OpenAI's Jalapeño chip beats Nvidia on speed and efficiency tests

OpenAI introduced its Jalapeño chip this week as a dedicated processor designed to accelerate AI inference workloads—the computational tasks required to run trained models and deploy intelligent agents. Built in partnership with Broadcom, the custom silicon aims to solve a fundamental engineering challenge: most AI systems must choose between fast response times or high processing throughput, but rarely excel at both.

Performance benchmarks#

Testing through the InferenceX platform revealed substantial improvements over contemporary alternatives. When measured against Nvidia's GB200 and GB300 superchips, Jalapeño demonstrated:

• 1.5 to 1.9× more AI work per watt across three major language models (GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T)
• 1.7 to 3.6× lower latency in end-to-end response times using the same models

According to OpenAI hardware vice president Richard Ho, this combination delivers what the company describes as the "best of both worlds"—faster user-facing responses alongside more efficient power consumption.

Deployment timeline and strategy#

OpenAI intends to begin deploying Jalapeño in limited quantities by year's end, with plans to scale production significantly throughout 2027. The company has not disclosed specific unit targets for next year. Ho emphasized that Jalapeño will supplement, rather than replace, OpenAI's existing chip partnerships. The firm remains committed to working with established partners like Nvidia while continuing development work on second and third-generation iterations of the processor.

First revealed in June, Jalapeño marks OpenAI's expansion into custom silicon as a means of optimizing performance for its inference-heavy workloads.