OpenAI says its Jalapeño chip beats Nvidia Blackwell in AI inference tests
OpenAI’s custom Jalapeño AI chip has outperformed Nvidia’s Blackwell-based GB300 in key inference benchmarks, according to results released by the company, as the ChatGPT maker works to reduce its reliance on Nvidia for running AI models.
The results, presented at the Hot Chips conference at Stanford University on Tuesday, showed Jalapeño delivering better performance per watt and faster response times in tests conducted using SemiAnalysis’ InferenceX benchmark.
OpenAI Jalapeño chip beats Nvidia Blackwell
Jalapeño is designed for AI inference, the stage at which trained models process user requests, rather than for training AI models.
OpenAI’s head of hardware Richard Ho said the chip was built to handle both high-volume workloads and applications requiring faster responses. It was tested with OpenAI’s smaller open-source model as well as models from DeepSeek and Moonshot AI.
The chip also operates at a lower power level of 700 watts, which Ho said could help reduce data centre costs.
“It’s a really good chip — it should drop it by a lot,” Ho told Bloomberg, referring to potential cost savings.
Jalapeño could challenge Nvidia in AI inference
SemiAnalysis, which tested the chip at OpenAI’s laboratories, found Jalapeño beat Blackwell on performance per watt in most scenarios.
However, the research firm said the comparison was not entirely like-for-like because Jalapeño uses newer HBM4 memory. Nvidia’s Vera Rubin platform also uses HBM4 and could provide a closer comparison.
“Jalapeño is really competing against chips like Rubin that also use HBM4,” CNBC quotes SemiAnalysis analysts .
“Vera Rubin systems are starting to ship to customers right now, while it will still be some time before OpenAI has anything beyond engineering samples of Jalapeño,” they added.
OpenAI targets lower AI computing costs
OpenAI unveiled Jalapeño in June with semiconductor company Broadcom. Early samples indicated potential cost savings of about 50% compared with conventional AI GPUs, the company said.
The chip was developed in roughly nine months before being sent to Taiwan Semiconductor Manufacturing Company for production. OpenAI plans limited deployment towards the end of 2026, with a broader rollout expected in 2027.
A second-generation Jalapeño is already in development.
Nvidia remains key to OpenAI
OpenAI’s move reflects a wider push by major technology companies to develop their own AI chips. Google, Meta and Amazon are also investing heavily in custom semiconductors for AI workloads.
Still, OpenAI says it will continue relying on Nvidia, particularly for demanding AI training workloads.
“Nvidia is a really good partner, and we continue to need a lot of Nvidia,” Ho said.
Analysts expect Nvidia’s GPUs to remain important because of their performance, flexibility and CUDA software ecosystem. Jalapeño, meanwhile, gives OpenAI greater control over the cost and efficiency of running AI models at scale.
