OpenAI has developed its first custom AI inference chip, internally named "Jalapeño," and the company reportedly claims it outperforms comparable Nvidia hardware in inference benchmarks.
What Happened
According to reporting by The Decoder, OpenAI's inaugural custom silicon—referred to internally as "Jalapeño"—has been tested against Nvidia's Blackwell and Rubin GPU architectures in inference benchmark evaluations. The report indicates that OpenAI's chip demonstrated performance advantages over both comparable Nvidia products during these benchmarks.
Why It Matters
OpenAI has historically relied on Nvidia GPUs for training and running its AI models, making this custom silicon development a notable shift in the company's hardware strategy. If validated through independent testing, a competitive inference chip could reduce OpenAI's dependence on a single supplier and potentially lower operational costs at scale. The move also positions OpenAI among a growing number of AI developers—including Google, Amazon, and Meta—that are pursuing custom silicon to optimize their specific workloads.
The Bottom Line
OpenAI has entered the custom AI chip market with its "Jalapeño" design, which reportedly outperforms Nvidia's Blackwell and Rubin in inference benchmarks. Independent verification of these claims remains limited, and broader availability or deployment details have not been disclosed.