OpenAI's Jalapeño Chip Is Here — And It's a Direct Challenge to Nvidia's AI Dominance

For the last four years, Nvidia has been the undisputed landlord of the AI economy. If you wanted to run a frontier AI model, you paid Nvidia's prices, waited in Nvidia's queue, and built your infrastructure around Nvidia's ecosystem. OpenAI spent billions doing exactly that.

That arrangement just changed.

On June 24, OpenAI and Broadcom unveiled Jalapeño — OpenAI's first custom AI inference chip. It is not a research prototype or a proof of concept. Engineering samples are already running real workloads in the lab, and initial deployment is planned by the end of 2026.

What Jalapeño Is

Jalapeño is an ASIC — an Application-Specific Integrated Circuit — designed from the ground up for one purpose: running large language models at scale. Unlike Nvidia's GPUs, which are general-purpose accelerators adapted for AI, Jalapeño is purpose-built for LLM inference, the compute-intensive process of serving AI model responses to users in real time.

OpenAI designed the chip around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models. Early testing shows Jalapeño will deliver performance per watt substantially better than current state-of-the-art hardware.

The chip was designed in collaboration with Broadcom, with chip implementation and networking handled by Broadcom, and production systems industrialised with partner Celestica.

Nine Months, Start to Finish

Jalapeño was co-developed from initial design to manufacturing tape-out in just nine months — what may be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors.

That speed was itself partly an AI story. The companies attributed this speed to a deep software-hardware co-development process that actively used OpenAI's own models to accelerate parts of the chip design. The same models being served to users helped design the infrastructure that will serve future models.

Why This Matters: The Economics of Inference

Training a frontier AI model is expensive but infrequent. Inference — running the model billions of times a day to answer user queries — is where the real cost lives. For OpenAI, which serves hundreds of millions of users across ChatGPT, Codex, and the API, inference costs are a significant drag on the path to profitability.

OpenAI president Greg Brockman explained the company's approach: "We have a deep understanding of the workload. We've really been looking for specific workloads that are underserved — how can we build something that will be able to accelerate what's possible?"

By designing its own inference silicon, OpenAI can optimise the chip precisely for its own models, reduce per-query costs, and move closer to profitability — a critical consideration as the company eyes a public offering. If it can drive down the costs of AI inference, it can recoup some of the losses spent on costly training runs.

The Nvidia Question

Jalapeño is not a direct Nvidia replacement — at least not yet. It is likely that more performance-intensive tasks like pre-training will still rely on Nvidia hardware. ASICs are less flexible than GPUs; you cannot retrain a frontier model on Jalapeño the way you can on an H100 cluster.

But inference is where volume lives. And controlling inference silicon means controlling the unit economics of serving AI at scale. OpenAI is attempting to rapidly rewrite its future unit economics of AI.

Broadcom CEO Hock Tan was unambiguous about the ambition: "This is just the beginning of a multi-generation roadmap. By co-developing our industry-leading silicon directly with OpenAI, we are enabling the deployment of gigawatt-scale data centres with Microsoft and other partners beginning in 2026."

The Bigger Picture

OpenAI is not alone in this move. Google has its TPUs. AWS has Inferentia. Meta is building its MTIA chips. Apple designs its own Neural Engines. The pattern is unmistakable: every major AI company is working to reduce its dependence on third-party silicon, and specifically on Nvidia.

Jalapeño is OpenAI's entry into that club. It signals that OpenAI sees itself not just as a software and model company but as a full-stack AI infrastructure player — owning the chips, the data centres, the models, and the products that run on them.

For the broader industry, that shift has profound implications. It means AI compute is becoming increasingly proprietary and vertically integrated. The open, interoperable GPU market that allowed thousands of startups to access frontier compute may gradually give way to a world where the biggest AI labs run on their own silicon — and everyone else competes for what is left.

Published June 27, 2026.