OpenAI 与 Broadcom 合作宣布了 Jalapeño，这是其第一款专为大型语言模型 (LLM) 推理加速器设计的客制化特殊应用积体电路 (ASIC)。该晶片突显了快速的开发周期，在 OpenAI 自身模型的局部协助下，仅用 9 个月（9 months）就完成了从设计到制造下片（tape-out）的流程。两家公司计划在 2026 年底前（by the end of 2026）开始于内部部署这些晶片及相关的伺服器系统，以建立多代、吉瓦规模（gigawatt-scale）的人工智慧基础设施，而非将其零售。
在 2025 年宣布的部署中，包含了庞大的 10 吉瓦（10 gigawatts）客制化 AI 加速器，并透过 Broadcom 的乙太网路（Ethernet）、PCIe 以及光学连接解决方案（包括 Tomahawk 网路晶片）进行整合。借由使用标准化的乙太网路架构，OpenAI 避免了完全依赖如 Nvidia 的 NVLink 或 InfiniBand 等专有解决方案。为了完成此机架级（rack-scale）的整合，Celestica 负责处理电路板（board）、机架（rack）和系统的协调，旨在优化每瓦性能（performance per watt），并缩小实际利用率与理论峰值性能之间的差距。
此项计划使 OpenAI 在客制化晶片趋势中，与 Google (TPU)、Amazon (Trainium/Inferentia)、Meta (MTIA) 以及 Microsoft (Maia) 等超大规模云端业者（hyperscalers）并列。Broadcom 执行长 Hock Tan 指出，该晶片的性能可与 Nvidia 的 Blackwell 和 Google 的 TPU 相比，尽管独立的基准测试和技术指标——例如记忆体频宽、制程节点和热设计功耗 (TDP)——仍未被公开。最终，这项合作伙伴关系有助于将推理成本内部化、减轻对 Nvidia 的单一来源依赖，并确立 Broadcom 作为关键机架级平台提供商的地位。
OpenAI, in collaboration with Broadcom, has announced Jalapeño, its first custom application-specific integrated circuit (ASIC) designed as a large language model (LLM) inference accelerator. Highlighting rapid development cycles, the chip progressed from design to manufacturing tape-out in just 9 months, partially assisted by OpenAI's own models. The companies plan to begin deploying these chips and associated server systems internally by the end of 2026, establishing a multi-generational, gigawatt-scale artificial intelligence infrastructure rather than retailing them.
The deployment announced in 2025 encompasses a massive 10 gigawatts of custom AI accelerators integrated via Broadcom’s Ethernet, PCIe, and optical connectivity solutions, including Tomahawk networking silicon. By utilizing standardized Ethernet architecture, OpenAI avoids sole reliance on proprietary solutions like Nvidia's NVLink or InfiniBand. To complete this rack-scale integration, Celestica handles the board, rack, and system coordination, aiming to optimize performance per watt and narrow the gap between real utilization and theoretical peak performance.
This initiative positions OpenAI alongside hyperscalers like Google (TPU), Amazon (Trainium/Inferentia), Meta (MTIA), and Microsoft (Maia) in the custom silicon trend. Broadcom CEO Hock Tan indicated that the chip's performance is comparable to Nvidia's Blackwell and Google's TPU, although independent benchmarks and technical metrics—such as memory bandwidth, process node, and thermal design power (TDP)—remain undisclosed. Ultimately, this partnership serves to internalize inference costs, mitigate single-source dependency on Nvidia, and establish Broadcom as a key rack-scale platform provider.