OpenAI and Broadcom have developed a custom AI chip called Jalapeno, designed specifically for large language model inference. The companies claim the processor can reduce costs by 50% and plan to deploy it at scale by late 2026.
OpenAI is moving beyond relying solely on third-party semiconductors. The company has partnered with Broadcom to create Jalapeno, a custom Application-Specific Integrated Circuit (ASIC) tailored for AI inference workloads.
Jalapeno represents OpenAI's first foray into custom chip design. Unlike general-purpose processors, ASICs are optimized for specific tasks—in this case, running inference on large language models. When users interact with ChatGPT or similar systems, inference processors handle the computational work of generating responses.
The chip addresses a critical pain point in AI infrastructure: cost. By designing hardware specifically for inference operations, OpenAI and Broadcom say Jalapeno can cut expenses by 50% compared to existing solutions. This efficiency gain matters significantly as AI companies scale operations and face mounting computational demands.
OpenAI is currently testing Jalapeno samples. Full-scale deployment is targeted for late 2026, giving the company time to validate performance and integrate the chip into its infrastructure.
The move reflects a broader industry trend. Major tech companies including Google, Amazon, and Meta have developed custom chips to reduce dependency on vendors like Nvidia and optimize for their specific workloads. For OpenAI, controlling its own silicon gives the company greater flexibility in pricing, performance, and supply chain management.
Broadcom's involvement brings manufacturing expertise and established relationships with foundries. The partnership allows OpenAI to leverage Broadcom's semiconductor knowledge while maintaining focus on AI development.
Custom chips alone won't solve all infrastructure challenges facing AI companies, but they provide meaningful advantages in cost and efficiency. Jalapeno's 50% cost reduction claim could significantly impact OpenAI's operational expenses and competitiveness as the AI market intensifies.
PsiQuantum CEO Victor Peng outlined the company's quantum computing strategy centered on photons as qubits, emphasizing built-in error correction capabilities and progress toward fault tolerance.
Ray-Ban Meta glasses collect visual data and audio through built-in cameras and microphones, raising privacy concerns for users and those around them. Not all security features are equally robust across the device's functionality.
SK Hynix is channeling AI boom windfalls into South Korea's corporate bond market while simultaneously committing $38 billion to domestic chip fab expansion.