A custom AI chip is a purpose-built processor designed from the ground up to handle specific artificial intelligence workloads, rather than relying on general-purpose hardware like traditional GPUs. On June 24, 2026, OpenAI and Broadcom jointly announced “Jalapeño,” OpenAI’s first custom AI inference chip. It’s a landmark move that signals OpenAI’s expansion from a company known for its software models and consumer products into one that now designs the physical hardware powering those products. Built as an application-specific integrated circuit (ASIC) rather than a general-purpose GPU, Jalapeño is tailored to run large language model (LLM) inference workloads more efficiently and at a lower cost than off-the-shelf alternatives.
In this article, we’ll discuss why OpenAI decided to build its own chip, how the Jalapeño was developed in a record-breaking nine months, what it means for the competitive landscape between AI companies and chip giants like Nvidia, and why this move could reshape the economics of running AI at scale. Whether you’re tracking the AI hardware race or trying to understand what drives the cost of your favorite AI tools, this is a story worth following.
TL;DR Snapshot
OpenAI has unveiled Jalapeño, its first custom-designed AI chip developed in partnership with Broadcom. The chip is purpose-built for LLM inference, meaning it’s optimized for the task of running AI models in real-time (such as answering ChatGPT queries or powering Codex). OpenAI says early testing shows the chip delivers significantly better performance per watt than current leading hardware, and Bloomberg reports that Broadcom CEO Hock Tan has indicated cost savings of roughly 50% compared to typical AI GPUs.
Key takeaways include…
- OpenAI is going full-stack: Jalapeño marks OpenAI’s entry into custom silicon, giving the company control over its infrastructure from chip design all the way up to consumer products like ChatGPT and Codex.
- The chip was developed at unprecedented speed: OpenAI and Broadcom took Jalapeño from initial design to manufacturing tape-out in just nine months, a timeline the companies believe is the fastest ASIC development cycle ever achieved in high-performance semiconductors. OpenAI’s own AI models helped accelerate the design process.
- It’s a competitive signal to Nvidia: While OpenAI isn’t abandoning Nvidia GPUs (especially for training), developing its own inference chip gives OpenAI leverage over hardware costs and reduces its dependence on a single supplier at a time when demand for AI compute far outpaces supply.
Who should read this: AI engineers, tech investors, startup founders, semiconductor industry watchers, and anyone curious about the infrastructure behind AI products.
Why OpenAI Is Building Its Own Chips
OpenAI has been one of the world’s largest buyers of Nvidia’s GPUs since the generative AI boom began in late 2022. But as CNBC reported, the company is experiencing such explosive growth in demand that it simply can’t get enough computing power from existing suppliers. OpenAI President Greg Brockman told CNBC that OpenAI “cannot get compute fast enough,” while Broadcom CEO Hock Tan described the compute demand from his largest customers as “simply insatiable.”
That supply crunch is one motivation, but the economics of inference are just as important. Training a new model is a one-time (albeit enormous) cost. Inference, on the other hand, is the ongoing expense of running that model every time a user sends a message to ChatGPT, makes a Codex request, or hits the OpenAI API. As TechCrunch noted, even small reductions in inference costs could significantly improve OpenAI’s bottom line. Jalapeño targets exactly that pain point.
By making their own chip, OpenAI gains what it calls a “full-stack advantage.” Rather than adapting general-purpose hardware to their needs, they can design silicon around the exact workloads they run (e.g. powering specific models, kernels, memory access patterns, and serving systems). As Richard Ho, who leads OpenAI’s hardware program explained in the announcement, the team optimized the chip’s architecture around the patterns that matter most for frontier AI models.
How Jalapeño Was Built in Record Time
Perhaps the most striking detail of the announcement is the development timeline. OpenAI and Broadcom took Jalapeño from a blank-slate design to manufacturing tape-out in just nine months. Both companies believe this is the fastest development cycle ever achieved for a high-performance ASIC.
That speed was possible because of a tight collaboration between three partners. OpenAI handled the core chip design, drawing on its deep knowledge of LLM fundamentals. Broadcom brought silicon manufacturing expertise and networking technology, including its Tomahawk networking chips. Celestica contributed board-level design, rack systems, and integration for scalable production.
Crucially, OpenAI’s own AI models played a role in accelerating the chip design process. As CNBC reported, Brockman said the degree to which their models sped up the process “was very surprising to us.” It’s a compelling feedback loop: the models that OpenAI serves to users are now helping improve the very infrastructure that will run future models.
Engineering samples of Jalapeño are already running ML workloads in the lab at production-target frequency and power levels, including GPT-5.3-Codex-Spark. While final performance numbers haven’t been released yet, OpenAI says early results show substantially better performance per watt than the current state of the art. A detailed technical report is expected in the coming months.
What This Means for Nvidia and the AI Chip Market
Jalapeño isn’t a direct replacement for Nvidia’s GPUs, and OpenAI has been careful not to frame it like one. Axios reported that this first generation of chips targets inference only, not training, and that Nvidia remains a key partner for OpenAI’s model training workloads. OpenAI is also reportedly still considering whether to expand its custom chip program into training down the road.
But the broader significance here is hard to miss. As Yahoo Finance reported, the announcement is part of a growing trend among the largest AI companies and cloud providers to develop their own custom silicon. Google has its TPUs. Amazon has Trainium and Inferentia. Microsoft has Maia. Meta has MTIA. Now OpenAI has Jalapeño. When every major buyer of AI chips is investing in alternatives, it puts long-term pressure on Nvidia’s pricing power, even if its GPUs remain dominant for now.
The deployment plan is very ambitious. OpenAI and Broadcom have committed to a multi-year agreement to co-develop and deploy 10 gigawatts of custom AI accelerators beginning in late 2026 and continuing through 2029. The Decoder reported that Microsoft is expected to purchase roughly 40% of the chips, and Broadcom CEO Hock Tan has said the company’s compute demand remains elevated through at least 2028.
Constellation Research analyst Holger Mueller offered a pointed take in his analysis, stating that Broadcom is the same partner that helped Google build its TPUs, so “the pedigree is there.” He called Jalapeño “a shot in front of the bow for Nvidia.”
The Bigger Picture: AI’s Vertical Integration Era
Jalapeño isn’t just a chip story, it’s a signal that the AI industry is entering an era of vertical integration, where the most competitive companies will control everything from the silicon in the data center to the product on your screen.
OpenAI’s announcement post describes a flywheel: better infrastructure drives compute efficiency, which enables better models, which power better products, which bring in more revenue, which funds the next generation of infrastructure. By owning more of that cycle, OpenAI can iterate faster and capture more of the value it creates.
Broadcom CEO Hock Tan put it more bluntly. According to a report from Futunn, Tan predicted that over time, every frontier model developer will eventually build its own dedicated AI accelerator and networking solution. If that’s true, the chip market is heading toward a world where the biggest AI companies each run on their own custom silicon, and general-purpose GPUs play a still important but increasingly commoditized role.
For now, there are still several major unknowns. OpenAI’s performance claims haven’t been independently verified. The chip hasn’t been deployed at scale. And as The Decoder pointed out, it’s unclear exactly which chips Jalapeño was tested against, on what tasks, and under what conditions. But even with those caveats, Jalapeño represents a clear strategic bet. OpenAI believes the future of AI isn’t just about building the best models, it’s also about building the best infrastructure to run them.
Frequently Asked Questions
OpenAI is the artificial intelligence company behind ChatGPT, Codex, and the GPT family of large language models. Founded in 2015, it has become one of the most prominent AI research labs in the world and one of the largest consumers of AI computing hardware.
Broadcom is a global semiconductor and infrastructure software company. In the AI space, it has become one of the leading partners for companies looking to develop custom AI chips, having previously worked with hyperscalers like Google on custom silicon solutions.
Celestica is a manufacturing and supply chain solutions company that contributed to the Jalapeño project by handling board-level design, rack systems, and scalable production integration. It plays a key role in turning chip designs into deployable data center hardware.
ASIC stands for application-specific integrated circuit. Unlike a general-purpose GPU, which can handle a wide range of computing tasks, an ASIC is designed and optimized for one particular type of workload. ASICs can be more efficient and cost-effective for their target task, but they sacrifice the flexibility that makes GPUs so versatile.
Inference is the process of running an already-trained AI model to generate outputs, such as responding to a ChatGPT message or completing a Codex task. It’s distinct from training, which is the computationally intensive process of building the model in the first place. Because inference happens every time a user interacts with an AI product, its cost and speed are critical to the economics of running AI services at scale.
Tape-out is the milestone in semiconductor development when the finalized chip design is sent to a manufacturing facility (or foundry) for fabrication. It marks the transition from the design phase to the production phase and signals that the chip is ready to be physically produced.
In this context, a full-stack strategy means that OpenAI is aiming to control and optimize every layer of the technology needed to deliver its AI products. That includes the consumer-facing applications (like ChatGPT), the AI models powering them, the software systems that serve those models, and now the physical chips that run those systems. The goal is to improve performance and reduce costs by integrating all of these layers.
Gigawatt-scale refers to data center infrastructure that consumes one or more gigawatts of power. To put that in perspective, a single gigawatt can power roughly 750,000 homes. The term highlights the extraordinary energy requirements of frontier AI and signals the industrial scale at which companies like OpenAI and its partners are planning to build future AI infrastructure.
Other Enterprise AI Articles You May Be Interested In
Oracle Cuts 21,000 Jobs as AI Transforms the Tech Workforce
Public Ownership of Big AI? Breaking Down Sanders’ Bold New Bill
SpaceX Acquires Cursor for $60 Billion: What Developers Need to Know
Fable 5 and Mythos 5 Shutdown: Why the U.S. Government Pulled Anthropic’s Most Powerful AI Offline
OpenAI Weighs Drastic Price Cuts as the AI War With Anthropic Heats Up
