio - Sr. Manager, Solution Architecture, OutSystems Ryan Horjus - Vp of Deployments, Hydra Host Devang Modi - Chief AI Officer, Bausch & Lomb Andrea Spielvogel - Professor, Nürtingen-Geislingen University (NGU) Every enterprise is grappling with the same question: should we build AI capabilities ourselves, or partner with vendors? This panel will unpack the strategic trade-offs CIOs and CAIOs face when making AI infrastructure bets. Panelists will compare open-source frameworks, cloud-native platforms, and proprietary technology solutions—highlighting differences in cost, flexibility, speed to market, and risk. They’ll also share approaches to avoiding vendor lock-in, maintaining security, and ensuring long-term scalability. Attendees will leave with practical decision frameworks to guide their own build vs. buy strategies in the AI era. [Source](https://www.youtube.com/watch?v=Z3lTpVMhAOM)

[5:09] Devang Modi: 将 AI 置于核心，彻底重塑端到端的工作流 [10:04] Mike Burgio: 如果你的 AI 供应商明天把价格翻倍，你能在 90 天内把应用无缝迁移到另一套系统吗？ [18:13] David Chun: 你绝不会希望在遥远的未来被困在别人的模型里。 [18:13] David Chun: 掌控自己的命运（Control over your destiny） [25:08] Mike Burgio: 随着成本与模型的演进，我们迁移系统的难度到底有多大？架构是否具有充分的可移植性？ [30:29] David Chun: 由于我们不加甄别地把所有数据都塞给模型训练，加上 AI 自身生成的合成内容，它彻底耗尽了上下文窗口（Context Window）。它开始遗忘最初精心设定的关键内容，甚至在某种程度上发生了退化……它缺乏判断哪些信息重要、哪些不重要的执行决策能力。 [30:29] Devang Modi: 这并不是什么令人遗憾的事，你应该大胆去做。你从整个过程中汲取了大量知识与经验……在选择供应商时，你需要看重的是能否借其研发能力加速你的交付和市场验证，未来你随时可以根据形势进行切换。 [33:53] Mike Burgio: 这就像在说：‘我知道我们才刚开始约会，但如果我们将来离婚了，你的所有东西都可以带走。’ [40:49] Devang Modi: 自建与采购并存（Build AND Buy）

[20:08] they refuse to be trapped into relying on someone else's model in the future—they want total control over their own destiny. [23:22] David Chun: current token prices are artificially low because venture capital firms, tech giants, and government grants heavily subsidize them. [23:22] David Chun: When an enterprise builds its own infrastructure in-house, it suddenly absorbs the real, un-subsidized costs—such as hardware procurement and massive electricity bills. [39:15] Because massive financial investments are at stake, organizations are willing to pay significant premiums to secure this specialized workforce and accelerate timeline execution. [39:15] Regardless of which large language model (LLM) wins out in the market, having an infrastructure-agnostic framework is crucial so enterprises can adapt no matter how the underlying technology evolves. [40:49] David Chun: Focus on identifying a compelling use case with clearly defined ROI. Enterprise leaders should target the hardest, previously unsolvable problems to maximize the impact of AI adoption. [40:49] Ryan Horjus: Determine where your organization needs to sit from a capital deployment perspective. [40:49] Mike Burgio: Adopt a flexible hybrid strategy by buying a platform you can build upon without getting locked in. [40:49] Mike Burgio: Given that the AI landscape shifts on a daily basis, establishing a predictable cost structure and flexible foundation must precede specific tool implementations. [40:49] Devang Modi: Evaluate the "build vs. buy" question at individual architectural layers rather than attempting to make a single binary choice across the entire technology stack.

[0:04] So hi everybody um thanks for being here. So uh  brief introduction from on my side I'm Andreas [0:11] Spilfog um so professor for applied AI in Germany  uh and uh my research area is uh user innovation [0:20] uh with AI and the all the upcoming uh topics  around uh governance risk and compliance. So I [0:27] know this is something nobody would like to hear  but um it's really interested to do research on. [0:33] So after decades um of um working in the industry  so I last year moved to academia and focusing now [0:43] more on um AI projects with have more purpose. Um  so my actual uh project is one with the university [0:52] hospital in Tubingan. So uh we are right now uh  do research on how um AI could help people with [1:01] brain cancer uh to do on uh just improve the  shared decision- making and also to improve [1:09] um how to um improve the um relationship between  them and their um doctors. So um I was on a little [1:22] AI journey the last week. So uh first I was in  uh Harvard just presenting then my uh there my uh [1:29] research according to user innovation and all the  upcoming blind spots according to uh governance [1:36] risk and compliance with AI. Uh and so we are uh  just talking on a conference called open and user [1:43] innovation. um uh the the topics around um how AI  affects um innovation and uh the user um the human [1:54] AI um interaction. So then I headed to Berkeley  uh on the weekend. So there was the agendic um AI [2:03] summit and that was all about agendic um how the  infrastructure need we need to have the research [2:10] on and also how we could develop um agents in  a sustainable and ethical way and now I'm here [2:18] um hosting that um awesome uh panel and I really  would like to um hear what uh you all think [2:28] um about the the the topics according to um build  and buy from the AI infrastructure. So that was [2:36] one thing from my side. So I know we are right now  facing you know the the question and how could we [2:42] just uh at AI and also AI because that's the big  topic within the environment um of the companies [2:53] um where should we take control um and also  uh where should we just make uh or build or [3:00] buy according to the infrastructure so that was  I know a long introduction from my side um but I [3:07] would like to uh overhand here to Ryan. Uh maybe  you just also introduce yourself a little bit [3:13] what you're doing. Yeah. So Ryan Horge with Hydro  Host. I lead our AI infrastructure deployments. [3:21] So right at the intersection between supply  and demand or what you could say capex versus [3:28] opex in a build versus buy scenario. So we work  with a number of different companies that some [3:32] own the infrastructure, some need to rent that  infrastructure and depending on their situations, [3:38] they might make one go one route or another. Hi um  I'm Mike Burgio. So I lead a team of AI solutions [3:47] consultants for a company called Outsystems.  Um Outsystems for those that don't know, [3:52] so we're an Aentic systems platform. um we  basically give you the capability to build, [3:57] deploy, manage, govern enterprise applications and  agents um across the enterprise stack. So nice to [4:04] be here. Hi, Dang Buri. I'm chief AI officer at  Bosch Alam. Um you know, I focus on applied AI at [4:12] the at the organization. I'm the buy side of this  uh buy or buy bill side of the decision- making [4:18] a lot. Um very excited to hear this uh discussion  today. I'll have different perspective more on the [4:25] buy side what I'm thinking how I'm thinking and  how frequently I change decisions also so it'll [4:30] be very interesting here. Sure. So um we miss  somebody so maybe we're just you know we are in a [4:38] improvised section but I think we can all do this  because we have really a lovely discussion before [4:45] uh and also we don't want to uh do that  questionnaires from my side because I recognized [4:52] uh you know listening these two these three it  was so awesome so I hope we we just um uh do that [5:00] conversation uh we had for preparation again.  So um maybe uh even just according to so let's [5:09] start with a real um example. So I will just ask  you uh uh could you just walk us through a recent [5:17] AI deployment um and ongoing infrastructure um  project on your side. So yeah um so yeah al to [5:26] be careful what I share so you know some of the  confidential information but you're looking for [5:30] something where you put AI in the center redesign  the workflow that's how we think about everything [5:35] right so not augment and put AI in the workflow  somewhere in between that that [clears throat] was [5:41] initial stages of the work that all of us did um  our recent project I would say is we are looking [5:47] at end to end source to procure for example  and looking at payment up to payment process [5:53] Right. And that's redesigning end to end process  using agent. Right? That's that's the example. Lot [5:58] of decisions to be made. It's still in progress.  A lot of decisions to be made. What I would say is [6:02] there's a lot of buy and build decision, right?  Like when we were dealing about 10 years or 15 [6:07] years ago, we were deciding do we want to build  CRM or buy Salesforce or something like that. [6:13] That was easier decision. I think the decision  here now is actually what layer you want to buy [6:19] and what layer you want to build. Do you really  want like companies like us, do you really want [6:24] to build LLM? Most of us, we don't want to do it.  We should, you know, buy it or rent it. That's [6:29] easier. Infrastructure side, you want to buy it,  build it. That's how you should be thinking it. [6:34] And from there onwards, you have orchestration  layer. Then you have memory, you have context, [6:38] now you have data. That's where your edges. That's  what you have what you have in your company, [6:43] right? That's that's what you want to own. Um,  so that's the decision you want to make. And [6:47] I'll add one more thing and then then continue  the conversation here is I think at the end of [6:53] the day for us we have resource and time to build  stuff and R&D it right and there are people who [7:02] will be doing it you see tons of vendors who does  spend a lot of time building certain things it's [7:07] going to be difficult can we build it absolutely  we can build it but do we have time and ROI that [7:12] we're going to get it to build it so that's the  decision you have to make but there are areas [7:16] where we think we should just build it in example  of source to pay for example that we talking about [7:21] we're going to end up building because it makes  sense it will create a board for us yeah and in [7:25] the end it's also a thing how to maintain that and  do do you have the people for it and um whatever [7:32] so um I think so the main concern so when I was  just uh working as a um um chief digital officer [7:39] so the main concern was uh when we just um migrate  to another system um So um according to we bought [7:51] it right and it was just like a thing about uh  vendor lockin. So this is one thing which uh most [7:57] of us are I would say afraid of. Um I know there  are two peoples they will say no no no that's fine [8:04] but um so because the main concerns are oh if we  just buy uh infrastructure from from um a vendor [8:13] so are we still stuck with them to the end of our  the company um uh whatever. Yeah. Yeah. So Ryan, [8:26] what do you think? So um how do you make sure that  customers stay flexible and in control uh when [8:35] they choose you? Yeah, I think what we saw in sort  of the boom of AI was we saw a shift back towards [8:44] the data center. It was um you had a lot of  workloads migrating into the cloud where you were [8:51] op you know other companies were managing that  infrastructure but GPU infrastructure was so much [8:57] different so much more expensive and it became a  more unique position where uh some companies could [9:05] own it or pay way more money than uh to the clouds  rent it. It really just depends on the position of [9:13] your company and maybe your consumption rates.  But what we saw was the workloads were much much [9:19] more agnostic. You anybody I shouldn't say anybody  but loading a model the model is kind of the key [9:26] thing that most things are built around. So uh  how you manage that infrastructure there's your [9:32] data that is on that infrastructure becomes a much  deeper question around uh how companies want to uh [9:40] control that information. Uh, and I think that's  where you can make that decision on a build versus [9:48] buy basis is can you use that infrastructure  for one use case and change out different [9:56] uh, workflows on that equipment via, you know,  driver updates or things like this? Exactly. So, [10:04] Mike, any any thoughts on that? Yeah, I mean I  I'll be the first one as a vendor to tell you [10:11] you should be concerned with vendor locking,  right? So, uh it's it's okay. Um and you know, [10:18] with that there's some cons considerations that  you need to take with that, right? There's some [10:22] flexibility, there's some portability, things  of that nature that you have to think about. [10:27] Um it was it was interesting. And I was reading an  article the other day and uh the the headline kind [10:33] of caught my eye and it was, you know, essentially  if your AI vendor doubled their prices, you know, [10:39] tomorrow. Um how quickly can you port your  applications over to to another system? Um can [10:45] anyone in here probably do it in under 90 days?  Probably not, right? So um there are many models [10:53] out there now that offer you this hybrid solution  where you can build your core infrastructure on [10:59] a platform that will allow you to you know pick  and choose either different models or different [11:07] vendors that you connect to different data sources  so on and so forth. Um so don't get caught up [11:13] totally by well if I buy it I'm locked in because  that's not always the case. A lot of the vendors, [11:21] ours included, will allow portability,  but also give you the flexibility to make [11:28] um changes to anything that you need to, whether  that be an LLM model or disparate data sources, [11:35] whatever that may be. So, just be cautious when  you're looking at the build part of it or the [11:41] buy part of it, my bad. um and make sure that  you're looking at a solution that you can easily [11:48] um move off of if for some reason that vendor  did decide that they were going to double their [11:53] prices tomorrow. I'm just going to add a couple  things. So I have my procurement head with me [11:59] here as well, right? So I have procurement inside.  Keep it with your side. I'm just more practical so [12:04] you guys know it. So if there is a vendor lock  in, you should understand what that means and [12:09] what what it means for you. understand the company  you're buying from, right? So, you can have a lot [12:14] of new AI companies that does not have backing or  they will potentially get acquired. So, you want [12:18] to understand what that looks like. That that is  a key when you make a decision. That's tough. If [12:24] they don't make investment or if they don't have  investment, you're going to lose out. So, those [12:28] are the things that you want to look out for. And  the build, again, I'm going to go back to again, [12:32] it's all about builder buy is really around the  layers. So I see no reason not to sign a contract [12:37] with any of the frontier models if that makes  sense for you or infrastructure companies. But [12:42] if you're thinking about your data memory layers,  that's where you want the most flexibility and you [12:48] want to own it and you want to be able to switch  it. You can switch to any model as as as we talked [12:54] about here. But if your data is stuck with quad,  let's say, you won't be able to move. I can tell [12:59] you today. So that's the challenge. So that's what  you to focus on and those are the questions you [13:03] want to ask. Yeah, I think like you're saying it's  naturally where you live in the stack. Obviously, [13:09] where we're at Hydro Host, it's bare metal. So,  but if you need to own higher parts of the stack, [13:16] then um you might buy bare metal. But if you  don't want to design, you know, the inference [13:23] layer or the token layer, you're going to be  buying tokens, right? So, that really has to [13:27] make a determination like what skill sets you have  in the company to manage that infrastructure. And [13:32] obviously we partner with people that need sort of  some management of the infrastructure layer. They [13:39] don't have like all of those engineers but make  it make it a little bit easier. But I mean we even [13:44] see a big move to like Kubernetes on bare metal.  So you're eliminating a virtualized layer in there [13:49] as well that also helps to eliminate lock in. Yep.  Perfect. So um also welcome right now David. So [13:57] maybe you introduce yourself. Um I apologize I was  in the wrong uh wrong room and while we get here [14:04] David Chun I'm a a researcher primarily in the  areas of CIO management uh at Columbia University. [14:12] So and originally your first question uh was yours  but we we already sorry for that. Yeah but you [14:20] can jump in. Okay. Yeah. So like at Colombia  we're very interested in responsible AI. So [14:25] uh we don't buy shelf off off offtheshelf product  and we do invest in large montial models like open [14:31] AAI and entropic and everybody has that but we're  very careful about how we select the tools and the [14:37] automation that it creates. So a lot of our work  is making sure we vet the responsibility component [14:42] and the transparency component. We're very big on  AI engines that are um explanatory that gives you [14:48] good explanation of it logics that it's auditable  in some way. Oh, thanks. Um, so, um, coming back [14:55] to our questions. Sorry for that. [laughter]  So, um, Mike, we we often see the tension [15:02] between flexibility and speed. Um, customuilt,  uh, building offers, uh, incredible flexibility, [15:09] sometimes a little bit too much. It was my  experience on that. Um, while leveraging uh, [15:15] established platform can reduce time to market  a lot. So, what is your recommendation on that? [15:22] I have a lot of thoughts. Um, look, I think that  the flexibility versus speed debate, right? It's [15:30] not just about deployment time anymore. Um, I  think everybody can rapidly deploy, you know, [15:36] prototypes, first iterations of applications  without any issues these days. So now the [15:42] conversation kind of switches to risk management.  Um, and you know, when you're doing a custom build [15:49] of something, you also have to custom build the  security, the guard rails, and everything else [15:54] around it. Um whereas if you buy a platform  that at least gives you that capability and [16:00] functionality then the build on top of what you've  already bought right will give you kind of some of [16:07] that flexibility to deploy things faster because  your guard rails are now built across your entire [16:14] um enterprise stack. Okay. So, um, you know, I  think it's really more of a conversation and I' [16:21] I'd love to hear from these guys a little bit that  are, you know, more in the office every day, but, [16:26] uh, you know, how how do you kind of weigh that,  um, now that speed's not really the the big [16:33] differentiator? It is more about minimizing your  risk. So, how do you go kind of go about that and [16:39] make sure that you're buying a solution that'll  kind of give you that governance across the stack, [16:45] but still give your teams kind of the flexibility  to custom build things? Totally agree. I think [16:50] I'm going to step back again. I'm a developer  at heart. I still code. Um, I can build agents [16:55] nowadays, right? And it's easy for everyone to  do it at some time. When I look at that mind, [17:00] it's like, I'll just build it. This is easy. Like,  just do it. It's easy. But ROI and what what what [17:06] you're describing Mike is really difficult to  look at end to end stack and saying how do you [17:11] put governance how do you make it consistent when  you have 13,000 employees building agents how do [17:16] you make that work that's where the critical  part is yeah one-off applications or agents [17:20] it's okay to build build it out that's fine but  if you're really looking for something that goes [17:25] to production wherein you have control you need  to have stoppage you need to make sure there are [17:29] human in the loop your IP is protected your PHI is  protected you're looking at all that it's not easy [17:34] your security like the passwords and everything  that can go into this very critical. So, [17:40] so that's why you need a base layer critically set  up as a buy make sure it's connected in your uh in [17:46] infrastructure and secured and it has governance  layer and then on top allow autonomy we want to [17:52] move fast. I'm stepping back again as a company  manufacturing company or companies like that it's [17:57] going to be difficult for us to hire bunch of FDES  or a bunch of developers who are just doing this [18:03] for for the living right we are not that's not the  purpose of the company want to get the product out [18:08] to our patients that's what the focus is so that  that's where we that's the decision needs to be [18:13] made so David any thoughts on that yeah so you  know building by I think we talked about this [18:19] before and we had like a very big relevance when  we were engineers right? Because you had the best [18:24] knowledge, you had all the data, you had all the  skills, you hired a good folk from a good school. [18:29] You like Columbia, right? And um and then you  can build great things in house. It's just not [18:34] possible now. AI technology is global, right? You  got open new open weight models that are coming [18:39] out almost every day that's outperforming some of  the frontier models. You have so many options. Um [18:46] you really have to kind of do a hybrid approach  and this is not my word. I think it's one of my [18:49] colleagues who mentioned that like your strategy  has to be build be able to pivot. Um we were also [18:54] talking about tokens right and as well as based  on cost before it was just easier and cheaper to [18:58] build things on house and control every element  and allow your users to work on it. Right now [19:04] that's really haven't been the case ever since  Apple iPhone you know consumers and users drive [19:11] some of the technology that you'll utilize. So the  enterprise kind of have to play catch-up. And if [19:17] you're Gemini has more freedom at your desktop  at home computer to do what they need to do in [19:23] the agent side. Um you know what what does large  complex organizations with all the lawyers and all [19:29] the compliance rules and all the things and how  fast can we really move and how fast can we build? [19:35] Because if we build something we also have to test  it and it's expensive, right? compliance is it's [19:41] it's a big part of our thought process, right?  Are we are we letting everybody know? And um [19:47] and we also talked about being black boxes, right?  When we buy these, it's a black box. I don't even [19:51] know what I'm buying, right? I know it's really  good. It's free um and we can use it in house, [19:56] but uh it just has to be a good combination of u  building things that you're still really good at, [20:02] keeping things protected like your data, and then  buying things that you really shouldn't have to, [20:08] right? Um uh I I don't want to go a long way.  I was just having a conversation with the Meta [20:12] folks and um they're obsessed with building the  A team like the old school days, right? They have [20:18] the most smartest a million billions of dollars  spent in building their own foundation model. And [20:23] I asked them like you're meta, why do you why do  you need to build foundation models, right? You're [20:28] like a advertising juggernaut. And um and he told  me I think which makes sense all to all of us is [20:34] that you don't want to be uh um trapped into using  somebody else's model in the way future. You want [20:42] to have some level of control over your destiny  and that's where the big balance is on building [20:46] by. But I'm an engineer. I love to build stuff.  Don't get me wrong. So just let's jump in because [20:51] we have a discussion over um the hidden costs of  AI infrastructure and also of AI operations. So, [20:59] uh, Ryan, uh, you mentioned, uh, uh, cool topics  about I I don't get it everything what you, [21:06] um, what you explained, but, um, it's it's worth  that we just go into that, uh, discussion because [21:13] I think it's also, uh, one thing we need to be  aware of, um, according to uh, builder buy. Yeah, [21:21] I think it was something around like where you  consume your tokens, right? Obviously at one one [21:26] scale you have you're buying it from claude or  open AAI directly. Um you're just consuming what [21:33] they give you. Then you've got this middle  section which is maybe you partner with an [21:37] inference provider and you have a bit more custom  control over that infrastructure but in the end [21:43] you're still buying tokens. those tokens might be  cheaper than buying it like you know retail value [21:48] tokens from you know the big the big juggernauts  but then you've also got the other phase where [21:56] okay well now you own the data center contract you  own the power and you can own the infrastructure [22:02] and you're just paying capex and you got a very  kind of clear maybe small opex component around [22:08] the rental of the data center but you've own  the equipment and maybe you consume those tokens [22:14] um with your own stack and you own the pipeline  and the data and there's a lot of great, you know, [22:21] software platforms you can plug onto that that can  be self-hosted, right? But now that infrastructure [22:28] component you've kind of brought inhouse and  maybe that's based on your own company's, [22:33] you know, status or like how profitable or if  you're a startup and your burn rate is high [22:39] um and you need a pure opex model. So that's where  we see kind of the dichotomy is like picking in [22:45] that window and you know like otherwise you know  you end up with these massive cloud bills or token [22:53] bills and you got to control that. That's like the  new thing or or just to add to it is security or [22:59] the 24/7 operations in manufacturing. You need  edge model running in the manufacturing so we [23:04] don't shut down. So that's different needs but  that that you have to consider right. So that's [23:09] where you're going to buy and build actually in  that one actually you're not renting it. We are [23:13] building it and putting in house but for a regular  use cases fine right you just rent it. Yeah, [23:22] I think we also talked about this whole cost  of tokens, right? I mean, does everybody really [23:26] understand what a token is, right? We kind of know  what what it costs, but then we realize all these [23:31] tokens are being heavily subsidized, right? We we  we get allocations, but most of that is subsidized [23:37] by the uh by the venture capitalist firms and  large companies and sometimes the government. So, [23:42] um and and you incur those costs when you build  inside, right? You pay the electricity bills, [23:46] you buy the hardware. So you're you're paying  the different things and um and you're trying [23:51] to really balance the total cost of ownership,  right? And I think that what you know buying [23:58] allows you to do is um you know do more opex  expenses which for my bosses and the CFOs love [24:06] um instead of me asking for like a teen amount of  money every 5 years and begging the board that we [24:11] definitely need to do that. So it's a complex  problem, right? um we we just don't know a good [24:16] fix. Everything's up in the bear and um you know  we're worried about that. Yeah. But coming back [24:22] to the infrastructure discussion uh not only into  the inter uh the AI uh operations because I think [24:29] we can spend uh I don't know we will sit here  uh until Christmas to talk about that. Um so um [24:37] according to uh what when is coming for a company  to to the decision from uh build and buy uh so [24:46] also procurement has completely changed right so  we talked about that and um so um what is right [24:53] now the new normal let's say it like this um when  uh you went into uh window evaluation and when it [25:01] comes to avoiding lock in and uh especially also  the managing the infrastructure cost. So maybe [25:08] um a question to your both sides or Ryan and Mike  um because you are working on the infrastructure [25:15] um side. So um what do you just uh what's  your thought about that? Sure. Yeah. And I [25:23] know we talked a little bit about vendor lock in  before um but you know there are some new kind of [25:29] non-negotiables I'd say you're probably getting  from your procurement department, right? I think [25:33] they used to be more concerned about SLAs's and  things of that nature and now it's like okay [25:39] um how do we pick the right kind of management  layer um again that'll allow us some portability [25:48] and won't get us in stuck into that kind of vendor  lock in right so from a procurement perspective um [25:55] and I wish we had your procurement person up here  but from that perspective I think they're looking [26:00] more at as costs change as models change. Um, how  easily can I move, right? How portable is this? [26:11] Um, and I think that's really the biggest thing  that you're going to see procurement asking. Um, [26:17] you know, maybe the only other thing I would say  is especially when we're talking about AI models, [26:24] um, the last thing that procurement is probably  thinking a lot about is data sovereignty, [26:29] right? So, how do I make sure that I, you know, I  was talking to someone last night from Australia [26:35] and he was talking me through some of his um  challenges that he has being in healthcare and [26:42] we had a really good conversation, you know,  about how they're trying to build everything [26:47] internally because they can't let any, you know,  um, PHI data or anything like that outside of, [26:54] um, you know, their infrastructure. Um, but  at the same time then they're not getting the [27:00] full benefit of what they might be getting from a  larger like LLM. So they're trying to balance that [27:05] a little bit. Um, but the idea scenario again goes  back to the hybrid approach, right? So how do you [27:13] kind of build a platform that allows that kind  of interconnectivity to be able to swap models [27:19] in and out as your cost structure changes and as  the need changes? because these models are not, [27:25] you know, one sizefits-all and those are really  some of the decisions that you have to think [27:30] about before implementing them. Yeah. I'm going  to add a few things, right? So think about going [27:37] back to simple you have same risk that you had was  a little bit more you know before AI and now. So [27:43] when I say procurement, it's a little bit broader  than procurement, right? So your procurement, your [27:47] legal, your compliance, data privacy, IT security  like as a vendor, you have to deal with all of [27:53] that. It's not just one of one of them, right?  So that's changing. But those are really your IT [27:58] security risk when you make a decision on buy you  have data and data sovereign you know that risk, [28:04] data side of the risk, uh privacy risk around  that. You have IP risk. I don't know how many [28:10] think about it but there's a lot of IP risk when  you making decision. Um there is a risk around [28:16] financial control and compliance on if you're on  GXP for us let's say those are risk because you [28:22] know regulatory may not accept what you're doing  right so there's a lot of risk to be assessed and [28:27] really financial control risk we talked about  token and token cost we need to understand [28:31] that also right on top of vendor lock in now  interestingly we went through couple generations [28:36] we got lucky in some where we we made a decision  to buy and they got acquired or these companies [28:41] got acquired by the larger corporation which made  it easy for us but once that did not and we feel [28:48] like we're going to switch it out. So those are  the things that you're going to start seeing and [28:52] most of our contract where we feel not comfortable  we just be upfront with the vendors saying listen [28:57] we are not feeling good about this but we won't  try you out give us 60 90 days out simple we'll [29:03] try it out if it makes sense we'll continue and  there are vendors where we say okay we'll lock in [29:07] because you've done enough um we'll lock in for  two to three years that's fine so and Ryan from [29:12] your perspective um are they do you recognize  any changes how companies uh just reach out [29:20] to you uh according to procurement. Yeah, I'd say  obviously we live a little bit down in the stack, [29:25] but procurement for us is a lot about cost and  lead time and that's whether you're capex or opex [29:31] um in your you know it's basically which financial  model makes most sense for your company but in the [29:36] end it's like well how quickly can I get this  infrastructure online right so part of why we [29:42] built our platform which is open-source uh here  in the next few months is to help eliminate even [29:48] vendor lock in amongst you know server or even  OEMs, right? So that you can control bare metal [29:55] no matter what vendor it is and then utilize it  in a more agnostic way. So that means when we're, [30:02] you know, we're prioritizing speed and cost to  enhance that procurement process so that instead [30:08] of having to wait for this one switch that's  the only one that works, you can now swap in [30:12] multiple different vendors if one's 4 days versus,  you know, 20 weeks, which obviously, you know, [30:18] the the green guys have some very long procurement  times that we're all dealing with right now. But, [30:23] you know, there's other ways to mitigate that  to get infrastructure online much faster. [30:29] So um thanks for that. Uh so coming back to the  customer side. So window platforms look great [30:35] um in a demo but uh the real test is to plug  them into existing texts right uh David. So, [30:44] um, have you ever reversed a build or also  buy a decision vice versa? Um, uh, just, uh, [30:54] midway through a project. Yeah, with a lot. Um,  we talked about the hybrid approaches before, but [31:03] um, nowadays we're also talking about procurement,  so it's kind of relevant. A lot of the AI agent [31:10] procurement process is no longer driven by uh  like a technical individual. He's the technology [31:16] officer or even a CIO. Well, what I see a lot  now with universities and specifically is that [31:22] is driven by business units because they see  these agents work really really well. So great, [31:26] we deployed a student agent. It's got talk to, you  know, I work for a very large university in New [31:31] Jersey, 40,000 students. We got 15,000, you know,  student is talking to this agent and um, you know, [31:39] and we ended up using over-the-c counter frontier  model. Very good one at that. Um, break the first [31:46] year, right? Everything's perfect. We have some  minor glitches. We're tweaking it. We're ragging [31:50] and modifying the output. Well, guess what?  In year two, um, we're starting the limits [31:56] of the technology, right? The context window. Um  because uh we were indiscriminate in training the [32:02] AI models on everything we had. Um and on top  of that the AI is generating its own synthetic [32:10] content on top. Uh it just ran out of context. So  he started getting really fuzzy about what we were [32:16] really really uh careful about in the earlier  phases. Um they just we just ran out of context [32:21] window to add more context to the conversation. It  was actually regressing in some way. Um and it's [32:27] just out remembering all the important stuff.  It wasn't able to have an executive function of [32:31] weighing what information is important or not.  And so how do you deal then with the challenges [32:36] to oh maybe it was not the right decision that we  should have made make it different. And then on [32:42] top of that every week they're changing the model  like we got you know 4.2 7.9. So we ended up kind [32:46] of saying okay we're going to have to freeze  this you know we don't want it to be smarter. [32:50] It's doing a good job. So, we ended up picking  one of the more open weight models that are [32:54] uh a weight uh relatively uh decent. It wasn't as  smart. It wasn't as fast, but um it was cheaper [33:00] anyway. And we ended up switching the model itself  um to lower the cost there, but we had, you know, [33:05] we incurred infrastructure cost there. Okay, got  it. So, yeah. I know there's about two years ago [33:11] we picked somebody some vendor very really good  I think state-of-the-art for what we were looking [33:17] to do and like year and a half now in I think we  can do it oursel so cheap faster actually not even [33:24] us marketing can do it in theirel and I'm like  okay what did we do and we stopped the project [33:29] right it's not not I would say not regrettable  you should do it it's you got lot of knowledge [33:35] and learning what you learned from that process  and that's the key part I would say think about [33:40] that investment that's what you learn and when you  pick the vendor that's what you look at are you [33:43] going to get knowledge or R&D from there that is  going to help you speed up the deliverables that [33:49] you need to get out and get to the market fast and  then you can make a decision if you need to switch [33:53] it in future okay so now we have the customer  perspective so now we just would I would like to [34:01] um go to the window perspective so when a customer  just said oh oh I was maybe the wrong decision So, [34:08] uh what are your um recommendations on that or  your experience with it when they just changing [34:16] their mind uh within the process? Let's say it  like this. Uh I've never heard a customer say [34:22] they made the wrong decision to buy our platform.  Well, good for you. [laughter] Hold up. Never. Um [34:28] of course. Right. So again, I'll go back to,  you know, without speaking too much about our [34:34] platform, you know, I'll go back to the whole  thing of we actually do have the capability to [34:39] um if you build applications,  Agentic applications on our platform, [34:44] um we have the ability to to give you that code  afterwards. Um, so if you, you know, it's it's [34:51] one of the things that we talk about on like  the first or second date, which you hate doing, [34:55] but it's like, you know, just so you know, I know  we just started dating, but if we get divorced, [35:00] you get to keep your stuff. Um, and it's a hard  conversation for us as a vendor to have, but at [35:06] the same time, it's an important conversation for  us to have. Um so sure we we've had some of these [35:12] um projects you know not go well and a lot of  times it's due as these guys probably will tell [35:17] you it's due to you know issues with scoping up  front um not really understanding the full aspect [35:23] of of um the full aspect of the project or the  capabilities needed to make the project successful [35:30] um and sometimes they just have to pull the plug  and and go at it themselves and do a custom build. [35:36] So and Ryan, yeah, I think for us it we're  pretty agnostic on either option. I mean, [35:42] where we live in the stack, again, it becomes  just a capital problem if they've decided, [35:47] I mean, we've worked with, you know, funny enough,  a few startups that they're just analyzing, oh, [35:52] how long can we use this infrastructure for?  And they run the numbers in their head like, [35:56] oh, actually, maybe I want to buy it. And it's  it makes no difference to us because we can help [36:00] them either way. That's the beauty of, you know,  open source tooling. if they've committed to using [36:05] that open source tooling, it can probably be used  for, you know, either a option, right? It's very [36:10] portable. So, like that's obviously our focus is  portability. Um, obviously like moving a server [36:15] from one data center to another is a challenge.  So, it's like basically like just who owns the [36:20] contract in that facility. Megawatts are scarce.  You know, everybody's chasing those. But, um, but [36:27] when it comes to like who owns the infrastructure,  that's where, you know, we can help either way. [36:34] So um so let's be honest so just keep let's uh  stay in that topic. So it was a challenging it [36:42] was a challenge to switch um the infrastructure  and all the also the build and buy decision [36:48] uh in the old times before AI. Um so with AI  will it be me be right now more challenging to [36:57] switch the infrastructure? um and what are your  recommendations uh to make the right decision [37:03] on the AI infrastructure? So uh especially first  uh at our um vendor um side I would like to know [37:13] um I I mean look I think AI has made it much  easier to make these these changes right so [37:19] um you know we're looking at tools out there  today that'll help you migrate you know these [37:25] legacy Cobalt applications legacy Lotus notes  applications things of that nature and you're [37:31] doing it now in the fraction of a time that you  were doing it 10, 15, 20 years ago. Um, and AI is [37:38] really the differentiator in that, you know, speed  change. Um, so I think AI has given you a lot of [37:46] that kind of um, insight and knowledge needed to  be able to kind of switch directions um, when you [37:55] need to. Um, so I I haven't seen any downside of  it, but I would love to hear if these guys are [38:01] seeing something. I'm I'm not from the vendor  side. I'm going to say something. Maybe it's [38:06] controversial, but uh I think the LLM models are  getting to a level where 70 to 80% of knowledge [38:13] workers work can be done with current LLMs. Like  yeah, research and everything else. Yeah. If you [38:18] want to do physics and crazy stuff, sure. Right.  So it's this point you can actually start locking [38:24] stuff. You can start saying hey it doesn't need to  be on cloud it could be running on kim if you're [38:28] allowed to do that but you can do a lot of those  stuff now right a lot of the knowledge work work [38:33] that happens practically you can start locking  that's the view I have okay Ryan so going back [38:40] to the vendor side yeah I mean for us kind of I  I guess what was the [laughter] original question [38:51] just hold on I need to uh I need to grab the right  thing So it's um uh it was challenging to switch [38:57] the infrastructure uh without AI or in the old  days just like let's say like yeah I mean I think [39:02] we're kind of back to the old school um in our  phase of the um or our level of the stack where [39:09] we're seeing obviously like we rely heavily on you  know the trades right these are physical people [39:15] running pipes people running cable right power  whips put the transformer in right so I guess in a [39:23] weird way return to old school and those jobs are  insanely in high demand and there's not enough of [39:28] them and that's why you pay way more to get those  things done and pay a lot more to get it done way [39:33] faster because there's so much money on the line.  So it's unique for us. I can't really as much [39:38] comment on which LLM are better because in the end  I don't care what LLM runs on the infrastructure [39:42] but you probably need it. So yeah, my our point is  to build an agnostic sort of framework that allows [39:51] the industry to progress into whatever you know uh  wherever it goes. So AI doesn't uh keep you away [40:00] from doing your homework, right? Right. Yeah. No,  you you still need to get who could do it fastest [40:05] and get the right labor in there in the right  amount of time. So and uh David, your thoughts on [40:10] that? It's a it's a great time to be a consumer. I  mean you have uh innovations that occurring every [40:16] day open source closed source you know proprietary  there's a massive competition to be the first you [40:21] have great vendors who have traditional you know  infrastructure values right they understand your [40:26] struggles I I care about my data I want it to  be portable you know I want it to be mine and [40:32] um and and there's a space for everyone and um  you know it's just a matter of having a lot of [40:38] conversations so um you know I I feel very lucky  to be in this space where there's so many people [40:43] helping you get to that that destination on the AI  side. Yeah, it's awesome. So, um according to the [40:49] time, so if you just would like to um give one  advice uh to the audience, I know that's pretty [40:58] hard, but it needs to be only once. So what you  would like to let them know out of your experience [41:04] and also your your um way of working and all what  you do right now. Uh so David you may start. Yeah. [41:14] So um start with a good use case. Um make sure  you you can define the value and pick the hardest [41:21] problems that you cannot solve right now. That's  my advice on tackling AI. Yeah. Ours would be, you [41:29] know, decide, you know, where you need to be from  a capital perspective and make those decisions. [41:35] If it's partnering with a higher level vendors or,  you know, pick a good consultative approach that, [41:40] you know, people can tell you use this or um like  where you need to live in the stack to get the [41:48] most value out of it. Yeah, mine's mine's still  going to be uh find a good hybrid solution. Um, [41:55] so you know, buy a platform that you can build  on that won't keep you kind of locked in. Um, [42:02] but the AI landscape, you know, it's shifting  daily, right? It used to be maybe weekly or [42:07] monthly or yearly, but now it's it seems like  it's daily. So, um, you know, make sure that [42:13] you have the right platform choices that you  make today that'll allow you to grow and have [42:19] a predictable cost structure on tomorrow. um and  set that foundation before you make any decisions [42:25] on how to use the AI. I I'll bring it back again  to the layer. Make a decision at the layer layer [42:33] layer level, not at the overall stack, right?  Saying should I buy this vendor's product or not? [42:38] It's just like what is the layer you're looking at  it and is it a make sense to buy or build? It's a [42:44] piece of puzzle. Okay, summarize my um what I just  learned from you right now. It's not about build. [42:52] versus by, right? Um, so it's build and buy. Um,  this is what what it's may the the main advice [43:01] uh we we could give you. Uh, and um, yeah,  make make and buy and do your homework. So, [43:10] um, when you would like to have some questions  because we are running out of time, these uh, [43:16] four guys are available for your questions. So  if you would like to um you will just we'll be [43:22] outside there anywhere uh and then just  reach out to them. Thank you. Thank you.