OpenAI has taken an important step toward controlling more of its AI infrastructure with the introduction of Jalapeño, its first custom AI processor developed with Broadcom. The chip targets inference workloads, complements OpenAI’s existing GPU deployments, and begins a multi-generation silicon roadmap. For ISVs, silicon developers, and enterprise IT organizations, the announcement signals a broader shift toward vertically integrated AI platforms where software, hardware, and infrastructure evolve together to improve efficiency, lower costs, and expand AI deployment.
OpenAI CEO Sam Altman and Broadcom CEO Hock Tan hold the new Jalepeño chip. (Source: © OpenAI)
OpenAI has introduced its first custom AI processor, Jalapeño, marking the company’s entry into AI silicon design. Developed with Broadcom, the application-specific integrated circuit (ASIC) targets inference workloads that power ChatGPT and other OpenAI services. The project extends OpenAI’s strategy beyond model development into hardware, systems, and infrastructure, giving the company greater control over performance, cost, and long-term compute capacity.
Jalapeño represents the first Intelligence Processor in a multi-generation compute platform that OpenAI and Broadcom plan to expand over the coming years. OpenAI designed the processor around the future requirements of its large language models rather than adapting existing accelerator architectures. The company completed the chip in only nine months, using its own AI models to accelerate portions of the engineering process. Greg Brockman, OpenAI president, noted that AI-assisted development shortened design cycles beyond the company’s expectations.
The processor focuses on inference rather than training. Inference now consumes an increasing share of AI compute as hundreds of millions of users generate prompts, images, software, and video every day. OpenAI stated that Jalapeño delivers performance comparable to leading AI accelerators while reducing latency and improving performance per watt. Those improvements directly influence operating costs across hyperscale AI deployments.
Unlike GPUs, which support a broad range of workloads, ASICs optimize specific functions. That specialization enables lower power consumption, simpler execution pipelines, and improved efficiency for targeted applications. OpenAI expects Jalapeño to complement rather than replace its existing accelerator fleet, allowing the company to reserve GPUs for training while shifting inference workloads to custom silicon.
The announcement reflects OpenAI’s broader objective to become a full-stack AI company. Beyond developing foundation models, OpenAI now designs processors, contributes to system architecture, develops software infrastructure, and optimizes complete AI platforms. Greater control over the stack allows the company to reduce dependence on third-party hardware suppliers, improve deployment efficiency, and respond more quickly to growing compute demand.
OpenAI continues to rely on a diverse silicon ecosystem. Nvidia remains its largest supplier of GPUs, and the company has expanded relationships with AMD, Amazon Web Services through Trainium, and Cerebras. Jalapeño adds another layer to that strategy by creating an internal inference platform optimized specifically for OpenAI workloads rather than general-purpose AI computing.
Broadcom continues to strengthen its position as a strategic silicon partner for hyperscalers and frontier AI laboratories. The company already develops custom AI processors for several cloud providers, and OpenAI’s selection reinforces Broadcom’s expertise in advanced ASIC development. Broadcom CEO Hock Tan indicated that customer demand for AI compute continues to outpace available supply and expects deployments to accelerate through 2028.
The roadmap extends well beyond a single processor. OpenAI and Broadcom previously announced plans to deploy custom AI racks capable of supporting approximately 10 GW of compute infrastructure. Initial Jalapeño deployments will begin during late 2026, with broader production during 2027 and continued expansion afterward. OpenAI has already begun running machine-learning workloads on early silicon inside its laboratories.
The announcement also illustrates how AI increasingly contributes to semiconductor development. OpenAI used its own models to accelerate portions of chip design, demonstrating that AI now supports the engineering tools required to build future generations of AI hardware. That recursive development cycle shortens product schedules, improves engineering productivity, and reduces development costs across increasingly complex semiconductor projects.
For ISVs and enterprise customers, Jalapeño represents more than another AI accelerator. The processor signals tighter integration between models, software frameworks, inference engines, networking, and hardware. Organizations deploying AI at scale will increasingly evaluate complete computing platforms rather than individual processors. Performance per watt, deployment efficiency, software compatibility, and infrastructure economics now influence purchasing decisions as much as peak compute performance.
Jalapeño marks the beginning of OpenAI’s transition from an AI software company into a vertically integrated AI infrastructure provider. Custom silicon, optimized software, and purpose-built systems strengthen OpenAI’s ability to scale future AI services while reducing infrastructure costs. The next several product generations will determine how effectively that strategy competes alongside established AI hardware platforms.
What do we think?
OpenAI’s move into custom silicon follows a logical progression toward infrastructure ownership. Jalapeño will not replace GPUs across every workload, yet it gives OpenAI greater control over inference economics and long-term capacity planning. Broadcom provides proven ASIC expertise, and OpenAI contributes workload knowledge. Together they have established a foundation for a sustainable multi-generation compute platform.
OpenAI’s custom processor may represent an inflection point in AI infrastructure. Model developers increasingly recognize that competitive advantage extends beyond algorithms into silicon, packaging, networking, and systems engineering. As more AI companies design purpose-built processors, the industry will shift toward vertically integrated platforms that optimize complete workloads rather than individual devices. That transition will reshape semiconductor development, cloud infrastructure, enterprise AI deployment, and software optimization throughout the AI ecosystem.
YOU LIKE THIS KIND OF STUFF? WE HAVE LOTS MORE. TELL YOUR FRIENDS, WE LOVE MEETING NEW PEOPLE.