Bristol Myers Squibb buys Nvidia AI system for drug discovery 百时美施贵宝收购英伟达AI系统用于药物研发

TL;DR

• Bristol Myers Squibb becomes the first life sciences company to deploy an Nvidia DGX SuperPOD based on the Vera Rubin architecture.
• The new infrastructure consists of eight DGX Vera Rubin NVL72 systems, significantly expanding BMS's AI computing capacity beyond its existing legacy clusters.
• BMS integrates the new system with its current infrastructure to support proprietary model training for drug discovery, including target identification and compound prediction.
• The deployment enables the "Predict First" methodology, allowing researchers to computationally screen dozens of candidates instead of just ten, accelerating early-stage development.
• Access to Nvidia’s BioNeMo Agent Toolkit and natural-language interfaces democratizes AI usage across global research sites, reducing reliance on specialized computational expertise.

Why It Matters

This acquisition marks a significant milestone in the convergence of high-performance computing and pharmaceutical R&D, demonstrating how next-generation AI hardware is becoming essential for scaling drug discovery pipelines. For the broader industry, it highlights the shift from experimental AI adoption to integrated, enterprise-wide infrastructure that directly impacts operational efficiency and candidate selection rates. It also underscores the growing necessity for life sciences companies to invest in cutting-edge silicon to maintain competitive advantages in speed and accuracy.

Technical Details

• Hardware Architecture: The cluster comprises eight DGX Vera Rubin NVL72 systems, integrating Nvidia Vera CPUs and Rubin GPUs, representing a generational leap over BMS's previous two-to-three-generation-old SuperPOD.
• Software Ecosystem: Utilization of Nvidia Mission Control for cluster provisioning, monitoring, and workload management, alongside the BioNeMo Agent Toolkit for protein structure prediction, molecular generation, and docking.
• Integration Strategy: The new Vera Rubin infrastructure is combined with existing DGX SuperPOD systems into a unified computing environment, managed by a software stack that schedules training and prediction workloads across both legacy and new hardware.
• User Interface Enhancements: Implementation of natural-language instruction capabilities to lower the barrier to entry for non-specialist scientists, enabling direct access to complex computing tasks without deep coding expertise.
• Workload Types: Primary focus on training proprietary foundation models, running large-scale predictions for small and large molecules, and executing multi-parameter optimization for compound synthesis prioritization.

Industry Insight

Pharmaceutical companies must prioritize scalable, next-generation AI infrastructure to handle the increasing computational demands of foundation models and large-molecule predictions. The integration of user-friendly interfaces like natural-language processing is critical for democratizing AI tools across diverse research teams, thereby maximizing ROI on expensive hardware investments. Furthermore, the ability to unify legacy and new systems suggests a hybrid approach may be necessary during transition periods, allowing firms to extend the utility of existing assets while adopting breakthrough technologies.

TL;DR

• 百时美施贵宝（BMS）成为首家采购基于Nvidia Vera Rubin架构的DGX SuperPOD的生命科学企业，以支持其全球药物研发工作。
• 新集群由8套DGX Vera Rubin NVL72系统组成，结合Vera CPU和Rubin GPU，旨在解决现有基础设施容量饱和及计算需求激增的问题。
• BMS采用“预测优先”策略，利用AI进行靶点识别、分子生成和多参数优化，将早期候选药物评估数量从十多个提升至数十个。
• 新系统将整合Nvidia BioNeMo Agent Toolkit，并通过自然语言界面降低使用门槛，实现全球研究站点的数据与模型共享。
• AI应用已使临床前候选药物的发现和制备时间缩短20%-30%，预计未来可达50%，显著加速药物研发流程。

为什么值得看

本文展示了顶级制药巨头如何大规模部署最新一代AI算力基础设施，标志着AI在生命科学领域的渗透已从实验性探索进入核心研发流程的基础设施建设阶段。对于AI从业者和行业观察者而言，它揭示了垂直领域对高性能计算的具体需求场景（如蛋白质结构预测、分子对接），以及软硬件协同（Nvidia硬件+BioNeMo软件栈）在解决复杂科学问题中的实际价值。

技术解析

• 硬件架构升级：BMS部署的DGX SuperPOD基于Nvidia最新的Vera Rubin架构，包含8套NVL72机架式系统。每套系统整合了Nvidia Vera中央处理器和Rubin图形处理器，相比其现有的两三代前的旧SuperPOD，提供了巨大的算力跃升，专门用于训练专有模型和处理大规模化合物、蛋白质数据。
• 软件工具链集成：新环境集成了Nvidia BioNeMo Agent Toolkit，提供蛋白质结构预测、分子生成、分子对接、序列分析和基因组学等工具。该系统能够连接不同的计算工具，形成统一的工作流，并支持通过Nvidia Mission Control进行集群配置、监控和工作负载调度。
• “预测优先”工作流：BMS实施“Predict First”方法，利用模型生成的预测结果在合成前排除不符合多参数优化要求的分子。这种方法通过计算筛选缩小实验室测试范围，确保宝贵的实验资源集中在成功率最高的分子上，从而优化研发效率。
• 人机协作与自然语言接口：尽管引入先进AI，BMS强调人类研究人员仍负责审查模型输出并决定推进哪些项目。同时，公司引入了自然语言指令界面，允许非计算专家的研究人员直接发起复杂的预测请求，降低了使用门槛。
• 全球分布式计算环境：新基础设施将与现有系统合并，形成一个共享的计算环境，允许全球不同站点（如新泽西州劳伦斯维尔和加州圣地亚哥）的研究团队共享数据集和模型输出，打破数据孤岛。

行业启示

• 算力即研发能力：在药物发现领域，高端AI算力已成为核心竞争力。制药公司正从单纯购买算力转向构建集成的软硬件生态，以支持从靶点识别到分子优化的全链条AI工作流，这要求供应商提供更垂直化的解决方案。
• AI重塑研发管线效率：AI不仅加速单个环节，更通过“预测优先”策略改变了整个研发漏斗的结构。通过大幅减少进入实验室阶段的候选药物数量，企业能够显著降低试错成本并缩短上市时间，预计未来几年可节省高达50%的前期开发时间。
• 民主化AI访问权限：通过自然语言接口和统一的云平台，大型制药公司正在努力消除计算科学与生物学之间的技能壁垒。让一线生物学家直接访问强大的AI工具，有助于释放组织内部的创新潜力，加速科学发现的迭代速度。