OpenAI Jalapeño — A New AI Chip Developed Together with Broadcom

OpenAI Jalapeño — A New AI Chip Developed Together with Broadcom

OpenAI and Broadcom have introduced Jalapeño — the first AI accelerator designed by OpenAI as part of its own infrastructure for running large language models. The new chip is intended to support primarily inference, the stage in which an AI model responds to user queries, powers applications, agents, APIs, and tools such as ChatGPT or Codex.

This is an important step in the development of OpenAI’s infrastructure. The company does not want to rely solely on ready-made hardware solutions available on the market. Jalapeño is intended to become part of a broader strategy to build a full technology stack: from models and products, through request-handling systems, to specialized chips, networks, servers, and data centers.

A Chip Designed for LLM Inference

According to OpenAI, Jalapeño is not a classic, general-purpose AI accelerator adapted to many different types of workloads. The chip was designed from the ground up for modern LLMs [Large Language Models] and for the way OpenAI operates its products in practice.

The goal is to combine high throughput, low latency, and better energy efficiency. This is especially important for interactive services where users expect fast responses while the system must simultaneously handle a very large number of requests.

OpenAI indicates that the Jalapeño architecture is designed to reduce unnecessary data movement and better balance computing power, memory, and network communication. In practice, this may translate into higher real-world performance compared with the theoretical capabilities of the hardware.

Collaboration Between OpenAI, Broadcom, and Celestica

OpenAI is responsible for the conceptual design, while Broadcom contributes its experience in silicon implementation, networking technologies, and chip production. Celestica is also involved in the project, supporting areas related to boards, racks, and system integration.

Broadcom is expected to play a key role not only in the chip itself, but also in the network infrastructure. OpenAI points, among other things, to the use of Broadcom networking technologies, including Tomahawk chips, which are important for building large computing clusters.

Nine Months from Design to Tape-Out

One of the most interesting elements of the announcement is the pace of work. OpenAI reports that Jalapeño went from the initial design stage to tape-out in nine months. Tape-out is the point at which a chip design is handed over for manufacturing.

The company emphasizes that OpenAI models were also used in the design process to accelerate parts of the engineering and optimization work. In other words, AI helped design the infrastructure on which future generations of AI will run.

Better Energy Efficiency

OpenAI states that early tests indicate significantly better performance per watt than current state-of-the-art solutions. At the same time, the company notes that final measurements are still ongoing, and a detailed technical report on performance is expected to be presented in the coming months.

This is an important caveat. At this stage, Jalapeño should be treated as a very significant infrastructure announcement, but not as a product whose full parameters have already been publicly and independently confirmed.

Gigawatt-Scale Infrastructure and Future Generations

Jalapeño is intended to be the first element of a multi-generation computing platform being developed by OpenAI and Broadcom. OpenAI says the platform is expected to be deployed from the end of 2026 and developed further in the following years.

The goal is to launch infrastructure at gigawatt scale together with data center partners. This shows how important the hardware layer has become for the development of AI. The more users rely on AI models, the more important the cost of serving queries, availability of computing power, reliability, and energy efficiency become.

What Could Jalapeño Mean for Users?

For end users, Jalapeño will not be a product they buy directly. Its significance may, however, be felt in AI services.

Better inference infrastructure may translate into:

faster AI model responses,
lower operating costs for API-based applications,
greater service availability under heavy load,
the ability for AI agents to perform more complex tasks,
smoother operation of tools such as ChatGPT and Codex,
further development of AI products for businesses, developers, education, and research.

OpenAI presents Jalapeño as part of a long-term strategy to increase access to advanced artificial intelligence. In practice, this means that AI development depends not only on new models, but also on physical infrastructure: chips, networks, data centers, and workload management systems.

Why Does It Matter?

Jalapeño shows that the largest AI players are moving deeper into designing their own hardware. This is a logical direction: language models are becoming increasingly advanced, and operating them requires enormous computing power. Software optimization alone is not enough if the hardware infrastructure cannot keep up with the scale of use.

For the market, this represents another stage of competition in AI. Advantage will be built not only by companies with the best models, but also by those that can run, scale, and deliver them to users most efficiently.

OpenAI Jalapeño is therefore more than just another chip. It is a signal that the future of artificial intelligence will depend on the tight integration of models, software, hardware, networks, and data centers.

 July 1, 2026
Source: OpenAI

Tags: OpenAI, Jalapeño, Broadcom, AI chip, AI accelerator, artificial intelligence, LLM, AI inference, ChatGPT, Codex, AI infrastructure, data centers, AI processors, generative AI

OpenAI Jalapeño, Jalapeño chip, OpenAI Broadcom, chip AI, procesor AI, akcelerator AI, Intelligence Processor, inferencja AI, infrastruktura AI, chip do sztucznej inteligencji, procesory do LLM, duże modele językowe, LLM, ChatGPT, Codex, OpenAI, Broadcom, Celestica, centra danych AI, hardware AI, układy scalone AI, AI inference, wydajność AI, efektywność energetyczna AI, generatywna AI, sztuczna inteligencja, własny chip OpenAI, akcelerator inferencji, infrastruktura dla ChatGPT,AI chip, AI processor, AI accelerator, Intelligence Processor, AI inference, inference chip, LLM inference, large language models, LLM, ChatGPT, Codex, OpenAI infrastructure, Broadcom AI, Celestica, AI data centers, AI hardware, AI silicon, custom AI chip, generative AI, AI computing, performance per watt, AI infrastructure, OpenAI chip, inference accelerator