OpenAI Unveils Its Jalapeno Debut Chip
Credit: OpenAI
OpenAI, in partnership with Broadcom, announced the release of its own specialized chip, Jalapeño. This new product promises to halve the computational costs of advanced language models and change the balance of power in the AI hardware market.
In October, OpenAI announced a partnership with Broadcom to develop accelerators optimized for running its AI models. According to OpenAI and Broadcom, the new chips were designed from the ground up in record time and are capable of delivering performance per watt that “significantly exceeds current inference standards.” Final versions of the chips will be deployed in large data centers at Microsoft and other OpenAI partners beginning later this year.
These chips are designed to improve the performance of AI models by reducing the amount of data transferred. They are designed to utilize the compute, memory, and networking resources critical for high-end AI models. The next version of Jalapeno is scheduled for release in 2028. The partners do not rule out the possibility of future chips designed for other workloads.
Tang expects OpenAI and Broadcom to exceed his previous forecast of deploying 1.3 GW of AI chips next year. “We want to believe we can achieve better results because demand is very high,” he said.
OpenAI’s head of hardware, Richard Ho, declined to disclose the payment mechanism with Broadcom, noting that the financial agreements will be finalized once the chip order is fully completed. Tan also declined to comment, but confirmed that Broadcom has created a funding mechanism for chip development in partnership with investment firms Apollo Global Management and Blackstone.
For regular users, the new processor could mean significantly faster ChatGPT responses, lower API costs, more stable service performance during peak load periods, and the ability to run increasingly complex models without a dramatic increase in costs.
