OpenAI claims its Jalapeño chip delivers faster responses than competing chips

Summary

OpenAI claims that the Jalapeño ASIC, manufactured by Broadcom, outperforms Nvidia's GB200/GB300 chips in InferenceX benchmarks. The company reports 1.5–1.9x higher workload per watt and 1.7–3.6x lower latency across three different models. Small-scale deployment is expected to begin by the end of the year, with capacity increases targeted toward 2027.

Why it matters

Inference cost and latency are among the biggest bottlenecks in scaling AI agents and consumer-facing products. Therefore, a reliable efficiency gain could fundamentally change the operating economics of frontier models. This move also signals OpenAI's intent to reduce its reliance on Nvidia. It adds a new variable to an already volatile AI chip supply chain, where Microsoft, Meta, and Google are also competing.

Highlights

• The benchmarks are self-reported values by OpenAI on its own chosen platform, without independent third-party verification.
• One of the three tested models is OpenAI's own GPT-OSS 120B model. This raises the possibility of hardware-specific optimization.
• Deployment is described only as 'small volumes.' Details such as chip count, data center location, or customer commitments are not disclosed.

Critical take

The headline numbers look impressive at first glance. However, it is evident that OpenAI selected the benchmark, the comparison set, and one of the three tested models—its own model. Nvidia's GB200/GB300 results are defined only as the 'best values recorded at that time.' This is a fairly favorable lower bound. Until an independent laboratory repeats the InferenceX test with the same workloads, this reads less like a verified engineering triumph and more like a well-crafted marketing bulletin timed to the chip race news cycle.