OpenAI develops 'Jalapeño' inference chip in nine-month sprint

OpenAI has developed a custom inference chip, internally codenamed "Jalapeño," in a rapid nine-month development cycle. The move marks a significant step toward vertical integration as the company seeks to optimize how its large language models are deployed.

Developed in partnership with Broadcom, the Jalapeño chip moved from initial design to manufacturing tape-out in just nine months. The hardware is specifically engineered for LLM inference, focusing on increasing performance and efficiency when running models. Notably, OpenAI did not rely solely on human engineers; the company used its own AI models to accelerate various stages of the design and optimization process.

The push for custom silicon

For years, OpenAI and other leading AI laboratories have operated under a heavy dependence on third-party hardware, specifically NVIDIA's H100 and B200 GPUs. While these GPUs are the industry standard for both training and inference, the sheer scale of demand has created significant bottlenecks. As the user base for OpenAI's models grows, the associated power requirements and procurement costs have become critical constraints. By developing its own silicon, OpenAI aims to gain tighter control over the hardware-software stack, ensuring the physical architecture is tuned to the specific mathematical requirements of its models.

Breaking the hardware bottleneck

This shift toward custom silicon is a strategic attempt to reduce systemic reliance on a single hardware vendor. If OpenAI can successfully deploy Jalapeño at scale, it could drastically lower the operational cost of intelligence. Furthermore, the use of AI to design the very hardware it runs on creates a recursive loop of efficiency. This capability suggests that future iterations of hardware could be developed at a pace far exceeding traditional semiconductor cycles, potentially accelerating the entire industry's trajectory.

The road to deployment

While the tape-out marks a major milestone, the industry will be watching for the actual deployment of Jalapeño in OpenAI's data centers. The primary challenge remains the transition from a successful design to mass production and integration into existing clusters. It remains to be seen how this custom silicon will perform against the latest generation of NVIDIA chips in real-world production environments, but the project establishes a blueprint for AI labs to move from software providers to full-stack hardware architects.