高通将模块化AI栈扩展至骁龙平台
高通将其Modular AI开发栈的支持范围扩展至面向骁龙平台优化的AI开发，并与微软合作以支持在Windows上进行原生开发。Chris Lattner代表高通参与此事。该消息涉及高通与微软两家公司，属于软件与开发工具层面的进展，未提及具体元器件型号、产能、价格或交期信息。
原文深度解读
本文报道高通在2026年9月23日于毛伊岛举行的骁龙峰会上宣布，将其收购的Modular公司的统一AI软件栈扩展至骁龙等产品组合，并与微软合作在Windows上原生支持Mojo语言。涉及主体包括高通、Modular、微软，关键人物为高通高级AI软件与平台执行副总裁Chris Lattner和微软Windows平台与开发者执行副总裁Logan Iyer。关键数字包括2026年9月23日、Mojo 1.0于8月发布。与电子元器件供应链的关系在于：该软件栈可针对CPU、GPU、NPU及定制ASIC进行优化，降低对特定硬件的依赖，可能影响AI芯片及终端硬件的需求结构和生态选择，但原文未提及具体元器件型号、产能、价格或交期。
- 高通在2026年9月23日毛伊岛骁龙峰会上宣布，Modular的统一软件栈正扩展至包括骁龙在内的高通产品组合，Chris Lattner称开发者将能用单一软件栈构建此前难以实现的AI体验。
- Modular由包括Lattner在内的工程师创立，Lattner曾任谷歌TensorFlow团队高级工程师；高通在6月表示已达成收购Modular的协议。
- Modular的统一AI开发软件栈由Mojo、MAX和Modular Cloud组成；Mojo于8月发布1.0并开源，MAX是实时运行AI模型的 serving 框架，Modular Cloud用于自动化配置、调度和扩展。
- 用Mojo编写并通过MAX运行的AI模型可针对CPU、GPU、NPU和定制ASIC优化，开发者无需为每种芯片重写代码，从而降低对特定硬件的依赖，并支持内存管理和并行处理分发用户请求。
- 高通在收购前已与Modular合作，将Modular栈应用于高通Dragonfly系列数据中心加速器，并展示在高通NPU平台上可提供行业领先性能。
- 高通与微软合作扩展Windows OS生态，微软计划与高通Modular团队合作在Windows上原生支持Mojo，使其无需虚拟化即可安装、用于系统级编程并直接在OS层面运行；由于Mojo是开放语言，在Windows PC上开发的软件不绑定特定硬件厂商或OS。
供应链影响
- 由于该软件栈可针对CPU、GPU、NPU和定制ASIC优化并降低对特定硬件的依赖，AI硬件平台的选择可能更加灵活，从而可能影响不同芯片架构在终端和数据中心中的采用节奏。
- 高通将Modular栈扩展至骁龙产品组合，可能增强骁龙平台对AI开发者的吸引力，进而可能影响采用骁龙平台的终端设备及相关元器件的需求结构，但具体影响取决于终端厂商的设计决策。
- 微软在Windows上原生支持Mojo，可能扩大基于Windows PC的AI开发与部署场景，进而可能影响PC端AI加速硬件及相关元器件的需求，但取决于开发者采用速度和OEM产品规划。
- 高通此前已将Modular栈用于Dragonfly数据中心加速器并展示NPU性能，这可能提升其数据中心AI加速产品的生态竞争力，但相关供应链影响取决于客户导入和量产进度。
- 原文未提及具体元器件型号、产能、价格或交期信息，因此对电子元器件供应链的短期直接影响尚不明确，需取决于后续产品化与生态落地情况。
系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。
Chris Lattner，高通高级AI软件与平台执行副总裁，宣布扩展Modular栈对骁龙的支持。（照片：THE ELEC，毛伊岛）
Modular正在扩展对针对高通骁龙平台优化的AI开发的支持。该公司还在与微软合作，以支持在Windows上进行原生开发。
Chris Lattner，高通高级AI软件与平台执行副总裁，于当地时间9月23日在毛伊岛举行的骁龙峰会2026上表示，Modular的统一软件栈正在扩展至高通的产品组合，包括骁龙。他说：“世界各地的开发者将能够构建此前使用单一软件栈难以实现的创新AI体验。”
Modular是一家集成计算平台公司，由帮助构建当今大部分AI基础设施的工程师创立，其中包括Lattner，他曾担任谷歌TensorFlow团队的高级工程师。高通在6月表示，已达成收购Modular的协议。
Modular提供一个开放的AI软件平台，允许AI工作负载无论硬件架构如何都能高效开发和运行。开发者此前不得不为不同类型的硬件使用类似但独立的软件栈，每个软件栈都有自己的缺陷。
Modular的统一AI开发软件栈由Mojo、MAX和Modular Cloud组成。Mojo是一种用于AI硬件的高性能编程语言，结合了Python的开发者生产力与C++和CUDA相关的速度和细粒度硬件控制。Modular在8月发布了Mojo 1.0，使其达到生产就绪状态，并将该语言开源。
MAX是用于实时运行AI模型的 serving 框架。（照片：THE ELEC，毛伊岛）
用Mojo编写并通过MAX运行的AI模型可以针对中央处理器（CPU）、图形处理器（GPU）、神经处理单元（NPU）和定制专用集成电路（ASIC）进行优化。
与传统方法不同，开发者无需为每种芯片重写代码。该框架降低了对特定硬件的依赖，并允许随着市场条件变化扩展硬件平台。它还支持内存管理以及通过并行处理分发用户请求，以保持系统不间断运行。
Modular Cloud是一个云软件栈，旨在让服务器持续运行，而无需开发者管理复杂的云服务器命令。它自动化包括配置、调度和扩展在内的任务。
配置为AI工作负载准备和分配所需的大规模计算资源。调度为AI工作负载分配优先级和处理顺序，而扩展在流量或计算需求急剧上升时动态添加资源。该平台还管理硬件延迟和故障，帮助服务器以最佳性能和总拥有成本（TCO）持续运行。
高通在收购前一直与Modular合作。“我们将Modular栈应用于高通Dragonfly系列数据中心加速器，”Lattner说。“我们展示了可以在高通的NPU平台上提供行业领先的性能。”
高通还在活动上公布了通过与微软合作扩展Windows操作系统（OS）生态的计划。微软Windows平台与开发者执行副总裁Logan Iyer在视频演示中表示，微软正在与高通的Modular团队合作，以在Windows上正式原生支持Mojo。这意味着Mojo可以安装、用于系统级编程，并直接在OS层面运行，而无需虚拟化。
微软计划与高通的Modular团队合作，在Windows上提供一流的Mojo开发体验。由于Mojo是一种开放语言，在Windows PC上开发的软件不绑定特定硬件厂商或OS。
“这对骁龙生态和Windows社区来说都是巨大的进步，”Lattner说。
Chris Lattner, Qualcomm executive vice president of Advanced AI Software and Platforms, announces expanded Snapdragon support for the Modular stack. (Photo: THE ELEC, Maui)
Modular is expanding support for AI development optimized for Qualcomm's Snapdragon platforms. The company is also working with Microsoft to enable native development on Windows.
Chris Lattner, Qualcomm executive vice president of Advanced AI Software and Platforms, said at Snapdragon Summit 2026 in Maui on Sept. 23 local time that Modular's unified software stack is being expanded across Qualcomm's product portfolio, including Snapdragon. "Developers around the world will be able to build innovative AI experiences that were previously difficult to implement using a single software stack," he said.
Modular is an integrated computing platform company founded by engineers who helped build much of today's AI infrastructure, including Lattner, who previously served as a senior engineer on Google's TensorFlow team. Qualcomm said in June that it had reached an agreement to acquire Modular.
Modular provides an open AI software platform that allows AI workloads to be developed and run efficiently regardless of hardware architecture. Developers previously had to use similar but separate software stacks for different types of hardware, each with its own bugs.
Modular's unified AI development software stack consists of Mojo, MAX and Modular Cloud. Mojo is a high-performance programming language for AI hardware that combines Python's developer productivity with the speed and fine-grained hardware control associated with C++ and CUDA. Modular released Mojo 1.0 in August, bringing it to production-ready status, and made the language open source.
MAX is a serving framework for running AI models in real time. (Photo: THE ELEC, Maui)
AI models written in Mojo and run through MAX can be optimized for central processing units (CPUs), graphics processing units (GPUs), neural processing units (NPUs) and custom application-specific integrated circuits (ASICs).
Unlike conventional approaches, developers do not need to rewrite code for each chip. The framework reduces dependence on specific hardware and allows hardware platforms to be expanded as market conditions change. It also supports memory management and distribution of user requests through parallel processing to keep systems running without interruption.
Modular Cloud is a cloud software stack designed to keep servers operating continuously without requiring developers to manage complex cloud-server commands. It automates tasks including provisioning, scheduling and scaling.
Provisioning prepares and allocates large-scale computing resources required for AI workloads. Scheduling assigns priorities and processing order for AI workloads, while scaling dynamically adds resources when traffic or computational demand rises sharply. The platform also manages hardware latency and failures, helping servers operate continuously at optimal performance and total cost of ownership (TCO).
Qualcomm had been working with Modular before the acquisition. "We applied the Modular stack to Qualcomm's Dragonfly family of data-center accelerators," Lattner said. "We demonstrated that we can deliver industry-leading performance on Qualcomm's NPU platform."
Qualcomm also unveiled plans at the event to expand the Windows operating system (OS) ecosystem through collaboration with Microsoft. Logan Iyer, Microsoft executive vice president of Windows Platform + Developer, said in a video presentation that Microsoft is working with Qualcomm's Modular team to officially support Mojo natively on Windows. That means Mojo can be installed, used for systems-level programming and run directly at the OS level without virtualization.
Microsoft plans to work with Qualcomm's Modular team to provide a first-class Mojo development experience on Windows. Because Mojo is an open language, software developed on Windows PCs is not tied to a particular hardware vendor or OS.
"This is tremendous progress for both the Snapdragon ecosystem and the Windows community," Lattner said.