OpenAI and Broadcom present Jalapeño, a chip designed to accelerate language model inference

OpenAI has taken a step that we have been seeing for a long time in the world of artificial intelligence infrastructure: getting fully into the design of its own chips. The company has presented together with Broadcom its first hardware accelerator designed specifically for the inference of large language models, baptized as Jalapeno.

The announcement comes from the heads of both companies. The chip was delivered to OpenAI CEO Sam Altmanand at President Greg Brockmanon the part of the Broadcom CEO Hock Tanand its president, Charlie Kawwas. Beyond the symbolism of the moment, the move confirms that OpenAI wants to control every layer of its infrastructure, from the models to the silicon that runs them.

A chip designed from scratch for inference, not adapted for other uses

The first thing that OpenAI highlights about Jalapeno is that it is not a generic accelerator converted for AI tasks, but rather a design built from the first sketch thinking exclusively about the inference of current and future language models. The company claims that the chip is informed by the systems it already operates daily in ChatGPT, Codex, the API and future agentic products.

According to OpenAI, early testing indicates that this first accelerator will offer a Substantially higher performance per watt to the current state of the art, although the company clarifies that it is still measuring the final performance and that it will publish a detailed technical report in the coming months. The architecture seeks to reduce data movement and balance computation, memory and networks to bring actual utilization closer to the theoretical maximum of the chip.

Engineering samples of the chip are already running in the lab at target production frequency and power, running machine learning workloads that include the model GPT-5.3-Codex-Spark.

Nine months from design to production, with the help of your own models

One of the most striking data from the announcement is the speed of development. OpenAI and Broadcom claim to have led Jalapeno from initial design to manufacturing tape-out in just nine monthsa timeline the companies describe as the fastest ASIC development cycle achieved yet in advanced high-performance semiconductors.

That speed is explained, according to OpenAI, by the combination of joint software and hardware development with the company’s engineering teams, Broadcom’s experience in silicon implementation and the use of OpenAI’s own models to accelerate parts of the design and optimization process. That is, the models that the company makes available to its users would also have participated in the design of the hardware that will run them in the future.

A multi-generation platform, built with partners

Jalapeño is not presented as an isolated projectbut as the first link in what OpenAI and Broadcom describe as a multi-generation computing platform. Also participating in this development Celestica, responsible for the integration of boards, rack systems and scale production.

The roadmap calls for an initial deployment in late 2026, with expansion in subsequent years. According to Hock Tan, the collaboration will allow deploy gigawatt-scale data centers with Microsoft and other partners starting later that year. Broadcom contributes to the platform, in addition to the implementation of silicon, its Tomahawk networking technologydesigned to sustain the connectivity of these large-scale systems.

Richard Ho, head of OpenAI’s hardware program, explained that the chip was optimized around the kernels, memory movement patterns, networks and service schemes that most affect the performance of frontier models. In his words, initial tests show that Jalapeño will run the company’s most relevant workloads close to the theoretical limits of the hardware.

The underlying argument: make inference cheaper to reach more people

OpenAI frames all of this development within what it calls its full-stack infrastructure strategy.– Not limited to developing models or building products on top of them, but also designing chip architecture, kernels, memory systems, networking, load planning, and deployment systems. The idea, according to the company, is that each layer can be optimized with the same objective: to make the models faster, more reliable and cheaper to operate.

Greg Brockman has summarized the logic behind the movement by noting that the world is heading toward a computing-driven economy, and that Jalapeño is part of a long-term infrastructure strategy aimed at making that computing more abundant. According to Brockman, designing more layers of the technology stack internally allows OpenAI to deliver more intelligence with greater efficiency.

At the moment, neither OpenAI nor Broadcom have detailed the price of these systems, nor have they confirmed which specific clients, beyond Microsoft, will be part of the gigawatt-scale deployment planned for 2026. Nor have definitive performance figures been provided, something that both companies have left for the technical report that they promise to publish in the coming months.