OpenAI’s ‘Jalapeño’ Chip Surpasses Nvidia Blackwell in Crucial AI Benchmarks
OpenAI has recently announced a significant advancement in the field of artificial intelligence hardware. Its first in-house inference chip, known as Jalapeño, has surpassed Nvidia’s leading Blackwell systems in several key benchmarks. This is an impressive achievement for a first-generation custom chip.
The Jalapeño chip, developed in collaboration with semiconductor giant Broadcom, was unveiled at the esteemed Hot Chips 2026 conference on August 25. During the event, OpenAI released its initial public performance data. Based on OpenAI’s benchmarks, Jalapeño delivers 1.5x to 1.9x more AI throughput than Nvidia Blackwell-generation systems, with end-to-end latency reduced by 1.7x to 3.6x. Each Jalapeño rack comprises 128 accelerators, 1.7 exaFLOPS of 4-bit compute, and 27.5 TB of HBM4 memory.
SemiAnalysis, an independent testing firm, ran its InferenceX benchmark suite alongside OpenAI engineers. They described the result as “huge news”, highlighting that it is rare for a first-generation custom chip to outperform Nvidia’s best in any category. Richard Ho, OpenAI’s VP of Hardware, stated that the chip “achieves high throughput and low latency simultaneously, a first in the industry.”
However, analysts caution that Jalapeño is exclusively designed for inference—the real-time execution of AI models—not training, where Nvidia continues to dominate. Volume production is not anticipated until 2027, suggesting that Nvidia’s forthcoming Vera Rubin platform may close the gap before Jalapeño is produced on a large scale.
Nevertheless, this announcement has caused a stir in the markets, coming just a day before Nvidia’s scheduled earnings report. It indicates a broader shift: the world’s largest AI companies are no longer satisfied with depending solely on Nvidia for their computational needs.
Source: CNBC – OpenAI’s Jalapeño AI chip brings new ‘threat’ to Nvidia margins