OpenAI Launches Jalapeño AI Chip

According to OpenAI, the company has unveiled Jalapeño, its first custom artificial intelligence (AI) accelerator developed in collaboration with Broadcom, marking a significant step in its strategy to expand AI infrastructure and reduce reliance on third-party processors.

As demand for advanced AI systems continues to surge, leading AI research organizations, including OpenAI and Anthropic, are seeking greater access to computing resources required to power next-generation chatbots and coding assistants. To address rising infrastructure costs and secure long-term compute capacity, OpenAI has increasingly invested in developing proprietary AI hardware as an alternative to GPUs supplied by NVIDIA.

The Jalapeño processor was jointly designed by OpenAI engineers and Broadcom specifically for AI inference workloads, enabling large-scale AI models to process data and generate responses for applications such as ChatGPT. According to Broadcom President and CEO Hock Tan, the chip delivers performance comparable to NVIDIA’s Blackwell architecture and the Tensor Processing Units (TPUs) developed by Alphabet.

OpenAI Hardware Lead Richard Ho stated that Jalapeño was engineered to efficiently accelerate large language models (LLMs), the foundation of modern generative AI applications. The processor is expected to support future generations of OpenAI’s AI models.

The company plans to begin deploying Jalapeño-powered systems before the end of the year, making it the first product in a broader multi-generation chip development roadmap. Server platforms based on the processor will be built by Celestica and will be used exclusively within OpenAI’s infrastructure.

OpenAI confirmed that engineering samples have already been tested internally using its GPT-5.3-Codex-Spark AI model. Initial results reportedly met the company’s targets for both power efficiency and computing performance.

The chip development cycle was completed in approximately nine months before being sent to TSMC for manufacturing. OpenAI noted that AI-assisted design techniques helped accelerate portions of the development process, contributing to the rapid turnaround.

The move reflects a broader industry trend toward custom AI silicon. Major technology companies, including Meta, Amazon, and Google, have also partnered with chip design specialists such as Broadcom and Marvell Technology to develop application-specific AI processors that offer greater control over performance, cost, and scalability.

Reports earlier this year also indicated that Anthropic is evaluating the development of its own AI chips, highlighting the growing importance of proprietary semiconductor solutions across the AI industry.

Despite strong demand, Broadcom noted that custom AI accelerators currently face profitability challenges due to rapidly increasing memory requirements. Hock Tan explained that advanced AI processors require substantial amounts of High Bandwidth Memory (HBM), which raises system costs and places pressure on margins. Broadcom sources HBM components from leading memory manufacturers including SK hynix and Samsung Electronics.

As AI workloads become increasingly compute- and memory-intensive, custom AI chips such as Jalapeño are expected to play a growing role in optimizing performance, improving energy efficiency, and supporting the next phase of large-scale AI deployment.