Jalapeño OpenAI and Broadcom Unveil First Jointly Developed AI Chip

Related Vendors

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

SIMTOS (Seoul International Machine Tool Show)

OpenAI and Broadcom have unveiled Jalapeño, a custom AI inference processor. It is OpenAI's first proprietary chip. The ASIC was developed jointly with Broadcom and is intended for use in large AI data centers.

OpenAI describes Jalapeño as an architecture developed from the ground up for modern large language models. The aim is to achieve lower latencies, better energy efficiency, and higher utilization of existing computing, storage, and networking resources.

OpenAI and Broadcom have not disclosed specific performance data so far. The companies only mention early internal tests, in which Jalapeño is said to achieve significantly better performance per watt compared to current solutions. Reliable benchmarks, memory specifications, manufacturing details, or comparison systems have not been published.

According to OpenAI, engineering samples are already running in the lab at the intended frequency and power consumption. As an example, the company mentions ML workloads, including GPT-5.3 Codex Spark. A technical evaluation of the performance data is to follow later.

Few Architectural Details, Initial Conclusions from Images

The announcement also remains reserved regarding the architecture. It is known that Jalapeño is designed to reduce data movement and align compute, memory, and network more closely with typical LLM inference patterns. Broadcom contributes, among other things, its experience with custom silicon and networking technology.

From the published images, only tentative conclusions can be drawn. Editors at Tom's Hardware have identified a large compute chiplet, several HBM stacks, and additional dies for I/O or structural purposes. However, such assessments remain speculative as long as the companies do not provide official package or die data.

The suspected use of HBM suggests that OpenAI and Broadcom are not solely aiming for low cost per component but for high bandwidth and low latencies. This aligns with interactive inference scenarios where response time, throughput, and energy demand must be optimized together.

Vertical Integration of AI Infrastructure

Strategically, Jalapeño is another step toward a more vertically integrated AI infrastructure and greater independence from Nvidia. OpenAI is not only developing models and products but is also increasingly delving into data center architecture, scheduling, networking, memory hierarchies, and now chip design.

For OpenAI, a proprietary inference accelerator can reduce dependence on standard GPUs, particularly from Nvidia. Jalapeño is unlikely to completely replace these systems initially. Computationally intensive training tasks are expected to continue running on GPU or specialized training platforms.

The companies state that the chip was developed up to the tape-out stage within nine months. OpenAI points out the use of its own models to support parts of the design and optimization process. It remains important to note that the actual development performance is also likely to depend on Broadcom's existing ASIC expertise, IP reuse, and manufacturing ecosystem.

Broadcom and OpenAI speak of a multi-year platform and multiple chip generations. Starting in 2026, Jalapeño is expected to be deployed in gigawatt-scale data centers with Microsoft and other partners. It remains unclear whether the hardware will be used exclusively for OpenAI workloads or could later become more widely available.