OpenAI said on Tuesday that its first custom artificial intelligence chip, Jalapeno, delivered faster responses and greater energy efficiency than leading Nvidia systems in company tests.
The ChatGPT developer said Jalapeno produced 1.5 to 1.9 times more AI processing work per watt at peak throughput, while reducing end-to-end latency by between 1.7 and 3.6 times.
For highly interactive AI workloads, OpenAI said the processor achieved 2.1 to 4.1 times higher performance than the systems used for comparison.
The tests were conducted using InferenceX, a public benchmark developed by semiconductor research firm SemiAnalysis, and involved OpenAI’s GPT-OSS 120B, DeepSeek R1 and Moonshot AI’s Kimi K2.5 models.
Jalapeno was compared with Nvidia’s GB200 and GB300 systems, although the tests did not include Nvidia’s newer Vera Rubin generation.
According to OpenAI, Jalapeno achieved around 1.9 times greater peak throughput per kilowatt than the GB200 when running GPT-OSS 120B. It recorded approximately 1.7 times better performance per watt than the GB300 on DeepSeek R1 and 1.5 times better performance on Kimi K2.5.
The processor is rated at 700 watts, although OpenAI said its sustained power consumption remained at or below 550 watts during the workloads tested.
Unlike chips primarily designed for training AI models, Jalapeno was developed specifically for inference — the process by which trained AI models respond to prompts and carry out tasks. OpenAI said its combination of high throughput and low latency could help reduce the cost of running models while enabling faster responses.
The chip was developed jointly with US semiconductor company Broadcom, with OpenAI also using its own AI models to accelerate parts of the design, verification and optimisation process.
OpenAI said the project progressed from its initial design to tapeout, the final design stage before manufacturing, in nine months.
The company plans to begin deploying Jalapeno across its computing infrastructure by the end of 2026.