# OpenAI development of Jalapeño AI inference chip

> Live situation record from CLSTR: https://clstr.news/situations/openai-development-of-jalapeo-ai-inference-chip
> Updated: 2026-08-27T15:53:07.000Z. Sources: 69. Developments: 2.

OpenAI has unveiled performance data for its custom AI inference chip, Jalapeño, developed in collaboration with Broadcom and Celestica. Designed specifically to handle large language model (LLM) workloads rather than model training, the hardware aims to lower computing costs and reduce the company’s dependency on Nvidia hardware.

Benchmark testing conducted via SemiAnalysis’ InferenceX tool indicates that the chip outperforms Nvidia’s GB200 and GB300 systems. Specifically, Jalapeño reportedly delivers between 1.5 and 1.9 times more AI work per watt and achieves 1.7 to 3.6 times lower latency. For interactive AI agent workloads, performance gains may reach up to 4.1 times.

Technical specifications note the chip utilizes a single die based on TSMC’s 3nm process, featuring 216 GiB of HBM4 memory and 13.4 PFLOPs of compute. While the chip is rated at 700 watts, testing showed actual power draw remained at or below 550 watts.

OpenAI utilized its own AI models to assist in the design and circuit verification process, which shortened the development cycle from concept to tapeout to nine months. Small-scale deployment of the chip is expected to begin by the end of 2026, with a full production ramp anticipated between 2027 and 2028.

## Claims

- OpenAI's Jalapeno chip is designed specifically for AI inference tasks rather than model training. (corroborated by 6 sources)
- Jalapeno produced between 1.5 and 1.9 times more AI work per watt at peak throughput compared to Nvidia systems. (corroborated by 4 sources)
- Jalapeno demonstrated latency cuts between 1.7 and 3.6 times compared to Nvidia GB200 and GB300 systems. (corroborated by 3 sources)
- The Jalapeno chip was developed in collaboration with Broadcom and Celestica. (corroborated by 3 sources)
- OpenAI used its own AI models to assist in the design of the Jalapeno chip. (corroborated by 3 sources)
- The design-to-tapeout process for the chip took nine months. (corroborated by 2 sources)
- The Jalapeno chip is rated at 700 watts, though measured power draw during testing stayed at or below 550 watts. (corroborated by 2 sources)

## Timeline

### 2026-08-27: OpenAI unveils Jalapeno custom AI chip to rival Nvidia

OpenAI has introduced Jalapeno, a custom AI inference chip developed with Broadcom that reportedly outperforms Nvidia's GB300 in energy efficiency and latency. Deployment is expected by late 2026.

8 sources. https://clstr.news/cluster/openai-unveils-jalapeno-ai-chip-to-compete-with-nvidia

### 2026-08-25: OpenAI unveils Jalapeño, custom AI inference chip

OpenAI has unveiled Jalapeño, a custom AI inference chip co-developed with Broadcom that claims to outperform Nvidia systems in power efficiency and latency for large language model workloads.

62 sources. https://clstr.news/cluster/openai-and-broadcom-unveil-jalapeo-custom-ai-inference-chip

---
Cite as: OpenAI development of Jalapeño AI inference chip. CLSTR, https://clstr.news/situations/openai-development-of-jalapeo-ai-inference-chip