OpenAI Unveils Custom Jalapeño Chip to Enhance AI Infrastructure
- tech360.tv
- 1 day ago
- 2 min read
OpenAI has showcased its first custom artificial intelligence chip, dubbed "Jalapeño," designed in conjunction with Broadcom. The development aims to accelerate OpenAI's infrastructure and address the increasing demand for computing power in AI labs.
AI organisations like OpenAI and Anthropic are facing challenges in securing sufficient computing power to operate their advanced chatbots and coding applications. This has led some, including OpenAI, to develop in-house chips.
Such custom chip development can reduce costs and offer an alternative to Nvidia's graphics processing units, which are commonly used for AI tasks. The "Jalapeño" chip was specifically designed by OpenAI engineers with Broadcom to perform AI inference, which involves processing data to generate a user's chatbot query response.
Broadcom Chief Executive Officer Hock Tan stated that the chip performs as effectively as Nvidia's Blackwell chips or Alphabet's Google tensor processing units. OpenAI hardware chief Richard Ho affirmed that the "Jalapeño" processor is engineered for speed and efficiency with the large language models powering many AI applications.
Ho noted, "It will be performant on, we think, all kind of future iterations of LLMs." OpenAI intends to deploy "Jalapeño" by the end of this year, marking it as the initial step in a multi-generation chip development strategy.
Canadian electronics manufacturer Celestica will construct the server systems, which, like the chips, are exclusively for OpenAI's use. The San Francisco-based company confirmed it has samples of the chip running in its labs, operating at target power and performance with its GPT-5.3-Codex-Spark AI model.
OpenAI engineers completed the chip design in approximately nine months, partly by utilising AI to expedite specific design aspects, before sending it to TSMC for manufacturing. Other major companies, including Meta Platforms, Amazon, and Google, have also turned to Broadcom and Marvell for specialised design services and intellectual property to create their in-house chips.
Anthropic is also considering building its own AI chip. However, Broadcom Chief Executive Officer Hock Tan indicated that the profit margin on custom chips is currently not as high as for other products, such as networking switches, due to an AI-related surge in memory demand.
AI chips require significant amounts of high-bandwidth memory, which impacts Broadcom's margins on these custom products. Tan added that South Korea's SK Hynix and Samsung Electronics supply memory chips to Broadcom.
OpenAI has unveiled its first custom AI chip, "Jalapeño," developed with Broadcom to enhance its AI infrastructure.
The chip is designed for AI inference tasks, processing data for chatbot queries, and is intended to be deployed by the end of the year.
"Jalapeño" aims to reduce costs and offer an alternative to traditional graphics processing units, with Broadcom's CEO asserting its performance matches leading market alternatives.
Source: Reuters
