OpenAI, in partnership with Broadcom, introduced Jalapeño - its first caste ASIC processor, fused under the interface of large language models and agent systems. The main sensation was a record-short development cycle: from concept to engineering samples took only nine months, whereas in the industry, creating a chip from scratch usually takes one and a half to two years. To accelerate the process, Broadcom’s prebuilt logic blocks were reused and OpenAI AI models were applied in the design stages.

According to the developers, the Jalapeño architecture was designed from scratch under specific patterns of memory consumption and network interactions of modern neuronetworks. The OpenAI claims that the chip is demonstrating unprecedented performance per watt and computing power disposal, approaching theoretical limits. Engineering samples are already successfully running loads such as GPT-5.3-Codex-Spark, but the company’s precise benchmarks and memory configurations have been kept secret for now, comparing the novelty only to abstract "advanced equipment".

Jalapeño’s mass deployment is scheduled for the end of 2026 in data centers on a gigawatt scale, created together with Microsoft and other partners. Notably, the chip was designed with an eye to supporting not only internal OpenAI models but also third-party language architectures. This opens the way for companies to diversify their business: in case of stable supplies from TSMC and Broadcom, OpenAI can start selling its own AI accelerators to external customers, competing with Nvidia Blackwell solutions and future chips on Rubin architecture.