OpenAI has taken a concrete step into hardware independence, unveiling its first custom-designed AI chip. Named Jalapeño, the processor is purpose-built for inference — the continuous, computationally intensive task of generating responses from already-trained models — rather than the more costly upfront work of training. Engineering samples are already running the company’s GPT-5.3-Codex-Spark model in its labs, and early testing indicates substantially better performance per watt than current off-the-shelf alternatives, though precise metrics have not been released. The chip was designed in roughly nine months, a timeline the company attributes partly to the use of its own AI models in the design process.

The silicon was co-developed with Broadcom and manufactured by Taiwan’s TSMC, while Canada’s Celestica will build the server systems, which are reserved exclusively for OpenAI’s own infrastructure. Broadcom’s chief executive, Hock Tan, described the chip as comparable in capability to Nvidia’s Blackwell GPUs and Google’s tensor processing units. He also flagged that the high-bandwidth memory required for such AI chips — supplied by South Korea’s SK Hynix and Samsung Electronics — pressures profit margins on custom silicon products. OpenAI’s hardware lead, Richard Ho, said the architecture is designed to perform efficiently across future generations of large language models, not just the company’s current offerings.

Viewed from Silicon Valley, the move mirrors a structural shift across the AI industry. Google, Amazon, Microsoft and Meta have all pursued proprietary chips to escape the cost and supply constraints of Nvidia’s dominant GPUs. Anthropic is also exploring its own design. Broadcom has emerged as the common partner in many of these efforts, providing design services and intellectual property that are difficult for software-native firms to replicate internally. For OpenAI, which has been one of Nvidia’s largest customers, bringing chip design in-house is a direct attempt to lower inference costs at the enormous scale required by ChatGPT and Codex, and to reduce exposure to a single supplier.

Initial deployment of Jalapeño is scheduled before the end of this year at data centres operated by Microsoft and other partners, with volume production expected to ramp through 2026 and 2027. The chip is the first in a planned multi-generation family, and the company’s longer-term ambition, outlined by executives, is to power 10 gigawatts of compute with its own silicon by 2029. The next factual milestone will be the release of detailed performance benchmarks and the start of commercial-scale operation, which will test whether the claimed efficiency gains translate into a meaningful shift in the infrastructure economics of generative AI.