OpenAI and Broadcom unveil Jalapeño, a new AI chip built for LLM inference

Thanks to the exponential growth of ChatGPT and other LLM-based applications, NVIDIA has grown from a $200 billion company into the first public company to reach a $5 trillion market cap.

Even though hyperscalers such as Google and Amazon have their own mature AI accelerators, NVIDIA still dominates the AI infrastructure market with multiple generations of GPUs. Microsoft, OpenAI, and Meta remain among NVIDIA’s largest customers, while Google and Amazon continue to be significant NVIDIA customers as they serve AI workloads for customers on their cloud platforms.

Today, OpenAI and Broadcom announced Jalapeño, OpenAI’s first custom “Intelligence Processor” designed specifically for large language model inference. The new chip is the first product from a multi-generation compute platform being developed by OpenAI. OpenAI highlighted that Jalapeño was built from the ground up for current and future LLM workloads, rather than being a general-purpose accelerator adapted for AI.

Despite heavy competition from Gemini, Claude, Copilot, and others, ChatGPT remains the most used AI platform in the world. OpenAI mentioned that it leveraged its knowledge of how its models and products run at scale, including ChatGPT, Codex, the API, and future agentic AI systems, to design this new chipset. Its chip architecture reduces data movement while balancing compute, memory, and networking resources.

Jalapeño will be deployed in production systems starting in late 2026; however, engineering samples are already running machine learning workloads in OpenAI’s labs at production target frequency and power. According to its internal testing, OpenAI claims this chip can deliver “substantially better” performance per watt, and a detailed technical report is expected in the coming months.

While OpenAI designed the chip, Broadcom handled silicon implementation and networking technologies, including Tomahawk networking silicon, and Celestica is assisting with board, rack, and system-level integration.

OpenAI pointed out that Jalapeño went from initial design to manufacturing tape-out in just nine months, which it claims is the fastest ASIC development cycle achieved for a high-performance advanced semiconductor. The company attributed the speed of development to its own LLMs, which were used during the chip design and optimization process.

Broadcom CEO Hock Tan stated that the company's plan is to deploy the Jalapeño platform at a gigawatt scale with Microsoft and other partners starting in 2026.

With Jalapeño, OpenAI joins Google, Microsoft, and Amazon to become a full-stack AI player. The company already develops models and products, and is now moving deeper into infrastructure, including chips, kernels, networking, scheduling, and deployment systems.