OpenAI's First Custom Chip 'Jalapeño' Is Real: What Broadcom's LLM Inference Accelerator Means

What Happened

OpenAI has a chip of its own. On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, the company's first custom AI accelerator — a piece of silicon designed from the ground up to run large language models rather than a repurposed graphics processor. Broadcom President and CEO Hock Tan and President Charlie Kawwas handed the first chip to OpenAI CEO Sam Altman and President Greg Brockman, a symbolic moment in OpenAI's push to control more of the hardware its models run on.

OpenAI calls Jalapeño an "Intelligence Processor," and it is the first in a planned family of homegrown chips. Unlike the training-focused hardware that dominates today's AI arms race, this first generation targets inference — the everyday work of running a finished model to answer user queries. In plain terms, Jalapeño is being built to make ChatGPT and OpenAI's other products faster and cheaper to serve at scale, and the companies are aiming to have it handling real traffic by the end of 2026.

What Jalapeño Actually Is

Jalapeño is an application-specific integrated circuit (ASIC) — a chip hard-wired for a narrow job. That makes it less flexible than a general-purpose GPU like Nvidia's, but also cheaper to run and far more efficient at the one thing it is built to do: serve large language models. Where a GPU has to be good at everything, an ASIC can be shaped around exactly how a specific class of models moves data and computes.

Crucially, OpenAI describes Jalapeño as a blank-slate design for modern LLM inference, not an older AI accelerator adapted to new workloads. OpenAI's researchers fed detailed insights about how frontier models actually behave into the design, so the chip is optimized around the things that matter most in production: compute kernels, memory movement, networking between chips, and the serving patterns that make a chatbot feel responsive. Early testing, the companies say, shows Jalapeño delivering performance per watt substantially better than current state-of-the-art hardware — a big deal when energy is the single largest constraint on scaling AI.

The economics are the headline. Broadcom CEO Hock Tan says Jalapeño can run inference for roughly half the cost of a typical AI GPU. At the volume OpenAI operates — hundreds of millions of users sending queries every day — even a fractional cut in the cost per token compounds into enormous savings on compute and electricity.

Nine Months From Design to Tape-Out

One of the most striking claims is the speed. Jalapeño went from initial design to manufacturing tape-out — the point where a finished chip design is sent to the foundry for production — in just nine months. The companies believe that is among the fastest ASIC development cycles ever achieved in high-performance, advanced semiconductors, a process that traditionally takes well over a year for chips of this complexity.

Part of that speed came from deep software-and-hardware co-development between OpenAI's engineers and Broadcom's silicon team. But OpenAI also says it used its own AI models to accelerate parts of the design and optimization work — the same models it serves to users helped build the infrastructure that will run future models. It is a neat, slightly recursive illustration of where the industry is heading: AI helping to design the chips that will run AI.

The 10-Gigawatt Partnership

Jalapeño is the first concrete product from a much larger arrangement. In October 2025, OpenAI and Broadcom announced a strategic collaboration to develop and deploy 10 gigawatts of OpenAI-designed AI accelerators — an amount of computing capacity so large it is measured in the power a data center draws rather than in chip counts. OpenAI designs the accelerators and the systems around them; Broadcom builds them and supplies the networking to tie them together.

Broadcom's role goes well beyond the chip itself. The full racks of accelerators will be scaled entirely with Ethernet and other Broadcom connectivity, which is central to how thousands of chips work together as one system. Deployment of the accelerator and network racks is targeted to begin in the second half of 2026 and run through the end of 2029, spread across OpenAI's own facilities and partner data centers. Hock Tan framed it as "a fundamental commitment to scaling the physical infrastructure required for the next decade of AI," including gigawatt-scale data centers built with Microsoft and other partners starting in 2026.

Building the Full Stack

The strategic logic is that OpenAI wants to own more of the "full stack" beneath its models — from the silicon up through the software. By embedding what it has learned building frontier models directly into the hardware, OpenAI can tune every layer of the system to its own workloads instead of accepting the trade-offs of off-the-shelf chips.

Tan put the reasoning bluntly: "At the end of the day, you cannot, should not rely on some other third-party GPU to do it for you, because it's such a key part." It is the same vertical-integration playbook that companies like Apple, Google, and Amazon have run for years — designing custom silicon to gain control over cost, performance, and supply. For OpenAI, whose entire business depends on the price and availability of compute, owning the chip is arguably the most strategic move it can make.

What It Means for Nvidia

The unavoidable subtext is Nvidia, whose GPUs power the overwhelming majority of AI training and inference today and whose scarcity and price have defined the economics of the entire field. OpenAI is one of the largest consumers of AI compute on the planet, so any move it makes to build alternatives is watched closely across the industry.

It is important to be precise about what Jalapeño is and is not. It is an inference chip, not a training chip, so it does not directly replace the GPUs OpenAI uses to build new models — at least not in this first generation. And OpenAI has been careful to frame Jalapeño as complementary to its ongoing purchases of Nvidia and AMD hardware, not a wholesale replacement. But the direction is clear: the biggest AI labs increasingly want custom silicon for the high-volume, cost-sensitive work of serving models, and every accelerator OpenAI deploys is one it does not have to buy from a third party.

Deployment and What's Next

OpenAI plans to start using Jalapeño to handle customer queries later in 2026, with capacity expanding in the years that follow as the broader 10-gigawatt build-out proceeds. Because Jalapeño is described as the first chip in a multi-generation platform, future versions are expected to broaden its capabilities over time.

For end users, none of this will be visible directly. There is no product to buy and no setting to change. What it should translate into over time is faster responses, higher usage limits, and lower prices as OpenAI's cost to serve each query falls. The heavy engineering happens in the data center; the payoff shows up as a smoother, cheaper ChatGPT.

Why It Matters

Jalapeño is more than one company's chip announcement — it is a marker of how the AI industry is maturing. The first phase of the boom was about who could get their hands on the most GPUs. The next phase is about who can run those models most efficiently, and that increasingly means custom silicon designed around a specific set of workloads.

If OpenAI can genuinely serve inference at half the cost of a standard GPU, it changes the math on everything from free-tier access to the viability of ever-larger models in production. It also deepens the ties between OpenAI, Broadcom, and Microsoft into a hardware supply chain that competitors will have to answer. The chip won't be handling your ChatGPT queries just yet, but when it does, it will quietly reshape the economics that decide how far AI can scale. The full announcement is published by Broadcom.