OpenAI is expanding beyond frontier models and into silicon design after it partnered with Broadcom to develop a custom inference processor, marking one of its biggest infrastructure moves yet. The companies introduced Jalapeño, calling it the first AI accelerator within a multi-generation compute platform and underscoring OpenAI’s ambition to control more of the technology stack powering future artificial intelligence services. Rather than relying exclusively on third-party accelerators, the company is positioning proprietary hardware as a strategic lever to improve inference efficiency while supporting the massive compute requirements that the AI industry will face over the next decade. The announcement also reflects a broader industry shift as leading AI developers increasingly invest in vertically integrated infrastructure to optimize performance, reduce costs and scale deployments.
Jalapeño Designed Around Large Language Model Inference
OpenAI architected Jalapeño using insights from developing and operating frontier large language models, while Broadcom transformed those specifications into manufacturable silicon. Celestica joined the collaboration as the board, rack and system integration partner, creating an end-to-end hardware platform instead of delivering a standalone processor. The partners say they engineered the accelerator specifically for inference workloads, where responsiveness, networking efficiency and power consumption increasingly determine the economics of large AI deployments. Consequently, the project represents more than another AI chip launch, it highlights an effort to redesign infrastructure around the operational demands of production-scale language models.
The company also expects Jalapeño to deliver higher performance per watt than current alternatives, an increasingly important metric as hyperscale operators prepare facilities that will consume hundreds of megawatts and eventually gigawatts of electricity. Higher inference efficiency could lower operating costs while allowing AI providers to serve more requests within the same power envelope. “The world is moving to a compute-powered economy,” Greg Brockman, President and Co-Founder of OpenAI, said in a press release. “Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems. By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access.”
Broadcom Eyes Multi-Generation AI Silicon Roadmap
For Broadcom, the partnership extends its strategy of co-developing custom AI silicon with hyperscale customers instead of competing directly in the merchant accelerator market. The company said Jalapeño represents only the first step in a broader hardware roadmap that will support next-generation AI infrastructure across multiple deployment cycles. AI companies continue driving demand for custom accelerators as they seek alternatives that better match their proprietary software stacks and operational requirements. Broadcom combines its manufacturing expertise with OpenAI’s model-level optimization to shorten development cycles and improve hardware utilization.
“Our collaboration with OpenAI represents a fundamental commitment to scaling the physical infrastructure required for the next decade of AI,” Hock Tan, President and CEO of Broadcom, added. “This is just the beginning of a multi-generation roadmap. By co-developing our industry-leading silicon directly with OpenAI, we are enabling the deployment of gigawatt scale data centers with Microsoft and other partners beginning in 2026.” The companies also revealed that Jalapeño progressed from initial design to manufacturing tape-out in just nine months. According to OpenAI, its own AI models assisted portions of the engineering workflow, accelerating design optimization tasks that traditionally require significant manual effort. The milestone illustrates how engineers increasingly apply AI to semiconductor development, creating a feedback loop where advanced models help build the hardware that will run future generations of those same models. Faster design cycles could become a competitive advantage as AI infrastructure evolves at unprecedented speed.
Custom Silicon Supports OpenAI’s Full-Stack Strategy
OpenAI’s hardware leadership emphasized that Jalapeño was engineered around the practical behavior of frontier AI workloads instead of adapting existing processor architectures. The company optimized memory movement, networking characteristics, serving patterns and computational kernels to maximize inference throughput while approaching the hardware’s theoretical performance limits. Those architectural decisions reflect OpenAI’s growing emphasis on building infrastructure tailored specifically to its production environment rather than depending entirely on generalized accelerators.
“Jalapeño was designed from the ground up for LLM inference using detailed insights from our close collaboration with OpenAI researchers,” said Richard Ho, who leads OpenAI’s hardware program. “We optimized the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models. Based on early testing, Jalapeño will efficiently execute our most important workloads close to the hardware’s theoretical limits.” The processor is expected to accelerate inference-intensive applications across OpenAI’s ecosystem, including faster ChatGPT responses and more efficient execution of Codex workloads. Those gains become strategically important as AI adoption shifts from experimentation toward production environments that demand predictable latency, lower operating costs and sustained scalability. Custom silicon also gives OpenAI greater control over hardware-software optimization, reducing dependence on external product roadmaps while strengthening its long-term infrastructure strategy.
Gigawatt AI Data Centers Become the Next Competitive Battleground
OpenAI and Broadcom plan to begin deploying Jalapeño inside gigawatt-scale data centers toward the end of 2026 before expanding availability over subsequent years. The announcement aligns with the industry’s rapid transition toward increasingly power-intensive AI campuses designed to support frontier inference at unprecedented scale. If deployment proceeds as planned, Jalapeño will become one of several custom AI accelerators entering a market historically dominated by NVIDIA’s platforms. Whether the processor ultimately challenges NVIDIA’s GB300 NVL72 platform will depend on real-world performance, software ecosystem maturity and production scale once commercial deployments begin. However, the launch demonstrates that the competitive landscape is evolving beyond standalone chips toward tightly integrated hardware, networking and software platforms. For AI infrastructure providers, success may increasingly depend on controlling the full compute stack rather than optimizing individual components alone.
