OpenAI 与 Broadcom 合作开发了一款名为 Jalapeno 的客制化人工智慧晶片。在最近的基准测试中，该晶片在能源效率（单位功率能处理的 AI 工作量）与回应速度（低延迟）两项指标上，均击败了辉达（Nvidia）目前的 GB300 处理器。

Jalapeno 晶片专为 AI 模型的推论阶段所设计，能在 700 瓦的低电压下展现强大效能。OpenAI 计划在今年稍晚开始使用这款新晶片来支援其模型，期望在同时提供高吞吐量与低延迟的优势下，大幅降低资料中心的基础设施与电力成本。

尽管 OpenAI 已经在使用自家 AI 模型加速开发第二代与第三代晶片，以进一步降低成本，但由于整体的运算需求依然庞大，该公司强调辉达仍是非常重要的合作伙伴，未来一段时间内仍将持续依赖外部供应商的技术。

OpenAI, in partnership with Broadcom, has developed a custom artificial intelligence chip named Jalapeno. In recent benchmark testing, the chip outperformed Nvidia's current GB300 processor in two key metrics: energy efficiency (the amount of AI work handled per unit of power) and response speed (low latency).

The Jalapeno chip is specifically designed for the inference phase of AI models, demonstrating strong performance at a low voltage of 700 watts. OpenAI plans to start deploying this new chip later this year to support its models, aiming to significantly reduce data center infrastructure and power costs while offering both high throughput and low latency.

Although OpenAI is already using its own AI models to accelerate the development of second and third-generation chips to further reduce costs, their overall need for computing power remains massive. The company emphasized that Nvidia is still a crucial partner and that they will continue to rely on external suppliers' technology for the foreseeable future.