OpenAI Jalapeño: A New Benchmark in AI Chip Performance
OpenAI has unveiled its new dedicated artificial intelligence chip, codenamed Jalapeño. According to a recent blog post and a briefing with reporters, this chip demonstrates significantly more efficient task completion and faster response generation compared to existing AI systems.
Jalapeño’s Innovative Performance Metrics
Richard Ho, OpenAI’s hardware vice president, stated that Jalapeño offers the “best of both worlds” by combining lower latency with higher throughput. This combination is crucial for AI systems that demand rapid data processing and response delivery.
Benchmarking conducted on Semianalysis’s InferenceX platform showcased Jalapeño’s superior capabilities. The chip registered:
- More tokens per user.
- Higher throughput per kilowatt.
These results position Jalapeño ahead of currently available state-of-the-art solutions, marking a significant advancement in AI hardware development.
Strategic Impact and Future Outlook
The development of Jalapeño involved a collaboration between OpenAI and Broadcom, with manufacturing utilizing TSMC’s N3P process node. This specialized ASIC (Application-Specific Integrated Circuit) is exclusively optimized for inference tasks of large-scale AI models, including GPT-4.5 and GPT-5. While the tech industry was focused on the acquisition of Nvidia Blackwell GPUs, Sam Altman’s team was secretly engineering this proprietary solution, poised to potentially redefine the inference chip market landscape.
The introduction of OpenAI’s Jalapeño chip, leveraging TSMC’s N3P process and a Broadcom collaboration, represents a pivotal shift in AI inference hardware. Its focus on optimized throughput per kilowatt and reduced latency for models like GPT-4.5/5 directly addresses the escalating operational costs and power consumption inherent in large-scale AI deployment. This proprietary ASIC could significantly disrupt the market dominance of general-purpose GPUs in inference workloads.