OpenAI Unveils Jalapeño, Its First Custom AI Chip for ChatGPT Infrastructure.

OpenAI announced on LinkedIn that it has designed and manufactured its first custom AI chip, codenamed Jalapeño. Developed in partnership with Broadcom, the chip is purpose-built for the large language model workloads that power ChatGPT, Codex, the OpenAI API, and what the company describes as future agentic products.

The move positions OpenAI alongside hyperscalers like Google and Amazon in building custom silicon for AI inference and training. According to the company, Jalapeño was designed from the ground up by OpenAI's internal team and brought to production with Broadcom's manufacturing capabilities.

OpenAI frames the chip as foundational to the AI economy and critical for scaling intelligence. The company said building its own silicon expands its full-stack platform from products to models to infrastructure, enabling it to serve more users and expand access to AI.

The announcement comes as OpenAI continues to scale its infrastructure to support 900 million ChatGPT users and growing demand for API access. Custom chips allow companies to optimize for specific workloads, reduce reliance on third-party GPU suppliers like Nvidia, and potentially lower long-term operational costs. Google has shipped multiple generations of its Tensor Processing Units (TPUs), while Amazon Web Services offers Trainium and Inferentia chips for machine learning workloads.

OpenAI did not disclose technical specifications for Jalapeño, such as transistor count, memory bandwidth, or performance benchmarks. The company also did not share timelines for when the chip will enter production at scale or whether it will be used exclusively for inference, training, or both. Broadcom, a major player in custom chip design and manufacturing, has previously partnered with Google on TPU development.

The announcement specifically names ChatGPT, Codex, and the OpenAI API as workloads Jalapeño will support. Codex, OpenAI's code-generation model, has seen significant expansion in recent months, including agent capabilities for autonomous workflows and Sites functionality for building full-stack apps.

OpenAI's reference to future agentic products suggests the chip is designed to handle more complex, multi-step reasoning tasks that require sustained compute over longer sessions. Agentic AI, which can autonomously plan and execute tasks, typically demands different infrastructure characteristics than single-turn chat completions, including lower latency for real-time decision-making and higher memory bandwidth for managing extended context windows.

The company has not indicated whether Jalapeño will be used in its recently launched GPT-5.5 models or specialized variants like GPT-Rosalind for drug discovery. However, the emphasis on LLM workloads suggests the chip is optimized for transformer-based architectures that underpin OpenAI's model family.

OpenAI's entry into custom silicon reflects broader industry trends. As AI workloads mature, companies with sufficient scale are designing chips tailored to their specific model architectures and deployment patterns. This vertical integration can improve performance per watt, reduce costs at scale, and provide more control over the supply chain.

For developers building on the OpenAI API, custom chips could translate to lower pricing, faster response times, or expanded rate limits as infrastructure costs decline. OpenAI recently introduced banked rate limit resets for Codex users, signaling ongoing infrastructure improvements.

The announcement did not address whether OpenAI plans to offer Jalapeño to external customers or keep it for internal use. The company also did not comment on its continued reliance on Nvidia GPUs or how Jalapeño fits into its existing data center footprint. OpenAI operates a mix of owned infrastructure and cloud partnerships with Microsoft Azure, which provides significant compute resources under a long-term agreement.

Broadcom's involvement suggests OpenAI is taking an ASIC (application-specific integrated circuit) approach rather than designing a fully custom chip from transistor to package. This model allows faster time to market and leverages Broadcom's expertise in high-performance networking and chip design.