OpenAI and Broadcom Revolutionize AI with Jalapeño Chip for LLMs

OpenAI and Broadcom Unveil Jalapeño, a Major Advancement for AI

In a strategic collaboration with Broadcom, OpenAI has introduced Jalapeño, an artificial intelligence chip specifically designed for the inference of large language models (LLMs). This processor, developed in a record time of nine months, promises to surpass current accelerators in terms of energy efficiency.

OpenAI, already recognized for its cutting-edge AI models, is now expanding its focus to the hardware that supports them. Jalapeño has been designed to meet the specific requirements of LLM inference, aiming to enhance performance, latency, and energy efficiency for services utilizing models such as ChatGPT, Codex, and future OpenAI agents.

An Optimized Architecture for Maximum Efficiency

The design of Jalapeño is based on a deep understanding of the constraints of language models. OpenAI has optimized the chip's architecture to improve interactions between computing units, memory, and the network, thereby minimizing data transfers that often lead to high energy consumption in AI infrastructures. The initial prototypes of this chip are already capable of handling workloads similar to those encountered in production, particularly with the GPT-5.3-Codex-Spark model.

Broadcom brings its expertise in silicon implementation and networking technologies, such as Tomahawk. Celestica is involved in the manufacturing of electronic boards, racks, and system integration. This division of roles allows OpenAI to maintain control over design while relying on industrial partners for rapid scaling.

Jalapeño: A Strategy to Control Inference Costs

With Jalapeño, OpenAI aims to control the entire technology chain, from chips to deployment systems, including execution software and user-facing products. This strategy seeks to manage inference costs, which represent an increasing share of operational expenses as the use of generative AI expands in businesses.

A better-suited chip reduces electricity consumption and increases the number of requests processed by the same infrastructure. According to OpenAI, initial tests indicate performance per watt superior to current benchmark accelerators, although detailed results are expected later. If these promises materialize, Jalapeño could enhance the profitability of AI services and provide better service quality to businesses using OpenAI's APIs or ChatGPT.

Prospects for Businesses with Jalapeño

Jalapeño is also the foundation of a computing platform expected to evolve over several years. OpenAI and Broadcom plan a gradual rollout starting at the end of 2026 with partners operating gigawatt-scale data centers.

One of the most remarkable aspects of the project is the speed of its development. The complete cycle from design to manufacturing of Jalapeño took only nine months. OpenAI utilized its own AI models to accelerate certain phases of design and hardware optimization, illustrating a virtuous circle where AI contributes to building the infrastructure necessary for its own operation.

For the market, this announcement shows that OpenAI is joining the ranks of companies investing in proprietary accelerators to better control their costs, performance, and technological independence. The company seeks to optimize the most critical inference loads with hardware specifically designed for its needs while continuing to use GPUs from suppliers like Nvidia.

This strategy could have an impact far beyond OpenAI's ecosystem. Lower computing costs, better energy efficiency, and more powerful infrastructures could enable faster, more reliable, and more affordable AI services for developers, SMEs, and large enterprises alike. This would also enhance OpenAI's competitiveness in the market.