On August 27, 2026, OpenAI officially launched its proprietary inference chip, Jalapeño, at the Hot Chips 37 conference. Test results show that this chip achieved material-level efficiency and latency improvements compared to NVIDIA GB200/GB300 systems in real-world workloads: peak throughput increased by 1.5 to 1.9 times per watt of workload, end-to-end latency decreased by 1.7 to 3.6 times, and performance for high-interaction workloads improved by 2.1 to 4.1 times. The chip has a rated power consumption of 700W, with actual testing showing it remained below 550W. OpenAI plans to deploy Jalapeño on its own infrastructure by the end of the year. The Gen 2 version is already in advanced development, while Gen 3 is still under research. Additionally, through optimization of the underlying kernel using GPT-Astra and Codex, the operation speed of some attention mechanisms and MoE blocks is 1.5 to 1.8 times faster than existing expert code.
At the 37th Hot Chips conference, OpenAI announced its proprietary inference chip, Jalapeño, claiming that it achieves material-level efficiency and reduced latency in real-world workloads compared to NVIDIA GB200/GB300 systems. Test data showed that Jalapeño achieved 1.5–1.9 times more work per watt, 1.7–3.6 times lower end-to-end latency, and 2.1–4.1 times better performance under high-interaction workloads; the chip’s rated power consumption was 700W, and it remained below 550W during testing. OpenAI plans to deploy it in its own infrastructure by the end of the year. Gen 2 has been thoroughly developed, while Gen 3 is still in development. Additionally, GPT-Astra + Codex helps optimize the underlying kernel, enabling some attention and MoE blocks to run 1.5–1.8 times faster than existing expert-written code.