OpenAI Developed Custom Jalapeño AI Inference ASIC

The company partnered with Broadcom to build proprietary silicon for internal model acceleration.

Updated on Sept. 28, 2026 in Semiconductors

Bold vector editorial illustration featuring a detailed silicon wafer and data center cooling components in teal and mustard tones. — OpenAI has developed its own custom Jalapeño AI inference ASIC to optimize compute efficiency and protect proprietary model research within its internal data centers. AI Illustration. Upload story photo >

Live Poll

Do you believe proprietary chips developed by AI companies will lead to better consumer technology?

In August 2026, OpenAI revealed it developed the Jalapeño application-specific integrated circuit (ASIC) to optimize internal AI inference workloads. The hardware is currently in use within internal data centers to run proprietary models.

Why it matters

The shift toward vertical hardware integration allows OpenAI to maintain proprietary research IP and improve system efficiency. This strategy marks a strategic departure from relying exclusively on third-party silicon merchants for large-scale compute.

OpenAI utilized the InferenceX benchmark platform to evaluate the Jalapeño chip against Nvidia GB200 and GB300 accelerators. Engineers demonstrated the ability to deploy complex models including GPT-OSS and DeepSeek R1 onto A0 sample chips within two months.

The players

OpenAI

An AI research organization focused on the development of large-scale generative models and proprietary hardware infrastructure.

Broadcom

A semiconductor design firm that provides custom ASIC development services for hyperscale data center operators.

Nvidia

A leading designer of graphics processing units and AI accelerators, including the GB200 and GB300 series, which dominate current inference benchmarks.

The details

The Jalapeño ASIC relies on hardware-software codesign, where the chip architecture is optimized specifically for the requirements of internal AI models. By managing the hardware stack directly, OpenAI avoids exposing proprietary model research to third-party silicon manufacturers. The deployment process for A0 samples—the initial physical hardware prototypes—was validated by successfully migrating models such as Kimi K2.5 onto the silicon in under 60 days.

OpenAI and Broadcom revealed their hardware partnership in 2025.
OpenAI presented Jalapeño benchmarks at Hot Chips 2026 in August 2026.
Details regarding the hardware were confirmed in an interview published September 28, 2026.

The Tech Race

This development positions OpenAI to challenge the dominance of Nvidia and its upcoming Vera Rubin platforms in the inference market. The company is betting that custom silicon will prove more efficient for its specific model architectures than general-purpose commercial accelerators.

This hardware is designed strictly for internal OpenAI workloads and is not currently available for commercial purchase or external integration. Developers should monitor future OpenAI infrastructure reports to see if these efficiency gains manifest as changes to API pricing or model availability.

The takeaway

The move suggests a long-term transition for major AI labs to become integrated device manufacturers to maintain control over compute costs. Watch for future performance disclosures from OpenAI or competing benchmarks from Nvidia to determine the real-world efficiency of this custom hardware.