OpenAI to Deploy In-House Chip This Year... "Estimated to Feature Samsung HBM4"

OpenAI announced plans to deploy its custom-developed AI chip, "Jalapeno," to power its AI models within this year.

In an interview with Bloomberg on the 25th local time, Richard Ho, OpenAI's head of hardware, stated that Jalapeno outperformed Nvidia's top-tier "GB300" in two key metrics: AI workload processing capacity per watt and response speed.

Ho noted that Jalapeno will significantly lower OpenAI's operational costs.

Jalapeno was developed in collaboration with Broadcom, a custom chip supplier.

The two companies, which announced their partnership last year, unveiled Jalapeno in June, emphasizing that it was completed at an unusually rapid pace.

While typical chips excel in either processing power or response speed, Jalapeno combines both strengths.

OpenAI will decide which AI models to run on Jalapeno, allowing customers to choose between cost reduction and performance enhancement depending on their needs.

Ho added that Jalapeno delivers high performance at a low power consumption of 700 watts, which will also help cut electricity costs—a major expense for data centers.

The IT-focused media outlet SemiAnalysis reported that Jalapeno is equipped with HBM4 memory delivering a bandwidth of 15.4TB/s per package, and estimated the supplier to be Samsung Electronics.

OpenAI stated on its company blog that it has publicly tested the new chip with its own small open-source models as well as third-party models from DeepSeek and Moonshot AI, noting that it showed an even greater performance advantage with Moonshot's "Kimi" model.

However, Jalapeno has not undergone comparative testing with Nvidia's next-generation "Vera Rubin" chip, which has just begun shipping.

It was also not designed for training AI models—an area where Nvidia excels—and is instead specialized for the "inference" stage, where pre-trained models respond to prompts and process tasks.

Ho mentioned that OpenAI will not immediately replace Cerebras chip technology currently used for relatively small models, stating, "Our computing demand is so massive that we have contracted with many suppliers, and this situation will continue for the time being."

At the same time, he added, "Nvidia is a really great partner, and we will continue to need a lot of Nvidia."

Ho also shared that OpenAI is utilizing its own AI models to accelerate chip development, noting that progress on the second-generation chip is already well underway, with "tape-out"—the final design stage—expected to be completed within the next few months.

OpenAI has also initiated conceptual design for a third-generation chip to further reduce AI infrastructure costs.
 
(Photo: Captured from OpenAI website, Yonhap News)
※ Please note: This article was translated by AI and may contain errors.