OpenAI details performance benchmarks for new custom artificial intelligence chip

OpenAI has released new performance details regarding its custom processor designed for artificial intelligence workloads. The chip, called Jalapeño, was developed in partnership with Broadcom and is classified as an Application-Specific Integrated Circuit — a specialized semiconductor built for a single, specific purpose.

Performance and Benchmarks

In benchmark testing conducted using the InferenceX platform, Jalapeño was compared against competing hardware systems including Nvidia's GB200 and GB300 superchips. According to OpenAI hardware vice president Richard Ho, the chip is designed to address a common industry compromise where systems must choose between minimizing delay or maximizing total output volume.

OpenAI stated that Jalapeño delivered between 1.5 and 1.9 times more work per watt across three tested artificial intelligence models. These models include GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T. Additionally, the processor achieved between 1.7 and 3.6 times lower end-to-end latency across those same models, representing a reduction in the time required to generate a complete response.

Deployment Plans

The processor was first introduced in June. OpenAI plans to begin small-volume deployments of the hardware by the end of the year, followed by a gradual increase in production volume throughout 2027.

Despite the introduction of its own hardware, the organization does not intend to completely replace its existing infrastructure. Company representatives noted that partnerships with external hardware providers will remain a component of its broader infrastructure strategy, while work on second and third generations of the processor is already underway.