electronics-journal.com
25
'26
Written on Modified on
OpenAI and Broadcom Unveil LLM-Optimized Intelligence Processor
OpenAI and Broadcom collaborate to engineer and deploy Jalapeño, a custom artificial intelligence processor optimized for large language model inference workloads.
www.broadcom.com
Photo: OpenAI
OpenAI and Broadcom have announced the development of Jalapeño, OpenAI’s first custom intelligence processor. The specialized application-specific integrated circuit (ASIC) represents the initial milestone in a multi-generation compute platform engineered to accelerate large language model (LLM) inference while enhancing system reliability and hardware accessibility.
Context of the Cooperation
As generative artificial intelligence and frontier AI models expand globally, the technology sector faces severe infrastructure bottlenecks related to computing availability, power constraints, and data movement latency. Traditional general-purpose accelerators, often adapted from older graphical or computational workloads, struggle to maintain high hardware utilization when handling the intense memory and networking demands of interactive, real-time LLM serving systems.
To scale infrastructure efficiently and establish a full-stack architecture behind its models, OpenAI required specialized manufacturing and hardware execution capabilities beyond its core software competencies. The collaboration addresses these complexities by combining OpenAI’s algorithmic understanding of LLM fundamentals, kernels, and serving patterns with Broadcom’s established silicon implementation, connectivity technologies, and networking expertise. Additionally, Celestica was brought into the ecosystem to manage complex electronic board and rack-level systems integration, ensuring the new architecture could scale reliably to industrial volumes.
Technical Solution and Responsibilities
The technological solution centers on Jalapeño, a custom-designed, blank-slate processor built specifically for modern LLM inference rather than general-purpose acceleration. Responsibilities are divided across the primary partners to align hardware development with software architecture:
OpenAI and Broadcom have announced the development of Jalapeño, OpenAI’s first custom intelligence processor. The specialized application-specific integrated circuit (ASIC) represents the initial milestone in a multi-generation compute platform engineered to accelerate large language model (LLM) inference while enhancing system reliability and hardware accessibility.
Context of the Cooperation
As generative artificial intelligence and frontier AI models expand globally, the technology sector faces severe infrastructure bottlenecks related to computing availability, power constraints, and data movement latency. Traditional general-purpose accelerators, often adapted from older graphical or computational workloads, struggle to maintain high hardware utilization when handling the intense memory and networking demands of interactive, real-time LLM serving systems.
To scale infrastructure efficiently and establish a full-stack architecture behind its models, OpenAI required specialized manufacturing and hardware execution capabilities beyond its core software competencies. The collaboration addresses these complexities by combining OpenAI’s algorithmic understanding of LLM fundamentals, kernels, and serving patterns with Broadcom’s established silicon implementation, connectivity technologies, and networking expertise. Additionally, Celestica was brought into the ecosystem to manage complex electronic board and rack-level systems integration, ensuring the new architecture could scale reliably to industrial volumes.
Technical Solution and Responsibilities
The technological solution centers on Jalapeño, a custom-designed, blank-slate processor built specifically for modern LLM inference rather than general-purpose acceleration. Responsibilities are divided across the primary partners to align hardware development with software architecture:
- OpenAI is responsible for the foundational architecture design, optimizing the chip from scratch around its internal roadmap of frontier models, kernels, and agentic product workloads. OpenAI also utilizes its own AI models within the engineering workflow to automate and accelerate parts of the physical chip design and optimization loop.
- Broadcom is responsible for the custom silicon implementation, physical fabrication tape-out management, and integrating its high-performance connectivity and networking silicon, including Tomahawk networking chips.
- Celestica provides the engineering support for board design, rack system integration, high-performance networking topography, and scalable production deployment systems.
At a system level, Jalapeño functions by balancing compute, memory, and networking resources directly inside the processor architecture to minimize data movement. This structural optimization ensures realized hardware utilization tracks close to the system's theoretical peak performance limits. According to Richard Ho, leader of OpenAI’s hardware program, the platform is engineered around the exact memory movement and serving patterns that matter most for interactive LLM products, combining high throughput with minimized latency. The hardware is designed with architectural flexibility to process diverse LLMs across the industry, and early testing indicates the platform delivers performance-per-watt efficiency parameters substantially ahead of current state-of-the-art accelerators.
Deployment or Implementation
The custom development program achieved a nine-month timeline from initial design concept to manufacturing tape-out. Engineering samples of the Jalapeño silicon are currently executing machine learning workloads—including GPT-5.3-Codex-Spark—under laboratory testing conditions at production-targeted frequency and power parameters.
The implementation roadmap positions Jalapeño as the first phase of a multi-generation compute platform. Initial physical deployment of the integrated accelerator racks is scheduled to begin by the end of 2026. Broadcom and OpenAI are co-developing this roadmap to enable the scalable deployment of gigawatt-scale data centers in coordination with Microsoft and additional infrastructure partners.
Edited by Romila DSilva, Induportals Editor, with AI assistance.
Deployment or Implementation
The custom development program achieved a nine-month timeline from initial design concept to manufacturing tape-out. Engineering samples of the Jalapeño silicon are currently executing machine learning workloads—including GPT-5.3-Codex-Spark—under laboratory testing conditions at production-targeted frequency and power parameters.
The implementation roadmap positions Jalapeño as the first phase of a multi-generation compute platform. Initial physical deployment of the integrated accelerator racks is scheduled to begin by the end of 2026. Broadcom and OpenAI are co-developing this roadmap to enable the scalable deployment of gigawatt-scale data centers in coordination with Microsoft and additional infrastructure partners.
Edited by Romila DSilva, Induportals Editor, with AI assistance.
