On June 24, OpenAI officially released its first self-developed AI chip, Jalapeño, co-developed with Broadcom, positioned as an inference processor optimized for large language models (LLMs). The chip is manufactured by TSMC, while Celestica, a Canadian electronics manufacturer, builds the server systems. Broadcom CEO Hock Tan stated that the Jalapeño's performance is comparable to Nvidia's Blackwell.
Jalapeño Technical Specs: 9-Month Tape-Out, Inference-Optimized Design, and Early Test Data
According to an official OpenAI article, the Jalapeño is a processor tailored specifically for inference workloads, not training chips. From initial design to tape-out, it took only 9 months, a pace attributed to OpenAI using its own AI models to help optimize chip design.
OpenAI's hardware chief, Richard Ho, noted that the chip efficiently handles key workloads; early official tests show significantly better performance per watt compared to similar products on the market. Broadcom CEO Hock Tan stated that its performance rivals Nvidia's Blackwell and Google's TPU.
Broadcom CEO Hock Tan's Performance Claims and HBM Margin Challenges
Hock Tan has delivered physical samples of the Jalapeño to Sam Altman. He also acknowledged that due to the large demand for high-bandwidth memory (HBM) in AI chips, the profit margins for custom chips are currently lower than products like network switches. HBM is mainly supplied by SK Hynix and Samsung Electronics.
He also pointed out that customer demand for computing power is exploding and is expected to continue through 2028. OpenAI President Greg Brockman told CNBC that the degree to which AI models have accelerated chip development is surprising, and the company's need for computing power is extremely urgent.
OpenAI's Diversified Computing Strategy: Collaboration with Broadcom, AWS Trainium, AMD, Cerebras
The Jalapeño is part of OpenAI's diversified computing strategy. Ben Barringer, technology research director at investment firm Quilter Cheviot, analyzed that tech giants are actively diversifying chip sources to avoid being constrained by a single supplier. Beyond its partnership with Broadcom, OpenAI has reached an agreement with AWS to use Trainium chips and has established collaborations with AMD and Cerebras, which will go public in May 2026.
OpenAI and Broadcom previously announced plans to deploy data centers with a total capacity of 10GW over the next few years, in partnership with Microsoft and others. The launch of a self-developed chip comes at a critical time as OpenAI prepares for an IPO, with market expectations of a valuation possibly reaching $1 trillion. A self-developed chip helps establish a full-stack technical advantage and reduce service costs, supporting higher valuation expectations.
Frequently Asked Questions
What is the origin and significance of the name Jalapeño?
Jalapeño is a type of Mexican chili pepper with intense spiciness. Tech media Engadget called it a "spicy start" for OpenAI entering the chip space; BusinessKorea noted that the name also reveals OpenAI's direct aim at the Nvidia-centered monopoly in the AI computing infrastructure market. In the semiconductor and software industries, R&D teams commonly use food, plants, or place names as development codenames. The OpenAI team used this codename during the 9-month secret development phase and kept it at launch.
Is the Jalapeño an inference chip or a training chip, and what's the difference?
According to reports, the Jalapeño is an inference processor optimized specifically for the inference needs of large language models, not a training chip. Inference refers to the process where an AI model generates responses during actual use, which is more sensitive to energy efficiency and latency; training requires extremely high compute density. Early tests of the Jalapeño show significantly better performance per watt compared to similar inference products.
What impact does OpenAI's self-developed chip have on Nvidia?
According to analyst Ben Barringer's assessment, tech giants (including OpenAI) are actively diversifying chip sources to avoid reliance on a single supplier. The launch of the Jalapeño allows OpenAI to reduce its dependence on Nvidia GPUs, complemented by a diversified supply chain including AWS Trainium, AMD, and Cerebras, building a more flexible computing mix. However, the Jalapeño's scale and maturity are still in the early deployment phase, making it unlikely to fully replace Nvidia's market position in the short term.
