Vultr CMO Kevin Cochrane Urges Healthcare Cloud Overhaul for AI

Vultr CMO Kevin Cochrane呼吁医疗行业重构云架构以支持AI 医疗服务提供商必须采用可组合的云基础设施，以便在消除遗留技术债务并满足严苛合规标准的同时，安全部署AI工具。 Kevin Cochrane指出，医疗行业必须消除遗留技术债务和运营复杂性，从而为患者护理和临床工作流引入实时的智能体AI（agentic AI）。结合显示AI支出快速增长但合规担忧持续存在的调查数据，他主张采用平台工程原则以及标准化、可组合的基础设施模板，而非脆弱的GPU配置。Cochrane阐述了即开即用的全球算力与安全的混合多云网络如何为持续模型重训和实时RAG流水线提供支持。这一可组合蓝图使机构能够在加速生物制剂发现等繁重科研工作的同时，为临床医生和患者提供差异化、个性化的体验。 Vultr 调查显示，93% 的医疗机构正在增加 AI 支出，89% 正在评估替代云平台；但在盲目追求裸机 GPU 算力之外，数据隐私与合规治理仍是落地的首要考量。 (4:05) 传统数字医疗在常规化验复查上面临长达数周的延误，医疗行业正从孤立的本地科研训练转向基于实时智能体（Agentic AI）的即时个性化临床推理。 (1:42) 为避免 GPU 集群沦为运维繁杂的“雪花”环境，医疗 IT 需采用平台工程模式，利用预先验证的可组合模板（Golden Paths）将合规与安全要求直接内嵌至下游应用。 (10:05) 构建基于私有直连的混合多云网络架构，可在不暴露于公共互联网的前提下安全传输边缘诊断与实验室数据，配合秒级开通的全球弹性算力保障低延迟推理与数据本地化合规。 (17:35)

Kevin Cochrane, Chief Marketing Officer at Vultr, contends that scaling artificial intelligence in healthcare requires replacing bespoke, complex infrastructure with standardized, composable cloud templates. He argues that this approach allows health organizations to overcome technical debt, maintain strict regulatory compliance, and support real-time clinical inference. Healthcare Priorities Shift Toward Real-Time Inference (4:05) Digital healthcare tools have successfully modernized patient scheduling and remote appointments, yet Cochrane points out that basic clinical workflows remain slow, citing typical scenarios where patients wait two weeks for a consultation merely to review standard lab tests. Cochrane asserts that interactive AI agents can eliminate this latency by providing immediate, individualized evaluations. While early healthcare AI focused heavily on specialized, on-premises research workloads like medical imaging and genomics, industry demand is rapidly pivoting toward production inference to assist doctor workflows and answer patient questions in real time. This architectural evolution is driving substantial shifts in IT budgets. According to a Vultr survey of 2,000 global CIOs and senior AI decision-makers—including 300 healthcare executives—93% of healthcare organizations are expanding their AI infrastructure spending, and 89% are evaluating alternative cloud providers to improve cost efficiency and development velocity. Nonetheless, respondents identified patient data privacy, security, data sovereignty, and regulatory compliance as the primary bottlenecks preventing wider deployment. Standardizing GPU Deployment via Platform Engineering (8:00) Cochrane argues that GPU clusters have traditionally been treated as delicate "snowflakes" that require scarce hardware expertise and intensive manual configuration. To lower costs and operational barriers, he maintains that launching and scaling AI compute must become as straightforward as deploying conventional web servers or mobile applications, enabling enterprise developers to build AI tools without low-level infrastructure engineering. To achieve this without compromising governance, Cochrane recommends a centralized platform engineering approach. In this model, a dedicated central team curates and pre-validates foundational models, microservices, containerized data pipelines, and infrastructure templates into a standardized "golden path." Clinical developers building distinct applications—such as workflow assistants for hospital physicians or diagnostic query tools for patients—deploy these validated templates directly, ensuring compliance policies and security standards are enforced across all deployments. Hybrid Networking and Sub-Second Compute Scaling (15:15) Cochrane critiques standard cloud deployment times that require hours to provision compute clusters, arguing that healthcare workloads demand elastic compute that can scale up and down dynamically within seconds. Meeting the low-latency requirements of medical workflows also requires geographic distribution, which Vultr delivers across 32 global data center regions to keep processing near the point of care while meeting local data residency regulations. Because raw patient data originates outside public clouds—such as on-premises hospital medical devices, clinic diagnostics, and commercial laboratories like Quest Diagnostics—Cochrane emphasizes the necessity of hybrid multi-cloud networking. Using private direct connects, healthcare organizations can stream diagnostic data directly from edge environments to cloud compute nodes, preventing sensitive patient records from traversing the unencrypted public internet. Closed-Loop Retraining and Multi-Persona Retrieval (19:44) Real-time clinical AI systems rely on continuous data feedback rather than static models. Cochrane describes an architecture where real-time diagnostics stream into a vector database to power Retrieval-Augmented Generation (RAG), while longitudinal time-series metrics feed continuous model retraining pipelines. This closed feedback loop allows algorithms to iteratively adjust recommendations based on documented health outcomes rather than fixed rules. The vector retrieval architecture is designed to support differentiated user experiences from a unified data foundation, delivering distinct technical contexts and terminology to clinician interfaces while generating clear guidance for patient applications. Cochrane adds that these composable deployment patterns extend beyond conversational agents to computational research, citing IPA Therapeutics' use of standardized high-density GPU clusters for biologics and drug discovery. Cochrane concludes that abstracting infrastructure complexity into composable cloud templates allows healthcare institutions to cut operational overhead and focus on clinical models and patient outcomes. However, the framework relies heavily on organizations successfully orchestrating edge-to-cloud private networks and validating automated agent guidance within strict clinical compliance regimes.

Vultr首席营销官Kevin Cochrane指出，医疗AI正在从早期的离线研发转向大规模实时临床推理，但传统架构的时延与定制化硬件运维构成了主要瓶颈；医疗机构亟需通过平台工程与可组合云基础设施，在保障数据主权与严苛监管合规的前提下实现即时、低延迟的智能化诊疗。 临床就医延迟与基础设施转型诉求 (1:42) 医疗数字化虽然普及了线上预约与远程问诊，但底层临床流程依然存在显著滞后。Cochrane结合自身经历说明，患者在常规血液检查完成且数据已存入数据库后，往往仍需等待两周的远程复诊才能得知结果。他主张，实时交互的AI智能代理具备消除此类流程延迟、按需提供个性化健康指引的潜力，但这要求医疗IT系统彻底革新其底层基础设施。 Vultr针对北美与欧洲2000多位CIO及AI预算决策者（包含300位医疗行业高管）的全球调研印证了这一紧迫性：93%的医疗机构正主动增加AI基础设施投入，89%的机构正在突破传统IT架构并评估替代云服务商，以追求更优的成本效率与开发速度。然而，调研同时显示，数据隐私、患者数据安全、数据主权及合规监管，仍是阻碍医疗企业全面应用AI的首要顾虑。 平台工程消解定制化GPU运维摩擦 (8:00) 长期以来，医疗行业往往将GPU集群视作脆弱且需要昂贵专属人才维护的特殊定制硬件。Cochrane批评了这一模式，指出AI算力必须向标准化、日用品化的公共资源演进，其部署与销毁应如同配置传统Web服务器一样简便，避免让普通工程团队承担底层硬件配置的重负或引发系统安全隐患。 为兼顾敏捷交付与严格合规，Cochrane提倡基于平台工程原则的可组合云基础设施架构。在该模式下，中心化平台工程团队通过规范的产品化流程，统一测试并封装AI模型、微服务、容器化数据流水线以及底层算力与网络模板。当下游开发者为医院临床医生、门诊专家或患者端构建应用时，直接调用预先验证的“黄金路径”（Golden Path），使得合规审查、数据主权与安全策略在基础设施模板层自动落地，消除重复造轮子带来的风险。 混合多云私有直连与全球即时扩缩 (17:35) 医疗数据广泛分布于医院设备、诊所检测硬件及Quest Diagnostics等商业检测机构，决定了医疗AI无法在孤立的单一云孤岛中运行。为保证数据传输安全，机构必须构建具备私有直连的加固型混合多云网络，确保患者原始诊断数据在从边缘设备流向云端节点的过程中完全隔绝于公共互联网。 针对行业中普遍存在的数小时集群调配延迟，Cochrane主张算力必须具备“秒级即时开启”（Instant-On）能力，动态匹配波动的临床负载。以Vultr为例，其分布在全球的32个云数据中心节点将高性能GPU与弹性CPU、低延迟网络及分布式存储整合，使医疗机构不仅能够在靠近诊疗点的位置实现亚秒级推理响应，还能切实满足各司法管辖区对数据驻留合规的法定要求。 闭环持续重训与医患双轨RAG流水线 (19:44) 医疗AI的效能取决于基于实际结果反馈的闭环更新机制。当系统给出的生活干预建议未能带来指标改善时，闭环数据流水线会采集纵向时序数据与遥测反馈并重新输入训练集，动态调整模型策略；同时，实时的检验数据会被摄入向量数据库以支持检索增强生成（RAG）。 该向量数据底座支持运行面向两类不同对象的双轨交互界面：系统根据同一底层事实数据源，分别向医护人员生成专业术语与诊断辅助信息，向患者生成通俗易懂的指导建议。此外，这种标准化模板同样适用于重度科研计算，如IPA Therapeutics即依托预置的大规模GPU集群模板进行高通量生物制剂与药物研发，免除了从零构建底层编排系统的运维负担。 Cochrane主张通过预置模板与可组合云平台将底层硬件复杂性彻底抽象化，使医疗IT与研发人员专注于模型调优、人机交互与临床结果本身；但该方案的全面普及，仍取决于跨机构边缘数据整合的实际难度以及医疗监管机构对实时推理模型的信任建立过程。

## Background 本内容由 YouTube 频道 Ai4 于 2025 年 8 月 28 日发布，记录了 Vultr 首席营销官（CMO）Kevin Cochrane 的主题演讲。根据页面介绍，该演讲围绕医疗机构如何通过可组合云基础设施（Composable Cloud Infrastructure）实现 AI 就绪展开，探讨了在确保安全、合规与互操作性的前提下加速医疗创新与业务转型的落地策略。 ## TL;DR Vultr CMO Kevin Cochrane 从自身看诊苦等化验报告的痛点切入，指出 AI 能够为数十亿人带来即时且个性化的医疗洞察。结合针对 2000 位技术决策者的调研数据，他指出 93% 的医疗机构正在扩大 AI 算力预算，但数据安全与架构复杂性仍是最大阻碍。为此，演讲提出了基于平台工程与“黄金路径”的可组合云基础设施方案，利用混合多云安全直连、实时 RAG 与持续重训机制驱动智能体 AI（Agentic AI），并以 IPA Therapeutics 的生物制剂研发为例，展示了一键编排高性能 GPU 算力以普惠全球医疗健康的实践蓝图。 ## Atmosphere & style 整场演讲开场风趣幽默，讲者通过抛洒吉祥物“莫蒂”和调侃打匹克球迅速拉近与观众的距离，随后话锋一转切入个人就医体验，引发强烈共鸣。叙事节奏明快且极具前瞻性，既有对行业痛点的一针见血和犀利批判，又有翔实的数据支撑与硬核的技术架构剖析，营造出一种兼具专业说服力与变革激情的现场氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=o6zbQThsuYs&t=4s) 医疗健康行业正在重构基础设施战略，在探索全新疗法与应用的同时，安全与合规是企业的核心生命线。 - [01:42](https://www.youtube.com/watch?v=o6zbQThsuYs&t=102s) AI 重塑医疗体验的关键在于打破传统看诊与化验报告的冗长等待周期，为患者提供即时且高度个性化的健康洞察与方案。 - [04:05](https://www.youtube.com/watch?v=o6zbQThsuYs&t=245s) 调研显示医疗行业转型速度领跑各行各业，93% 的机构正在增加 AI 算力预算，89% 积极寻求更具性价比的替代型云服务商。 - [04:05](https://www.youtube.com/watch?v=o6zbQThsuYs&t=245s) 数据安全、隐私保护与合规驻留仍然是阻碍医疗机构最大化释放 AI 潜能的首要挑战。 - [04:05](https://www.youtube.com/watch?v=o6zbQThsuYs&t=245s) 医疗 AI 正从早期的医学影像与药物研发训练，加速延伸至融入医生工作流管理与患者即时交互的生产级推理阶段。 - [08:02](https://www.youtube.com/watch?v=o6zbQThsuYs&t=482s) 通过平台工程铺设“黄金路径”（Golden Path），将模型、合规策略与微服务预置为模板，大幅降低下游开发者构建 AI 应用的门槛。 - [08:02](https://www.youtube.com/watch?v=o6zbQThsuYs&t=482s) 新一代可组合云基础设施要求实现跨区域 GPU 与 CPU 资源的秒级就绪与弹性伸缩，推动算力演进为随取随用的大宗商品。 - [17:35](https://www.youtube.com/watch?v=o6zbQThsuYs&t=1055s) 医疗智能体（Agentic AI）的规模化落地依托于混合多云与专用直连通道，确保边缘生理数据在传输与推理全程与公网完全隔离。 - [19:44](https://www.youtube.com/watch?v=o6zbQThsuYs&t=1184s) 建立双路径数据闭环：长期时间序列数据驱动模型周期性微调重训，新鲜化验数据即时接入向量数据库支持实时 RAG 响应。 - [19:44](https://www.youtube.com/watch?v=o6zbQThsuYs&t=1184s) 统一底层架构可根据角色差异化分发输出：对医生提供专业临床决策支持，对患者提供通俗实用的健康干预建议。 - [20:43](https://www.youtube.com/watch?v=o6zbQThsuYs&t=1243s) 基础设施“一键部署”预制模板消除了手工组装算力集群的风险与成本，加速生物制药与药物发现等高性能计算应用落地。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=o6zbQThsuYs&t=4s) 破冰互动与医疗基础设施转型的重要性 Vultr 首席营销官（CMO）Kevin Cochrane 开场介绍了 Vultr——全球最大的专注于 AI 基础设施的独立公有云平台。为了活跃现场气氛，他拿出了公司的吉祥物“莫蒂”（Morty）向观众抛洒。莫蒂是一只喜马拉雅兀鹫（Himalayan vulture），它的标志性口号是“莫蒂专吃技术债”（Morty Eats Tech Debt）。Kevin 幽默地分享说，这些毛绒玩具手感蓬松、深得宠物狗喜爱，而且形状和弹性极佳，用来在公寓里打无声匹克球（pickleball）甚至不会吵到邻居。 在玩笑之余，Kevin 切入正题指出，抛出莫蒂的真正寓意在于：医疗健康行业必须彻底反思现有的基础设施战略，重新思考如何利用云的力量推动全新的医疗成果。当前算力基础设施赛道异常火热，几乎每家公司都在融资并提供“GPU 即服务”（GPU as a Service）。但对于医疗从业者而言，安全与合规才是核心生命线；只有建立在安全稳妥的基础设施之上，企业才能在先锋探索全新疗法与拓展医疗服务边界的同时，切实保障公众利益与数据安全。 --- ### [01:42](https://www.youtube.com/watch?v=o6zbQThsuYs&t=102s) 个人看诊经历与 AI 重构医疗体验的愿景 Kevin 分享了自己参会前的一段真实亲身经历：在来会场之前，他刚刚完成了年度常规血液检测。抽完血后，他脑海中浮现的第一个念头就是：“我什么时候能拿到化验结果？我到底健不健康？”在早期 Web 应用时代，医疗机构通过数字化重构了客户体验，提供了更便捷的预约途径；疫情期间，远程医疗（Telehealth）又进一步普及。就在演讲开始前 30 分钟，Kevin 收到医生的通知，让他在手机上预约两周后的远程视频问诊，以解读化验报告。 Kevin 对此提出质疑：虽然数字化预约和免去奔波的远程医疗非常方便，但化验报告明明已经出来了，为什么患者还必须苦等整整两周，才能得到一个基础问题的答案？为什么不能直接问一句“我的身体状况到底好不好”？他强调，这正是 AI 赋予医疗的核心力量——让全球数十亿人都能在瞬间获得高度个性化的医疗洞察与健康方案。为了实现这一目标，医疗行业需要像当年拥抱 Web 数字化一样，全面重构技术栈并投资新一代基础设施架构。 --- ### [04:05](https://www.youtube.com/watch?v=o6zbQThsuYs&t=245s) 医疗行业 AI 演进阶段与 2000 位决策者调研洞察 在医疗 AI 发展的早期阶段，基础设施主要局限于各机构在本地私有部署 GPU 集群，用于训练医学影像分析或药物发现模型。而当前的发展趋势则是推动 GPU 基础设施大幅降低成本与复杂度，以加速医学影像、药物研发以及一系列能为全球患者带来实效的新型应用落地。 ```mermaid flowchart LR accTitle: 医疗行业 AI 支出与驱动因素 subgraph S1["支出与架构意向"] A["93% 机构增加 AI 算力预算"] B["89% 探索替代型云服务商"] end subgraph S2["核心考量与阻碍"] C["数据安全与隐私保护"] D["数据合规与主权落地"] end S1 --> S2 ``` 为了量化这一趋势，Vultr 近期对北美和欧洲的 2,000 多位负责 AI 预算的 CIO 及高级决策者进行了全球调研（北美 1,000 位、欧洲 1,000 位，涵盖多个行业，其中 300 位专门来自医疗行业）。调研显示出几项关键结论： - **医疗行业转型速度领跑各行各业**：医疗行业在探索 AI 实际应用和加速公共健康服务方面处于最前沿。 - **93% 的医疗机构正在增加 AI 基础设施支出**：算力投资进入全面扩张期。 - **89% 的机构正在打破传统基础设施路径**：积极寻求更具成本效益、迭代更快的替代型基础设施提供商。 - **首要阻碍仍是安全合规**：数据保护、数据安全、隐私与合规性是阻碍企业最大化释放潜能的核心挑战。 随着行业从早期的影像、基因组学与药物发现扩展到生产级推理（Production Inference），AI 的应用场景正深度融入临床医生的工作流管理、诊疗建议优化以及面向患者的即时健康交互。 --- ### [08:02](https://www.youtube.com/watch?v=o6zbQThsuYs&t=482s) 可组合云基础设施与平台工程的“黄金路径” 传统 GPU 基础设施往往像“定制雪花”（snowflake）一样复杂脆弱，不仅面临多地区数据驻留（Data Residency）与安全风险，还需要极高门槛的专业技能。Kevin 质问：“为什么构建和部署 AI 原生应用不能像搭建网站或移动应用一样简单？算力在本质上难道不应该是一种随取随用、快速扩缩的大宗商品（commodity）吗？” ```mermaid flowchart TD accTitle: 可组合基础设施与黄金路径架构 subgraph Central["中心化平台工程团队 (Platform Engineering)"] T1["预构建 AI 模型资产"] T2["标准化微服务 / 容器"] T3["安全合规基础设施模板"] end subgraph Developers["下游开发人员 (Downstream Developers)"] D1["医生工作流 Agent"] D2["临床诊疗辅助 Agent"] D3["患者即时问答 Agent"] end Central -->|"提供'黄金路径' (Golden Path)"| Developers ``` 为了彻底消除这种复杂性，Vultr 提出了**可组合基础设施模型（Composable Infrastructure Model）**，通过引入平台工程（Platform Engineering）方法实现集中管控与敏捷赋能的平衡： 1. **产品化模板管理**：由中心化平台工程团队将核心 AI 模型、微服务、容器化数据与底层算力/存储/网络模板进行预构建、预打包、预测试与预验证，将合规与安全策略内置于模板之中。 2. **铺设“黄金路径”（Golden Path）**：下游普通开发人员无需重复造轮子或操心 GPU 集群运维，只需调用统一的智能体（Agent）模板，即可快速上线面向医生、临床人员或患者的各类 AI 服务。 3. **秒级算力调度**：Kevin 批评了展厅中宣称“数小时内拉起 GPU 集群”的传统做法，指出新一代基础设施的标准必须是全球 32 个数据中心区域内任意规模 GPU 与 CPU 资源的“瞬时秒级就绪与自动弹性扩缩”。 --- ### [17:35](https://www.youtube.com/watch?v=o6zbQThsuYs&t=1055s) 混合多云架构与智能体 AI（Agentic AI）落地实践 在医疗场景中，医疗设备、医院系统和第三方化验室（如 Quest Diagnostics）始终分布在本地边缘侧，源源不断地采集实时生理数据。为了在保证绝对安全的前提下进行大模型推理，医疗机构必须采用硬化的混合多云与专用直连通道（Private Direct Connect），确保数据在从本地传输至 Vultr 等云平台的过程中完全隔离于公网之外。 Kevin 举例说明了智能体 AI 在日常生活中的实际价值：当患者在 Quest Diagnostics 完成化验后，数据经由安全专用网络汇总到云端，患者向专属 Agent 询问健康状况，Agent 可以立即反馈：“你的血糖水平再次偏高，因为你每晚总是忍不住吃一夸脱冰淇淋。”这种面向规模化智能体 AI 的端到端落地架构，依托于将本地实时流式数据安全注入云端的分类处理管道之中（首先是进入持续训练与微调的数据存储桶，以便根据治疗与干预结果动态优化模型建议）。 ### [19:44](https://www.youtube.com/watch?v=o6zbQThsuYs&t=1184s) 数据闭环驱动模型重训与实时 RAG 体验 在个性化医疗服务中，模型并非一成不变，而是需要根据观察到的时间序列数据持续进行重新训练。凯文·科克伦（Kevin Cochrane）以自身为例风趣地解释道，如果系统一开始建议“早餐多吃麦片”，但实际监测发现这对健康毫无帮助，模型就必须在持续迭代中捕捉到规律——“原来吃麦片对凯文·科克伦不管用，戒掉冰淇淋才见效”，以此逐步优化治疗效果。这种机制依赖于两条并行的数据路径：一条是从日常收集的用户行为与体验数据中流式提取训练集，用于模型的周期性重训；另一条则是将实时的关键数据（如最新的化验报告与实验室检查结果）即时写入向量数据库（Vector Database），以支持实时检索增强生成（RAG）和智能体（Agent）的高效运行。 ```mermaid flowchart TD accTitle: 医疗实时 RAG 与模型持续微调架构流程图 subgraph DataSources["数据输入源"] A["用户体验与时间序列数据"] B["实时化验与体检报告"] end subgraph Processing["核心处理与存储"] C["模型持续训练与微调"] D["向量数据库 (Vector DB)"] end subgraph DualExperience["差异化终端体验"] E["医生端体验 (临床决策支持)"] F["患者端体验 (个性化健康建议)"] end A --> C C --> D B --> D D --> E D --> F ``` 通过向量数据库驱动的智能体验能够同时支撑两类截然不同的服务对象：临床医生与患者。针对同一批数据，系统会调用不同的模型策略并生成风格迥异的输出——对医生提供专业的诊疗决策支持，对患者则输出通俗易懂的个性化健康指导。 - **持续重训机制**：依赖长期追踪的时间序列数据，动态修正不恰当的干预策略（如用“戒掉冰淇淋”替代无效的“吃麦片”建议）。 - **实时 RAG 架构**：化验结果等新鲜数据直接接入向量数据库，确保智能体在毫秒级内响应临床与患者端的查询。 - **双重交互体验**：同一套底层架构能够根据医生与患者的角色差异，分发针对性、差异化的模型响应。 --- ### [20:43](https://www.youtube.com/watch?v=o6zbQThsuYs&t=1243s) 基础设施一键编排与生命科学计算实践 为了支撑复杂的 AI 医疗管线，底层基础设施的搭建必须极致简化。如果让医疗机构的每一个开发团队都自行研究如何组装和连接整套云端基础设施，不仅会推高研发成本，还会引入巨大的系统复杂性与合规风险。Vultr 提供的解决方案是通过预置模板实现“一键部署”（Single click），将所有底层网络连接与计算资源全部自动化打通。这种“开箱即用”的设计让开发团队无需操心集群扩展与硬件连通细节，从而能够将全部精力聚焦在模型算法、模型训练以及最终能够为用户带来实际价值的临床体验上。 除了面向交互式的 Agentic AI 应用，在硬核的医学研究领域，高性能算力同样不可或缺。以 IPA Therapeutics 为例，该机构利用大规模 GPU 集群进行前沿的生物制剂研发与药物发现（Biologics Discovery）。Vultr 将行业顶尖实践固化为可复用的基础设施模板，帮助医疗研发机构快速启动并降低试错门槛。在医疗行业快速迭代、安全与合规要求极高的背景下，通过可组合的云架构降低成本与技术复杂度，最终目标是实现让全球每个人都能享受到高质、精准的个性化医疗成果。 - **基础设施一键编排**：消除手工搭建底层架构的风险与成本，提供即开即用的预制模板。 - **生物医药科研落地**：IPA Therapeutics 借助大规模 GPU 集群加速先进生物制剂发现。 - **医疗创新愿景**：在满足严苛安全与合规要求的前提下，以更低门槛驱动全球个性化医疗普惠。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=o6zbQThsuYs&t=4s) 莫蒂专吃技术债 — Kevin Cochrane > [01:42](https://www.youtube.com/watch?v=o6zbQThsuYs&t=102s) 虽然数字化预约和免去奔波的远程医疗非常方便，但化验报告明明已经出来了，为什么患者还必须苦等整整两周，才能得到一个基础问题的答案？为什么不能直接问一句“我的身体状况到底好不好”？ — Kevin Cochrane > [04:05](https://www.youtube.com/watch?v=o6zbQThsuYs&t=245s) 医疗行业在探索 AI 实际应用和加速公共健康服务方面处于最前沿。 — Kevin Cochrane > [08:02](https://www.youtube.com/watch?v=o6zbQThsuYs&t=482s) 为什么构建和部署 AI 原生应用不能像搭建网站或移动应用一样简单？算力在本质上难道不应该是一种随取随用、快速扩缩的大宗商品吗？ — Kevin Cochrane > [17:35](https://www.youtube.com/watch?v=o6zbQThsuYs&t=1055s) 你的血糖水平再次偏高，因为你每晚总是忍不住吃一夸脱冰淇淋。 — Kevin Cochrane > [19:44](https://www.youtube.com/watch?v=o6zbQThsuYs&t=1184s) 原来吃麦片对凯文·科克伦不管用，戒掉冰淇淋才见效 — Kevin Cochrane ## Mentioned Kevin Cochrane, Vultr, 莫蒂 (Morty), Quest Diagnostics, IPA Therapeutics, Ai4 ## Video description Kevin Cochrane, CMO, Vultr Abstracts Is your healthcare organization prepared to harness the full potential of AI? Composable cloud infrastructure offers the flexibility, scalability, and performance needed to support AI adoption and drive innovation across clinical workflows, patient engagement, and operational efficiency. In this session, you'll learn how forward-thinking healthcare providers are modernizing their IT environments to become AI-ready—while maintaining compliance, security, and interoperability. Discover practical strategies for building a composable architecture that accelerates innovation and simplifies integration of new technologies. Join Vultr, the world’s largest privately held cloud platform, to explore the blueprint for transforming healthcare and positioning your organization for the future. [Source](https://www.youtube.com/watch?v=o6zbQThsuYs)

[0:04] Kevin Cochrane: 莫蒂专吃技术债 [1:42] Kevin Cochrane: 虽然数字化预约和免去奔波的远程医疗非常方便，但化验报告明明已经出来了，为什么患者还必须苦等整整两周，才能得到一个基础问题的答案？为什么不能直接问一句“我的身体状况到底好不好”？ [4:05] Kevin Cochrane: 医疗行业在探索 AI 实际应用和加速公共健康服务方面处于最前沿。 [8:02] Kevin Cochrane: 为什么构建和部署 AI 原生应用不能像搭建网站或移动应用一样简单？算力在本质上难道不应该是一种随取随用、快速扩缩的大宗商品吗？ [17:35] Kevin Cochrane: 你的血糖水平再次偏高，因为你每晚总是忍不住吃一夸脱冰淇淋。 [19:44] Kevin Cochrane: 原来吃麦片对凯文·科克伦不管用，戒掉冰淇淋才见效

[0:04] Kevin Cochrane: Morty eats tech debt. [0:04] Kevin Cochrane: GPU as a service [1:42] Kevin Cochrane: Am I doing okay or not? [8:00] Kevin Cochrane: snowflakes [8:00] Kevin Cochrane: dead simple [10:05] Kevin Cochrane: golden path [15:15] Kevin Cochrane: GPU clusters in hours. [15:15] Kevin Cochrane: instant on [19:44] Kevin Cochrane: eat more cereal for breakfast [19:44] Kevin Cochrane: eating cereal doesn't help Kevin Cochrane; actually, it's cutting out the ice cream that's going to help him.

[0:04] All right. Well, everyone, I'm Kevin [0:05] Cochran, chief marketing officer here at [0:07] Vulture. Vulture is the world's largest [0:10] independent public cloud specializing in [0:12] AI infrastructure. And I happen to have [0:14] here our mascot. Our mascot's name is [0:17] Morty. He's a Himalayan vulture. And [0:20] Mort's uh claim to fame is Morty Eats [0:23] Tech. And so before we begin, I want to [0:26] see who wants an actual Morty in the [0:28] crowd here. All right, let's throw some [0:30] of these things out. Uh, they're super [0:33] fun. They're super fluffy. Dogs love [0:36] them. Uh, they're also the perfect shape [0:39] for pickle ball and have a perfect [0:41] bounce but have no noise. So, all of the [0:44] people in my condominium complex love me [0:46] for it. So, let's throw them all out. [0:49] And again, Morty eats tech debt because [0:51] we need to rethink our infrastructure [0:54] strategies. We need to rethink how we're [0:56] leveraging the cloud so we can start [0:59] advancing new health care outcomes, [1:01] leveraging the power of infrastructure. [1:04] As you all know, the infrastructure [1:06] space is very hot. Every single company [1:08] seems to be raising capital and offering [1:10] GPU as a service. But you as uh [1:13] practitioners in the health care space [1:15] know that security compliance really do [1:18] matter and that you need to have safe [1:20] and sane strategies for your [1:22] infrastructure if you're going to [1:23] protect the public while you're actually [1:25] starting to advance and pioneer new [1:28] therapies and new access to health care [1:31] services. So I'll stop throwing these [1:34] things out, but you can always get more [1:35] at our booth or come up and see me here [1:37] after the stage. And now let me tell you [1:40] a little bit about a personal story. [1:42] Okay. So obviously I'm here at this [1:44] conference and just prior to coming to [1:47] the conference I actually got my annual [1:50] blood test and after I got my blood test [1:53] the first thing that goes through my [1:54] head is when do I get my lab results and [1:58] when can I find out if I'm healthy or [2:00] not? Now in the early days of [2:04] infrastructure when we had web [2:06] applications, we had the power as [2:09] healthcare providers and practitioners [2:11] to reinvent the customer experience and [2:13] provide [2:15] immediacy of access to information and [2:18] services. And right now here today, I'm [2:22] being pinged literally 30 minutes ago by [2:25] my clinician to schedule an appointment [2:30] on my mobile phone for a tea health [2:33] session with my doctor in 2 weeks to [2:37] find out my lab test results. Now, that [2:39] was wonderful in the early days of the [2:41] web that as healthcare providers, we [2:43] actually provided the new services, the [2:45] digital services so you can schedule an [2:47] appointment. And with co we introduced [2:50] new services like tellaalth but let me [2:54] ask you this I actually just want to [2:57] understand how I'm doing and I'm [3:00] traveling and as easy as it is to [3:03] schedule an appointment as wonderful as [3:05] it is to have tellaalth so I don't have [3:07] to actually make a physical appointment [3:09] I still have to wait two weeks my lab [3:12] results are in why can't I ask a simple [3:15] question why can't I just understand am [3:19] I doing okay or not? This is the power [3:22] that AI can bring to us. How do you [3:25] unlock personalized healthc care [3:28] insights and personalized health care [3:31] outcomes at global scale for every [3:35] person on the planet? Just as we [3:38] invented new digital services and new [3:41] means of accessing clinicians and [3:44] physicians, how do we do that again and [3:47] take it to the next level and reinvent [3:51] our technology stack and invest in new [3:54] infrastructure strategies so that we can [3:56] truly unlock the power of delivering [4:00] personalized healthc care outcomes to [4:02] every single person on the planet. Now [4:05] in this early days of the AI revolution [4:08] in healthcare, infrastructure was [4:10] largely people building and deploying [4:13] GPU clusters on premises to build models [4:18] for medical imaging or for drug [4:21] discovery. But now as we're moving [4:24] forward, we need to start taking the [4:27] next step and reducing the cost and [4:30] complexity of scaling up GPU [4:32] infrastructure so we can start [4:35] accelerating medical imaging, start [4:37] accelerating drug discovery, and start [4:40] accelerating a new class of applications [4:42] that can deliver demonstrable benefits [4:45] to people all around the planet. This is [4:48] what we're seeing here at Bolure, which [4:51] is a rapid evolution of AI in healthcare [4:56] and specifically the consequent [4:58] implications it's having on [5:00] infrastructure strategies and as a [5:03] specialist in AI infrastructure that's [5:05] focused on the healthcare industry and [5:07] focused on the security and compliance [5:09] and governance that healthcare requires. [5:11] This is what makes this moment so [5:14] important for us. Now, we recently [5:16] concluded a worldwide survey of over [5:18] 2,000 CIOS and senior decision makers [5:21] responsible for AI budgets across North [5:23] America and Europe. A thousand North [5:25] America, thousand in Europe, across a [5:27] variety of industries, 300 in healthcare [5:30] industry specifically. And the most [5:32] important thing that came out of that [5:34] survey, the biggest insight is you all [5:36] here in this room [5:38] are moving faster than any other [5:41] industry. Healthcare is leading the [5:43] charge in terms of thinking through the [5:47] practical applications of AI and how it [5:50] can accelerate new services to citizens [5:54] around the planet. What we learned is [5:57] that fully 93% of healthcare [6:00] organizations are increasing their spend [6:02] on AI infrastructure. 89% are looking to [6:07] move beyond traditional approaches that [6:10] they've had for infrastructure, [6:12] specifically looking at alternative [6:14] infrastructure providers that can help [6:16] them be more costefficient and move [6:19] faster. But the not surprisingly [6:22] the number one requirement and the [6:24] number one thing holding enterprises [6:27] back is concerns around data protection, [6:31] data security, data privacy and [6:34] compliance. This is the biggest [6:37] challenge that we need to face is how do [6:40] we help you move fast with the best cost [6:43] efficiency, eliminate the complexity, [6:46] but also eliminate any risk with [6:49] leveraging new strategies for building [6:52] and deploying new AI models globally. [6:56] And what we're working on here at Bolure [6:58] is helping people in the healthcare [6:59] industry across a variety of spectrum uh [7:03] to advance different AI initiatives. As [7:06] I mentioned in the early market, a large [7:09] part of the innovation you were seeing [7:11] in AI model was in medical imaging. It [7:14] was in genomics. It was in drug [7:16] discovery. And now as we're starting to [7:19] move beyond and mature as an industry, [7:22] we're starting to see health care [7:24] providers and practitioners start [7:26] thinking about the physician experience, [7:28] the clinician experience, and the [7:30] patient experience and how we can start [7:33] building and deploying models for [7:35] production inference to help the doctor [7:39] manage their workload to help the [7:42] clinician [7:43] provide better recommendations and [7:45] better quality of care and to help [7:48] patients like myself just have easy [7:52] access to understanding like am I [7:54] healthy? Am I doing okay? And get a [7:58] valid response. [8:00] So [8:02] what we're here at Bolture trying to do [8:04] is help try to understand what are the [8:07] things that are holding back the [8:08] healthcare industry from a pure [8:11] infrastructure perspective from moving [8:14] even faster. How do we help you secure [8:17] your data, ensure the privacy, ensure [8:21] data residency in different geographic [8:23] regions when you're trying to serve a [8:26] diverse perhaps global population? [8:29] There's a lot of challenges at that [8:32] security and computational complexity [8:35] when you're standing up new [8:36] infrastructure [8:38] that engenders a bunch of security risk. [8:41] This is new infrastructures in new [8:43] locations. It also brings a lot of [8:46] complexity. GPU infrastructure [8:49] is often times a snowflake. It requires [8:52] specialty skills, specialty [8:55] practitioners. How do we make that [8:58] simple? How do we make that secure? How [9:00] do we make it so that every single [9:01] developer in your IT organization can [9:04] have the power and the efficiency at [9:07] their fingertips to build, deploy, and [9:10] scale AI native applications just as [9:13] easily as they can spin up a website or [9:15] a mobile application? Why does it have [9:18] to be so hard? Why can't it be simple? [9:21] Isn't compute fundamentally something [9:24] that should be commoditized and easy to [9:27] access and easy to spin up and easy to [9:30] spin down? Why do we have to have the [9:32] complexity? Why do we have to have the [9:34] cost? Why do we have to have the [9:36] specialty skill set in our workforce and [9:40] why do we have to have such challenges [9:43] adopting broadly new infrastructure to [9:47] deploy AI native applications? I would [9:50] suggest to you that there is no reason [9:52] for it. Building and deploying AI [9:55] applications should be as simple as [9:58] building a website. There should be no [10:00] additional cost or complexity. It should [10:02] be dead simple. [10:05] And the way to make it dead simple is [10:08] following a new model for [10:10] infrastructure. what we call a [10:13] composable infrastructure model where [10:15] the complexity of the infrastructure and [10:18] the build process gets abstracted away [10:22] by putting in place a centralized team [10:25] that can build and deploy core [10:28] artifacts, your AI models, your [10:31] application code, your core [10:32] infrastructure templates, and then [10:35] rapidly [10:37] enable every single developer within [10:39] your IT organization [10:42] to be able to build and deploy an AI [10:44] application. [10:46] Just like with the adoption of cloud, IT [10:49] organizations put in place the [10:51] principles of cloudnative engineering [10:54] and made it simple for everyday [10:56] developers to be able to build and [10:58] deploy applications in the cloud. So too [11:01] do those same principles now need to [11:03] apply to AI. And here at Vulture, we are [11:06] pioneering this new model to make [11:09] certain that everyone can quickly and [11:12] easily spin up all of the core digital [11:15] services that are AI powered um here in [11:19] the future. So let's talk a little bit [11:21] about this because this is critical to [11:24] take what we call a platform engineering [11:27] approach. [11:28] While you may have decentralized [11:30] infrastructure to support your [11:33] physicians, your clinicians, to support [11:35] your patients, [11:37] you're also going to need to have some [11:38] amount of centralization so that you can [11:41] enforce the security, so you can enforce [11:43] the governance and so you can ensure the [11:45] efficiency of your operations. [11:48] And it starts first by talking about you [11:52] know cloudnative composable templates. [11:55] So for every type of application, for [11:58] every type of model you're going to [12:00] build, you have to think of it as you're [12:03] delivering a product. You have to take a [12:06] product management approach to your core [12:11] applications, [12:13] your new application for the patient so [12:16] they can interact with a digital agent [12:19] to ask questions about their latest lab [12:22] results. That would that's a type of an [12:25] application [12:26] and the core model the core application [12:30] artifacts the core infrastructure that [12:33] needs to support that should be [12:35] centrally built and governed by a team [12:40] within your organization. [12:42] So the whole notion is to pre-build, [12:46] preackage, pre-est, pre-validate [12:50] to ensure the safety and security of [12:53] what you're enabling and then to take [12:56] these product managed artifacts and then [12:59] enable every downstream developer to be [13:02] able to pick them up, leverage them, and [13:06] then build something net new. [13:09] If you pre-build a template for a [13:11] generic agent, a developer who needs to [13:15] build an agent for doctors in the [13:17] office, a developer who needs to build [13:19] an agent for a clinician in a local [13:22] healthcare facility, a developer who [13:24] wants to build a separate agent for uh [13:27] the patient. It's all the same template. [13:31] They don't need to reinvent the wheel. [13:33] They don't need to think about the [13:35] complexity in the infrastructure and [13:37] scaling the infrastructure. They don't [13:39] need to think about the security. They [13:41] don't need to think about the [13:42] compliance. It's all built into the [13:44] template. They just want to create an [13:46] agent experience. Take the pre-built [13:49] application artifacts pre take the [13:51] pre-built infrastructure templates and [13:54] simply run and run fast and innovate at [13:58] scale with safety and security. This is [14:02] what we call a composable cloud [14:04] infrastructure where you can take all of [14:07] your core uh model artifacts, [14:11] microservices, [14:13] uh you know, and containerized data and [14:16] marry them to all of your pre-built [14:19] templates for accelerated compute, [14:21] storage, and networking. And the [14:23] wonderful thing about this is this [14:26] provides the golden path. This provides [14:29] the golden path for every one of your IT [14:33] developers to be able to build and scale [14:39] all without again the risk and with [14:42] compliance all built in. This is what [14:45] we're pioneering here at Bolure working [14:48] with our customers which is how can we [14:50] help take [14:52] an approach where a deployment at one [14:55] customer can define a golden path for [14:59] other customers to follow. So how can we [15:02] share best practices across the industry [15:05] so that all of you can learn from your [15:07] peers as well as help your internal [15:10] teams learn from the experts in your [15:12] platform engineering team. And again the [15:15] goal of this and I don't want to spend [15:17] too much on this is the complexity of [15:21] spinning up the infrastructure to [15:22] support your AI applications today. The [15:26] art of the possible is pretty hard. I [15:29] was down in the exhibition hall and I [15:31] won't name the company, but right in [15:33] their booth they say GPU clusters in [15:36] hours. In hours? Really? Why isn't it [15:41] instant on? Why can't you just have a [15:43] GPU cluster right now before I finish [15:45] the sentence? And why can't it just [15:47] scale up and down without us having to [15:49] worry about it? We don't have to wait [15:51] hours to spin up a compute instance. [15:54] That was ridiculous. We don't have to [15:56] wait hours to spin up more capacity for [15:59] that compute instance. It just happens [16:01] automatically. We don't even think about [16:02] it. It just works. Well, this is what we [16:06] are pioneering here at Vulture, which is [16:08] all of the complexity of scaling up and [16:12] scaling down instantaneously. [16:15] your GPU resources anywhere in the [16:18] world, 32 data center regions around the [16:21] planet, coupled with all of your CPU [16:24] resources for delivering the actual [16:26] application experience with all of the [16:29] associated storage and all of the [16:32] networking. So you have minimal latency [16:36] and the best possible experience again [16:39] because this is about the customer [16:40] experience. This is about the doctor [16:42] experience. This is about the clinician [16:43] experience. We have to deliver a good [16:45] experience. We have to help them work [16:47] more efficiently. We have to help them [16:48] drive better outcomes. This is a whole [16:50] new stack. And it has to be as easy as [16:55] we made the core cloud native stack. And [16:59] the complexity of enabling all of our [17:03] teams needs to go away again. I don't [17:08] want to come back here next year and see [17:09] another person say it takes hours to [17:11] spin up a GPU cluster. The [17:12] state-of-the-art in the industry must be [17:14] everyone gets immediate access within 1 [17:16] second to whatever scale of GPUs they [17:20] need anywhere in the world because if we [17:22] can't democratize and make access to [17:24] GPUs easy, we're never going to be able [17:26] to accelerate personal personalized [17:28] healthcare outcomes at scale for [17:30] everyone on the planet, which I think [17:32] should be all of our goals in this [17:34] industry. [17:35] What it also means is that we need to [17:37] think very critically about a hybrid [17:38] multi cloud approach. And what I mean by [17:40] this is we're always going to have [17:42] medical devices on premises. We're going [17:44] to have medical devices in the hospital. [17:46] We're going to have medical devices in [17:47] the doctor's office. And it's going to [17:49] be capturing real-time data that needs [17:51] to be safely and securely provisioned to [17:54] the cloud for purposes of inferring an [17:58] AI model. So in my particular case, my [18:00] lab results, they're being captured at [18:03] Quest Diagnostics. [18:05] And then that data needs to make its way [18:07] from Quest Diagnosis to my to my doctor. [18:11] And then from my doctor, it needs to be [18:13] make its way to a cloud service [18:15] provider, Vulture, so that I can askly [18:18] ask an agent a question, how am I doing? [18:22] And have the agent respond, your glucose [18:25] levels are again too high because you [18:26] can't stop eating ice cream at night. [18:28] Which is a true fact. A quart a night, [18:31] it's got to stop at some point. And to [18:34] do this, this means that you need to be [18:36] able to set up private direct connects. [18:39] You need to actually ensure that that [18:41] data can be safely transmitted from the [18:43] on-prem environment to the cloud service [18:45] provider without ever being exposed over [18:47] a public cloud. You have to have the [18:49] hardened networking infrastructure to [18:51] deliver a private cloud, a hybrid cloud [18:55] to enable agentic AI at scale. And where [18:59] this comes together is how we help [19:02] people today drive industry success. [19:04] Now, unfortunately, I can't name this [19:06] one particular customer at this point in [19:07] time because we're still in the middle [19:08] of deployment, but this is exactly what [19:11] Agentic AI looks like at scale. And what [19:14] it means is all of your data that you're [19:17] collecting on prem is being safely [19:20] transmitted, streamed real time into one [19:23] of three buckets. a bucket for training [19:27] a model because you're constantly [19:29] training and retraining based on [19:32] specific outcomes. So, for example, if [19:34] you have an agent that's giving [19:36] recommendations for lifestyle changes, [19:38] giving given massive spikes in glucose [19:41] levels every single time I get a lab [19:43] test result, maybe the agent's [19:44] recommending, you know, eat more cereal [19:46] for breakfast, but that doesn't help. [19:49] So, you've got to be able to retrain the [19:51] model consistently based on what you're [19:53] observing. in terms of you know time [19:57] series data so that the model can pick [20:00] up things like ah actually eating cereal [20:02] doesn't help Kevin Cochran actually it's [20:04] cutting out the ice cream that's going [20:05] to help him right so we get better [20:07] outcomes over time so there's a training [20:09] data set that's being streamed from that [20:11] data that you're collecting but it's [20:13] also collecting data from the actual [20:14] user experience itself and then there's [20:16] the real time data like my actual lab [20:19] results that need to get fed into a [20:22] vector database to support real time rag [20:25] a real-time agent and the data that's [20:27] actually being sent [20:30] into that vector database is actually [20:33] supporting two types of experiences a [20:34] clinician experience as well as the [20:37] patient experience and the models and [20:39] the responses are going to be different. [20:43] So having the ability to quickly spin up [20:47] a template that spins up all of the [20:50] wiring to make this happen, that's what [20:53] we need. Because for every single [20:55] developer in your organization to figure [20:57] out how to spin up and wire all of the [21:00] infrastructure to make this happen, it's [21:03] only going to add risk, cost, and [21:04] complexity. It has to be super easy. And [21:07] this is what we do. Single click, spin [21:08] up all this infrastructure. It just [21:10] works. So you can focus on the model, [21:13] the training of the model, the user [21:15] experience and how the user is getting [21:19] benefit from the new services that [21:20] you're providing in all the complexity [21:22] of all the wiring that should all just [21:24] be done. It should just automatically [21:26] happen and you shouldn't worry about the [21:27] scaling. [21:29] Now in the other space um beyond just [21:33] aentic AI experience obviously in the [21:36] pure medical research firm uh you know [21:39] we have lots of examples uh I'll just [21:41] highlight one example here IPA [21:44] therapeutics using large scale GPU [21:46] clusters to do advanced uh you know [21:49] biologics uh discovery uh in the [21:51] interest of time we'll just make sure [21:53] these slides are available to you but [21:54] it's a wonderful use case and tremendous [21:56] learnings and another good example of [21:58] where we can give you a template to help [22:01] you get started and learn from the best [22:05] in the industry. So with that, I just [22:08] leave you and here at Vulture, we're [22:10] super excited about supporting uh our [22:12] clients in the healthcare space. [22:14] Healthcare is moving faster. Healthcare [22:16] is setting the pace and security uh [22:19] compliance uh means everything in the [22:22] healthcare arena. And what we want to do [22:25] is help all of you move faster uh with [22:28] less cost, less complexity and help [22:30] deliver a promise of a future where [22:32] every person on the planet can get [22:34] personalized healthcare outcomes. So [22:35] thank you.