OpenAI says its Jalapeno AI chip delivers faster responses than rivals like Nvidia

OpenAI has unveiled the performance of its custom Jalapeño AI chip, which it developed in collaboration with Broadcom. According to benchmark results, the chip significantly outperforms rival chips in terms of speed and efficiency. When measuring AI work per unit of power, Jalapeño delivered 1.5 to 1.9 times more processing, and also exhibited 1.7 to 3.6 times lower latency, leading to faster response times.

The advantage becomes even more pronounced in highly interactive tasks, reaching up to 4.1 times better performance. Unlike most chips that prioritize a trade-off between speed and efficiency, Jalapeño manages to maintain both by minimizing data movement within the system during processing. OpenAI has previously relied on Nvidia GPUs, making this the company's first significant step towards building hardware in-house.

Interestingly, OpenAI's AI models were instrumental in designing certain parts of the chip itself, allowing a small team to bring three new AI models to full performance within just two months. Despite the promising results, OpenAI isn't giving up on Nvidia just yet, as Jalapeño is only focused on handling inference, not the computational-heavy process of training new models.

The company plans to debut the chip in small volumes by the end of 2026 and scale up production in 2027, with subsequent generations already in development. OpenAI's move to develop its own AI chip joins a growing list of tech companies that are also investing in custom silicon, including Google, Amazon, and Anthropic. Moreover, Samsung reportedly utilized Anthropic's Claude to expedite its own chip design efforts, suggesting that AI-assisted hardware development is becoming increasingly common in the industry.