OpenAI Says Jalapeño Beats Nvidia Systems in Inference Tests
The custom chip delivered up to 1.9 times the peak throughput per watt and cut latency by as much as 72% in company tests.
Topics
News
- Apple Tightens Mac Access Rules as AI Agents Gain Reach
- Trump Puts Intelligence Chief Jay Clayton in Charge of AI Task Force
- IBM Widens AI, Quantum Research With IIT Bombay, IISc
- OpenAI, Synopsys Build AI Model That Could Slash Chip Design Time
- Pentagon Plans Four-Star Command for AI, Drone Warfare
- California Draws the Line on AI Firings, Layoffs and Workplace Surveillance
OpenAI said its first custom inference chip, Jalapeño, outperformed Nvidia’s GB200 and GB300 systems across several efficiency and latency measures, as the company prepares to deploy its own processors by the end of 2026.
Tests using GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T found that Jalapeño delivered 1.5 to 1.9 times as much inference throughput per watt at peak performance, according to OpenAI. End-to-end latency was about 41% to 72% lower, while performance on highly interactive workloads was 2.1 to 4.1 times as high.
For GPT-OSS 120B, Jalapeño produced about 1.9 times the peak throughput per watt of Nvidia’s GB200. It delivered about 1.7 times as much as Nvidia’s GB300 when running DeepSeek R1 and 1.5 times as much with Kimi K2.5.
The comparisons were run using InferenceX, a public benchmark from SemiAnalysis. On GPT-OSS 120B, Jalapeño produced about 1.9 times higher peak throughput per watt than Nvidia’s GB200, while on DeepSeek R1 it delivered about 1.7 times the throughput per watt of Nvidia’s GB300, according to OpenAI’s results.
The figures are based on OpenAI’s own testing and do not establish that Jalapeño is faster than Nvidia hardware across all workloads.
Jalapeño is rated at 700 watts, although OpenAI said sustained power consumption stayed at or below 550 watts during the workloads it tested.
The company designed the processor specifically for inference, optimizing the chip, memory, networking and software together to reduce the delays created when data moves across different parts of an AI system.
AI also helped build the chip. OpenAI said its models helped engineers move from initial design to tapeout in nine months, while newer models were used to optimize software for the processor.
“We plan to begin deploying Jalapeño within OpenAI’s compute infrastructure by the end of the year,” the company said.
Jalapeño is only the first step. OpenAI said its second-generation chip is already deep in development, while a third generation is taking shape. The company will, however, continue using accelerators from Nvidia and other suppliers for both training and inference.
