OpenAI has officially unveiled Jalapeño, a custom-built chip developed with Broadcom to handle AI inference tasks. This application-specific integrated circuit is designed to process user requests for services like Codex and ChatGPT rather than training new models. The announcement marks the first time the company has revealed its own hardware architecture nine months after confirming the partnership with the semiconductor giant. OpenAI stated the processor will power current systems and future large language models, aiming to reduce reliance on standard graphics processing units. By dedicating silicon specifically to inference workloads, the firm expects to improve efficiency and lower costs for running its services. This move signals a shift from renting computing power to owning the infrastructure required to deliver AI responses.
The significance lies in controlling the supply chain and optimising performance for specific tasks. Custom chips often deliver better speed and energy efficiency compared to general-purpose hardware. OpenAI can now tailor the architecture to match the exact needs of its models without compromise. This reduces the need to purchase vast quantities of off-the-shelf components from third-party vendors. The strategy also positions the company to manage its own data centre requirements more effectively as demand grows.
- Designed specifically for AI inference workloads
- Developed in partnership with Broadcom
- Targets cost reduction for running large language models
