OpenAI与Broadcom发布首款定制芯片"Jalapeño"，专为LLM推理设计
阅读原文· the-decoder.comOpenAI与Broadcom联合发布首款定制芯片"Jalapeño"，专为大语言模型推理而设计，从零打造而非修改通用芯片。OpenAI称其性能功耗比“显著优于”当前最先进硬件，但为自报数据，尚未独立验证。芯片开发耗时9个月，OpenAI自身模型辅助加速设计。工程样片已运行包括GPT-5.3-Codex-Spark在内的ML工作负载。计划于2026年底进行大规模部署，微软预计将购买其中40%的芯片。
OpenAI and Broadcom unveil "Jalapeño," a custom chip built for LLM inference
Key Points
- OpenAI and Broadcom have unveiled "Jalapeño," a custom chip built specifically for large language model inference.
- OpenAI says the architecture delivers better performance per watt. Development took just nine months, with OpenAI's own models helping speed up the process.
- Custom hardware is meant to make running AI models cheaper and more reliable. Large-scale deployment is planned for late 2026, with Microsoft expected to buy 40 percent of the chips.
OpenAI is adding custom hardware to its tech stack. The "Jalapeño" chip, developed with Broadcom, is tailored for large language model inference and is set to run at scale by late 2026.
According to a joint announcement, OpenAI and Broadcom have unveiled "Jalapeño" - OpenAI's first so-called "Intelligence Processor." It's a custom accelerator built specifically for large language model inference, and the first chip in a multi-generation platform the two companies are building together.
Broadcom CEO Hock Tan and President Charlie Kawwas handed the first wafer to OpenAI CEO Sam Altman and President Greg Brockman. For OpenAI, this marks its first step into custom hardware after years of focusing on models and products.
OpenAI says Jalapeño isn't a modified general-purpose chip. It was designed from scratch for modern LLM inference. OpenAI handles the chip design, Broadcom contributes silicon manufacturing and networking technology including its Tomahawk networking chips, and Celestica takes care of boards, racks, and system integration.
Performance claims lack independent verification
Early tests showed performance per watt that's "substantially better" than current state-of-the-art hardware, according to OpenAI. These are self-reported numbers that haven't been finalized. Take them with a grain of salt. A technical report is supposed to follow. Right now, it's unclear which chips Jalapeño was tested against, on what tasks, and under what conditions.
The architecture reportedly cuts data movement and pushes utilization closer to its theoretical max. Engineering samples are already running ML workloads in the lab, including the GPT-5.3-Codex-Spark model. That model currently runs on Cerebras hardware, which also specializes in inference.
OpenAI says the process from design to tape-out took just nine months, what the company calls the fastest ASIC development cycle for high-performance semiconductors it's aware of. OpenAI's own models helped speed up parts of the design process. The rumors about chip plans, though, have been circulating since 2023.
The announcement reflects OpenAI's argument that controlling the full stack from chip to product lets it run models faster, more reliably, and at lower cost. Broadcom CEO Tan says the first deployment is planned for late 2026 at gigawatt scale, together with Microsoft and other partners. Broadcom has reportedly demanded that Microsoft guarantee it will buy 40 percent of the chips to secure the first phase.