OpenAI Just Built Its Own AI Chip — And It’s Called Jalapeño

The Chip That Could Change Everything About AI’s Economics

For years, OpenAI has been one of Nvidia’s most important customers. Every ChatGPT query, every image generated by DALL·E, every line of code written by Codex — all of it has run on Nvidia GPUs, the gold-standard hardware that powers the modern AI industry. That relationship has made Nvidia one of the most valuable companies in the world and given it remarkable leverage over every major AI lab.

On Wednesday, OpenAI took a significant step toward changing that dynamic.

The company unveiled Jalapeño — its first custom-built inference processor, designed in collaboration with semiconductor giant Broadcom. It is a chip built by OpenAI, for OpenAI, optimized specifically for the unique demands of its AI systems. And if early results hold up, it could meaningfully shift the economics of running the world’s most widely used AI products.

This is a big deal — not just for OpenAI, but for the entire AI industry. Here’s everything you need to know.

What Is the Jalapeño Chip?

Jalapeño is OpenAI’s first custom silicon — a purpose-built AI chip designed from the ground up to handle inference workloads. The chip was created in collaboration with Broadcom, one of the world’s leading semiconductor companies, and the partnership was officially announced back in October.

What makes Jalapeño notable is not just that it exists, but how it was built. OpenAI says its own AI models actively assisted in the chip’s design process — an early and striking example of AI being used to design the very hardware it will eventually run on.

The chip is still in testing, but OpenAI reports that early results are promising: Jalapeño is showing significantly better performance-per-watt than current state-of-the-art alternatives. In the world of data center economics, where electricity costs are one of the largest line items, performance-per-watt is not a trivial metric — it is one of the most important measures of a chip’s real-world viability.

Why Did OpenAI Build Its Own Chip?

Running AI models at scale is extraordinarily expensive. Every time a user sends a message to ChatGPT, a cluster of Nvidia GPUs springs to life to process the query and generate a response. Multiply that by hundreds of millions of daily queries, and you’re looking at a hardware bill that runs into the billions of dollars annually.

Nvidia’s GPUs are exceptional general-purpose AI accelerators, but “general purpose” is also their limitation. When you know your workload inside and out — as OpenAI does with its own models — there is a compelling case for building something more targeted.

OpenAI president Greg Brockman articulated this logic on the company’s in-house podcast: “We have a deep understanding of the workload. We’ve really been looking for specific workloads that are underserved, and asking: how can we build something that will be able to accelerate what’s possible?”

What Is Inference, and Why Does It Matter?

Jalapeño is specifically designed for inference — the process of running a pre-trained AI model in response to a user command.

Pre-training is the computationally intensive process of teaching a model by exposing it to vast amounts of data. Training a frontier model like GPT-4 costs hundreds of millions of dollars and runs for months on thousands of GPUs. Inference, by contrast, is what happens after training — when the model responds to real users in real time. Every single user query triggers an inference operation.

At the scale OpenAI operates, even modest improvements in inference efficiency translate to enormous savings. A chip that performs the same inference work while consuming 20% less power could save the company hundreds of millions of dollars per year. OpenAI specifically highlighted Jalapeño’s performance when running real-time coding models — central to products like Codex and the company’s push into agentic AI.

OpenAI Is Not Alone: The Custom Chip Race

OpenAI’s move into custom silicon is not without precedent. Google’s TPUs (Tensor Processing Units) have powered the company’s AI workloads for nearly a decade, from Google Search to the training of Gemini. Amazon’s Trainium is AWS’s answer to the same challenge — a custom chip designed to accelerate machine learning on Amazon’s cloud infrastructure.

Both companies have found that custom silicon, despite the enormous upfront investment, pays off at sufficient scale. The message is clear: at the frontier of AI, owning your silicon is increasingly a strategic necessity, not a luxury.

AI Designing Its Own Hardware: A New Frontier

One of the most fascinating details in OpenAI’s announcement is the role its own AI models played in designing the chip. Chip design has traditionally been one of the most demanding engineering disciplines in the world. Using AI to assist in chip design is an emerging practice, but OpenAI’s use of its own frontier models to help build the chip those models will run on is particularly striking.

It is an early, tangible example of recursive self-improvement — AI systems contributing to better AI infrastructure, which in turn enables more capable AI systems.

The Bigger Picture: OpenAI’s Full-Stack Ambition

Jalapeño is the latest expression of a much larger strategic vision. OpenAI’s own announcement made this explicit: “OpenAI is not only developing frontier models or building products on top of them; it is designing the infrastructure underneath them: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience.”

When you control every layer of the stack, you can optimize each layer around the same goal. A chip designed with specific knowledge of the model architecture it will run, feeding into a memory system designed with knowledge of that chip, scheduled by software with knowledge of both — the cumulative efficiency gains could be enormous.

“Because OpenAI operates across the stack,” the company wrote, “each layer can be optimized around the same goal: making its models faster, more reliable, and more affordable for users.”

What Does This Mean for Nvidia?

The honest answer: not immediately, and probably not dramatically, but the directional pressure is real.

OpenAI is not trying to replace Nvidia across the board. Jalapeño is an inference chip — designed for a specific, high-volume workload. The most computationally demanding tasks, like pre-training frontier models, will almost certainly continue to rely on Nvidia’s highest-end hardware for years to come.

But inference is where the volume is. Every query, every API call, every ChatGPT conversation — that’s all inference. If OpenAI can serve those workloads at meaningfully lower cost using its own chip, the number of Nvidia GPUs it needs to purchase goes down. The more significant long-term risk for Nvidia is if Jalapeño succeeds and accelerates a broader industry shift toward custom silicon.

What Does This Mean for Users?

For everyday users of ChatGPT, Codex, and OpenAI’s other products, the near-term impact of Jalapeño is likely to be invisible. But the medium and long-term implications are real. Cheaper inference costs mean OpenAI can offer more capabilities at lower price points, faster response times, and more investment in developing new models rather than routing capital to hardware vendors.

If Jalapeño delivers on its early promise, the beneficiaries will ultimately be the millions of people and businesses that use OpenAI’s products every day — even if they never know a chip named Jalapeño had anything to do with it.

What Comes Next?

Jalapeño is still in testing, and OpenAI has not announced a production timeline. The chip will need to pass further validation before being deployed at scale, and the company will likely run it in parallel with Nvidia hardware for some time.

But the direction of travel is clear. OpenAI has shown its hand: it intends to build not just the best AI models, but the best infrastructure to run them. Jalapeño is the beginning, not the end. Expect more custom silicon to follow.

Jalapeño is currently in testing. OpenAI has not announced a production deployment timeline. This article is based on OpenAI’s official announcement and publicly available information.