← Back to feed

OpenAI and Broadcom on Wednesday delivered engineering samples of Jalapeño, OpenAI’s first custom inference processor. The chip was handed to CEO Sam Altman and President Greg Brockman by Broadcom CEO Hock Tan and President Charlie Kawwas, marking the public close of a partnership first announced in October 2025.

The name is the least surprising thing about it. The rest of the story is a direct challenge to Nvidia’s dominance of the AI inference market.

What Jalapeño Is

Jalapeño is an ASIC — purpose-built silicon, not a general-purpose GPU. That distinction matters: ASICs sacrifice flexibility for efficiency. They can’t adapt to new workloads the way a GPU can, but for a fixed, well-understood workload like running frontier LLMs at scale, they can deliver substantially better performance per watt. OpenAI knows its workloads better than anyone. It trains and serves the models, writes the kernels, builds the serving infrastructure. Jalapeño was designed from the ground up around that stack.

The architecture targets three specific constraints: data movement, compute utilization, and memory-network balance. The stated goal is to bring realized utilization much closer to theoretical peak — the gap between what a chip can do and what it actually delivers is where inference costs live. Broadcom’s Tomahawk networking silicon handles the interconnect layer. Celestica joins as the third partner, responsible for board, rack system integration, and scalable production systems.

Early samples are already running ML workloads in the lab at production target frequency and power, including GPT-5.3-Codex-Spark. OpenAI says it will publish a detailed technical performance report in the coming months.

Why Nine Months Matters

Nine months from design to production tape-out is fast for advanced AI silicon. The usual timeline for a chip of this complexity runs 18-24 months. OpenAI attributed part of the compression to its own AI models accelerating design work — the first documented instance of OpenAI’s models contributing meaningfully to hardware engineering output at this scale.

The partnership structure spread the risk. Broadcom handled chip implementation; Celestica handled system integration. OpenAI provided the model architecture and kernel knowledge. The division of labour let each party move in parallel rather than sequentially.

The Stack Play

OpenAI’s framing is explicit: Jalapeño is an infrastructure move, not just a cost-reduction exercise. The company now controls models, kernels, serving systems, data centres (in development), and — as of today — silicon. Greg Brockman described the goal as making “compute more abundant,” translating to lower costs and broader access.

The deployment plan is gigawatt scale across multiple chip generations. OpenAI’s data centre partners will host the hardware. The chip is designed to work with LLMs broadly, not only OpenAI’s own models, signalling possible future licensing or partnership arrangements.

Google has been running custom silicon since the first TPU in 2016. Amazon has Trainium for training and Inferentia for inference. Apple runs its Neural Engine. For inference specifically, the GPU still dominates the market. Jalapeño is OpenAI’s entry into that market, starting at the workloads it already owns.

What It Doesn’t Cover

Training. Jalapeño is an inference chip. Pre-training of frontier models will still run on Nvidia hardware for the foreseeable future. The market OpenAI is targeting with Jalapeño is the serving side: the continuous, high-volume compute required to respond to hundreds of millions of ChatGPT and Codex requests daily.

Inference costs have become the primary constraint on OpenAI’s economics. Every percentage point reduction in cost per token at inference scale flows directly to margins or to lower prices. The chip is solving the right problem for a company that already spends billions annually serving live traffic.