OpenAI's first custom chip Jalapeño is here — built from scratch just for LLM inference!

Hi everyone, it's Shiichan! Today's news is a big one: OpenAI has built its very own chip! OpenAI is best known as a software company, so seeing it step into hardware is really exciting.

OpenAI News openai.com

What was announced?

Over on OpenAI's News, the company introduced "Jalapeño," a chip built specifically for LLM inference. It's OpenAI's first in-house chip design, and the company calls it an "Intelligence Processor."

It wasn't built by OpenAI alone. It's a joint effort with Broadcom, which handled silicon implementation, networking, and connectivity, and Celestica, which contributed the board, rack, and system work. The announcement came on June 24, 2026.

Why it matters

Today, most AI runs on Nvidia GPUs for both training and inference. GPUs are general-purpose, though, so for a well-defined workload like LLM inference, a purpose-built chip has room to run things more efficiently.

OpenAI chose that purpose-built path. Google's TPUs and Amazon's own accelerators show that big players building custom AI chips is already a trend, and OpenAI is now part of it.

OpenAI's Greg Brockman put it this way:

We have a deep understanding of the workload. We've really been looking for specific workloads that are underserved.

Because OpenAI understands its own workloads deeply, it can target the parts that aren't yet well optimized and build a chip just for them.

What changes

OpenAI owns the whole stack, from model design to kernels to serving, and it poured all of that knowledge into Jalapeño. So this isn't a training chip repurposed for inference, and it isn't a general-purpose AI processor. It's a design created from scratch for LLM inference.

The exciting part is efficiency. Early testing suggests performance per watt substantially better than the current state of the art. For a company running inference at massive scale, that means both more performance and lower cost and power.

One note: this chip handles inference only. Heavy work like pre-training a model from scratch is expected to keep running on Nvidia hardware.

Dive Deep

Jalapeño is a large, reticle-sized ASIC (an application-specific integrated circuit). From the images shared, the package appears to hold one large compute chiplet surrounded by six HBM modules, plus another chiplet that likely handles input/output (the detailed specs aren't official yet).

The development speed is the surprising part. From initial design to manufacturing tape-out, the team got there in nine months, which is described as possibly one of the fastest cycles ever for high-end semiconductors.

There's real hardware behind it, too: engineering samples are already running ML workloads in the lab at production target frequency and power, including GPT-5.3-Codex-Spark. A detailed technical report on performance is planned for the coming months.

And Jalapeño isn't a one-off. It's the first step in a multi-generation compute platform. Initial deployment is planned for the end of 2026, expanding into gigawatt-scale data centers with Microsoft and other partners.

Wrap-up

• OpenAI announced Jalapeño, its first in-house chip, an ASIC built specifically for LLM inference.
• It's a joint effort with Broadcom (silicon and networking) and Celestica (systems).
• It's a large, reticle-sized chip, taken from design to tape-out in nine months.
• Engineering samples are already running workloads like GPT-5.3-Codex-Spark, with performance per watt expected to be well ahead of the state of the art.
• It's inference-only, with training staying on Nvidia, and it rolls out from late 2026 at gigawatt scale with Microsoft and others.

You can't go and try this one yourself, but it's a landmark for the "optimize the entire stack in-house" trend in AI infrastructure. If you care about the hardware, cost, and power efficiency behind AI, this one's for you!