OpenAI says its 1st custom AI chip surpasses Nvidia systems in key tests

OpenAI said Tuesday that its first custom artificial intelligence chip, Jalapeno, outperformed leading Nvidia systems in company tests, delivering greater energy efficiency and faster responses across several large language models.

The ChatGPT developer said Jalapeno produced 1.5 to 1.9 times more AI work per watt at peak throughput than the comparison systems while reducing end-to-end latency by 1.7 to 3.6 times.

For highly interactive workloads, the chip delivered 2.1 to 4.1 times higher performance, according to results published by OpenAI.

The company tested Jalapeno using InferenceX, a public benchmark developed by semiconductor research firm SemiAnalysis.

The processor was evaluated using OpenAI’s GPT-OSS 120B, DeepSeek R1 and Moonshot AI’s Kimi K2.5 models.

The comparison included Nvidia’s GB200 and GB300 systems, although Jalapeno was not tested against Nvidia’s newer Vera Rubin generation.

OpenAI said Jalapeno achieved approximately 1.9 times higher peak throughput per kilowatt than the GB200 system while running GPT-OSS 120B.

It delivered around 1.7 times higher performance per watt than the GB300 on DeepSeek R1 and 1.5 times higher performance on Kimi K2.5, the company said.

Jalapeno is rated at 700 watts, but its sustained power consumption remained at or below 550 watts during the tested workloads, according to OpenAI.

Unlike processors used to train AI models, Jalapeno was designed specifically for inference, the process through which trained models respond to prompts and perform tasks.

OpenAI said the processor’s architecture combines high throughput with low latency, potentially enabling customers to access lower-cost models or select options offering faster responses.

The chip was jointly developed with US semiconductor company Broadcom, while OpenAI’s own AI models were used to accelerate parts of its design, verification and optimization.

The company said it moved from initial design to tapeout, the final stage before manufacturing, in nine months.

OpenAI plans to begin deploying Jalapeno in its computing infrastructure by the end of 2026.

A second-generation processor is already at an advanced stage of development, while work on a third generation has begun, the company said.

Despite developing its own processor, OpenAI said it would continue deploying Nvidia accelerators and hardware supplied by other partners for both AI training and inference