OpenAI's Custom Chip Outperforms Nvidia in AI Model Tests

OpenAI has announced that its custom-designed chip, Jalapeño, outperforms chips from Nvidia and others when running DeepSeek R1, a massive AI model with over 1 trillion parameters. The company says it plans to have a limited number of Jalapeño-based systems in place this year, with greater capacity next year. However, OpenAI's VP of hardware, Richard Ho, noted that the chip is not designed for training new models, meaning the company remains heavily reliant on Nvidia and others for compute needs.

OpenAI's competitors are also developing their own custom chip designs, including Google's TPU line and Amazon and Microsoft's AI chip efforts. Anthropic has also announced plans to build its own in-house chip design effort. The development of these chips is significant as it could reduce the reliance on external vendors for compute power.

OpenAI currently uses a combination of Jalapeño, Cerberus, Nvidia, AMD, and other chips for its compute needs. However, Ho stated that the company has no plans to offer Jalapeño to others due to its own high demand for it.