OpenAI disclosed its first custom artificial intelligence chip, Jalapeño, on Wednesday, a processor the company said it designed together with Broadcom to accelerate a specific AI workload known as inference. The announcement highlights OpenAI's move to build bespoke hardware to support the high-compute needs of modern chatbots and coding assistants.
Engineers at OpenAI collaborated with Broadcom on the chip design. OpenAI said Jalapeño is targeted at inference - the phase when trained models process incoming user queries and produce responses - and that the processor is tailored to work quickly and efficiently with the large language models (LLMs) that power many contemporary AI services.
Broadcom's chief executive Hock Tan told Reuters the Jalapeño design is comparable to Nvidia's Blackwell chips and Alphabet's tensor processing units in performance. OpenAI's hardware lead Richard Ho said the processor was engineered to be "performant on, we think, all kind of future iterations of LLMs." OpenAI also described Jalapeño as the first release in a planned multi-generation chip roadmap.
The company reported it already has chip samples operating in its laboratories and that the devices met target power and performance metrics when running OpenAI's GPT-5.3-Codex-Spark AI model. OpenAI intends to deploy Jalapeño by the end of this year.
Production of server systems that will incorporate the chips will be handled by Canadian electronics manufacturer Celestica, OpenAI said. Like the processors themselves, those server systems are intended for OpenAI's internal use only.
OpenAI said it completed the chip design in roughly nine months before sending the design to Taiwan Semiconductor Manufacturing Co. (TSMC) for fabrication. The company also noted it used AI to accelerate specific stages of the design process.
The expansion into custom silicon follows a pattern among large AI labs and cloud companies seeking alternatives to general-purpose graphics processing units that are commonly used for model training and inference. OpenAI and other labs have faced challenges securing sufficient compute capacity to operate the most powerful models.
Broadcom's Hock Tan cautioned that margins on custom AI chips differ from the company's higher-margin networking products. He explained that AI accelerators demand large amounts of high-bandwidth memory, which has weighed on Broadcom's profitability for those specific chips. Tan identified SK Hynix and Samsung Electronics as memory suppliers to Broadcom.
Key points
- OpenAI and Broadcom jointly designed the Jalapeño processor, optimized for inference tasks on large language models.
- Samples of the chip are running in OpenAI's labs with the GPT-5.3-Codex-Spark model, and deployment of Jalapeño is planned before year-end.
- Celestica will build the server systems for use exclusively by OpenAI; Broadcom noted AI-related memory demand reduces margin on such custom chips compared with other product lines.
Risks and uncertainties
- Profitability pressure for custom AI chips - Broadcom has indicated that high-bandwidth memory requirements reduce margins on custom AI accelerators, which could affect suppliers in the semiconductor and memory sectors.
- Supply chain reliance - OpenAI has sent its designs to TSMC for manufacturing and is relying on SK Hynix and Samsung for memory components, creating vendor dependencies in the production chain.
- Limited external availability - Celestica will produce server systems that, like the chips, are intended only for OpenAI use, so market impact is constrained by the closed deployment plan.
Several technology companies have explored in-house chip development to control costs and secure specialized computing capacity. OpenAI's chip effort was first reported as an exploration in 2023. Other major cloud and AI firms have similarly worked with contract designers such as Broadcom and Marvell to obtain design services and intellectual property that are not straightforward to replicate inside a company.
OpenAI emphasized Jalapeño represents the opening step of a longer-term chip strategy. The company said its engineers used AI tools to accelerate parts of the nine-month design cycle before moving to TSMC for fabrication. OpenAI has not provided broader commercialization plans for the chip beyond internal deployment.
Broadcom acknowledged the bright demand for memory driven by AI workloads, which affects the economics of bespoke AI processors compared with its network switch business. The company named South Korean and Taiwanese memory suppliers as partners in the program.
The move positions OpenAI among a growing set of organizations building custom hardware to support modern AI systems, while also underscoring ongoing supply and margin dynamics within the semiconductor and memory supply chain.