OpenAI's In-House Chip 'Jalapeño' Outperforms Nvidia's Blackwell in Efficiency and Speed
OpenAI's in-house AI inference chip 'Jalapeño' has demonstrated faster performance while using less power than Nvidia's latest GPU, 'Blackwell.' According to performance data released by OpenAI on the 25th (local time), the chip showed superiority in both power efficiency and response speed compared to Nvidia's high-performance systems. The benchmarks were conducted using 'InferenceX,' a public inference benchmark by the semiconductor analysis firm SemiAnalysis. In tests involving three open-source models—GPT-OSS 120B, DeepSeek R1 (670B), and Moonshot AI's Kimi K2.5 (1 trillion parameters)—Jalapeño achieved 1.5 to 1.9 times higher throughput per watt and 1.7 to 3.6 times better end-to-end latency compared to Nvidia systems.
SemiAnalysis released an architecture analysis via newsletter. Jalapeño is equipped with six HBM4 stacks, providing 15.4TB/s of memory bandwidth per package, a figure that surpasses currently shipping HBM3E-based accelerators. Even though competitor chips were tested with all performance optimization techniques applied—such as Multi-Token Prediction (MTP), Speculative Decoding, and Prefill-Decode Separation (PDD)—Jalapeño achieved superior results using only Single-Token Prediction (STP).
However, there are caveats regarding the interpretation of the figures. SemiAnalysis noted that all data was provided by OpenAI; while the firm verified the measurements in its lab, it did not independently run the entire benchmark suite. The agency added that Jalapeño's actual competitor is not the already-shipping Blackwell, but the next-generation Rubin chip, which also uses HBM4. They also stated that on a Total Cost of Ownership (TCO) basis, the two chips are comparable.
OpenAI plans to begin small-scale deployment of Jalapeño within 2026 and significantly increase volumes starting in 2027. The company announced that a successor stepping (B0) with 25% improved performance is currently in fabrication, and the conceptual design for a third-generation chip is already underway.
