OpenAI Jalapeño Chip Outperforms Nvidia Blackwell in Power Efficiency Tests

OpenAI revealed its first custom inference chip, codenamed Jalapeño, on August 25, 2026, showing it outperforms Nvidia's GB200 and GB300 accelerators in power efficiency benchmarks.

The chip was developed in partnership with Broadcom and integrated with systems by Celestica. In testing using SemiAnalysis' InferenceX benchmark, the 700-watt Jalapeño chip outperformed Nvidia's Blackwell-generation hardware, which is rated between 1,200 and 1,400 watts, across models including GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T.

Key performance advantages include 1.5 to 1.9 times more AI work per watt at peak throughput, end-to-end latency reduced by 1.7 to 3.6 times, and ultra-low-latency interactive inference workloads running 2.1 to 4.1 times faster. OpenAI noted that sustained power draw remained at or below 550 watts during testing, suggesting the chip's efficiency edge may be greater than normalized figures indicate.

The Jalapeño chip is designed exclusively for inference, not model training, an area where Nvidia maintains a dominant position. OpenAI's head of hardware, Richard Ho, confirmed the company will continue to use Nvidia hardware extensively.

Industry analysts say the benchmark results represent a significant engineering milestone but do not pose an immediate threat to Nvidia's market share. Nvidia's Data Center revenue reached $75.25 billion in Q1 FY2027, supported by massive supply commitments.

The chip is scheduled for limited deployment by the end of 2026, with volume production not expected until 2027. By then, Nvidia's newer Vera Rubin architecture is expected to be in wider circulation. OpenAI is already working on second and third generations of its custom hardware.