Nvidia CEO Jensen Huang Responds to OpenAI's Custom 'Jalapeño' AI Chip
During Nvidia's Q2 earnings call, CEO Jensen Huang responded to OpenAI's custom AI inference chip, Jalapeño, by stating that Nvidia is building a complete "AI factory platform" rather than a single-purpose ASIC. This comes after reports showed OpenAI's Jalapeño chip outperformed Nvidia's GB300 in power efficiency and response speed for specific workloads like DeepSeek R1. The development highlights the growing competition between general-purpose AI platforms and custom ASICs designed by hyperscalers to reduce reliance on Nvidia and lower operational costs. If custom silicon like Jalapeño continues to outperform GPUs in inference efficiency, it could threaten Nvidia's high profit margins in the AI hardware market. OpenAI's Jalapeño chip, co-developed with Broadcom, reportedly achieved 1.7 times the AI throughput per watt of Nvidia's Blackwell-class GB300 processor when running DeepSeek R1. Huang argued that Nvidia's platform remains superior because it can run any model across any cloud environment, offering flexibility that custom ASICs lack.
## BACKGROUND
AI inference is the process of running trained machine learning models to make predictions or generate responses, which requires massive computational power. While general-purpose GPUs like Nvidia's Blackwell series (including the GB300) are highly versatile, Application-Specific Integrated Circuits (ASICs) are custom-built to execute specific algorithms with maximum efficiency.