- Published on
DeepSeek R1 AI Throughput Per Watt is 1.7x That of GB300
- Authors
- Name
- DeepHorus
On August 25, OpenAI released performance benchmark results for its first self-developed AI inference chip, Jalapeño. The tests showed that the chip outperforms the current market-leading NVIDIA systems in key energy efficiency metrics, marking a significant step for OpenAI in custom computing hardware.
The tests were announced in an official blog post by OpenAI, using the public benchmark tool InferenceX developed by semiconductor research company SemiAnalysis. The subjects tested were three mainstream large language models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, with NVIDIA's GB200 and GB300 as reference systems for comparison.
Specific data indicates that Jalapeño's per-watt AI throughput performance is outstanding. When running the GPT-OSS 120B model, its per-kilowatt peak throughput was approximately 1.9 times higher than the GB200 system. When running the DeepSeek R1 model, its per-watt performance was about 1.7 times higher than the GB300; for Kimi K2.5, this advantage was approximately 1.5 times. Additionally, its end-to-end latency was only 28% to 59% of the comparison systems.
This chip was led in development by Richard Ho, OpenAI's hardware lead, who previously worked on Google's Cloud TPU project for nearly nine years. Jalapeño is an ASIC (Application-Specific Integrated Circuit) co-developed by OpenAI and Broadcom, manufactured using TSMC's 3nm process, and first unveiled in June 2026. The chip adopts a systolic array architecture and is equipped with high-bandwidth memory, optimized for large language model inference, and does not support model training.
Notably, OpenAI claims that Jalapeño took only about 9 months from architectural design to tape-out completion, stating this is one of the fastest ASIC development cycles in the history of high-performance advanced semiconductors. During this process, OpenAI's own AI models were deeply involved in the chip design workflow, helping to accelerate design optimization.
The deployment plan for this chip has been clarified. OpenAI plans to begin small-scale deployment of Jalapeño by the end of 2026, gradually expanding the deployment scale by 2027. The chip and its accompanying server systems are intended solely for OpenAI's internal use and will not be sold externally, aiming to reduce the company's own GPU procurement and computing operation costs.
Currently, the second-generation Jalapeño chip has entered the later stages of development and is expected to complete tape-out within the next few months; the third-generation chip has also initiated conceptual design. However, the comparison targets for this test were NVIDIA's GB200 and GB300 systems, and no comparison has yet been made with NVIDIA's latest generation Vera Rubin chip.
OpenAI emphasized in the blog post that NVIDIA remains an important partner, and the company will continue to adopt a multi-supplier strategy to meet its vast computing needs. This demonstration of the self-developed chip's performance highlights the trend of top AI companies seeking greater autonomy and control in computing infrastructure.
