According to OpenAI, the Jalapeno chip surpassed the GB300 in both power and speed.

      OpenAI has announced that its Jalapeno inference chip, created in collaboration with Broadcom, has surpassed Nvidia’s GB300 in terms of AI tasks performed per unit of power and speed of responding. It has not been evaluated against Nvidia’s latest Vera Rubin hardware and is not intended for model training.

      The company has released performance metrics for the chip, which aims to reduce reliance on Nvidia. Jalapeno reportedly outperformed Nvidia’s GB300 based on two criteria: the volume of AI tasks processed per unit of power and the response time efficiency.

      OpenAI did provide its own disclaimers. Jalapeno has not been tested against Vera Rubin, Nvidia's recently launched hardware, and it is incapable of training models.

      The Next Web pointed out a key issue when the chip was introduced. In June, we noted that vendor benchmarks should be approached with skepticism until independent evaluations are available, and these figures are still from OpenAI.

      The power consumption statistic carries significant importance. Jalapeno achieves its performance at 700 watts, a crucial factor since electricity represents a major operational cost in data centers.

      Chip head Richard Ho mentioned that Jalapeno performs two functions typically divided between different components. It offers high throughput for servicing many customers at a low cost and provides low latency for those prioritizing quick response times.

      It's also important to highlight the models used in testing. OpenAI evaluated a small open model it developed alongside third-party models from DeepSeek and Moonshot, with the most significant advantage seen in Kimi, the largest model tested.

      The development timeline is already in progress. A second version is set to be completed in the coming months, and a third-generation concept is also in the works.

      Ho was cautious about his comments. “Nvidia is a strong partner, and we still depend heavily on Nvidia,” he stated.

      Other major buyers are pursuing similar paths. Anthropic is working on its own silicon design, while Google has engaged Marvell for customized inference chips.

      In Europe, there is a push towards procuring merchant products. The EU has allocated around €20 billion for AI gigafactories that will source hardware from the open market, shortly after Nvidia announced that server prices would increase by over 15%.

      However, Europe’s solution is smaller in scale compared to the challenge. Axelera, the continent's most notable AI chip company, backed by Samsung, is focusing on accelerators for edge devices rather than the data centers that the gigafactory funding aims to establish.

      Published August 25, 2026 - 3:11 pm UTC

      Back to top