OpenAI turns up the heat in chip market with Jalapeño

ChatGPT maker OpenAI has stepped beyond the world of models by announcing its first-ever custom-built chip named Jalapeño.

This piece of hardware is what OpenAI is calling an Intelligence Processor, which is a specialised type of computer chip designed from the ground up to handle the specific needs of large language models (LLMs).

What is Jalapeño?

In the world of artificial intelligence, there are two main stages of work. The first is training where a model learns from vast amounts of data. The second is inference where the model uses what it has learned to answer a user’s prompt.

While training has often received more attention, inference is what happens every time someone uses a tool like ChatGPT. It is the part of the process that must happen constantly and at a massive scale.

Jalapeño is an Application-Specific Integrated Circuit (ASIC), which is a chip built for one particular job: to make the inference process as fast and cheap as possible.

The development marks a significant step in OpenAI’s broader AI strategy.

For much of its history, the company relied on third-party hardware from companies like NVIDIA or used cloud infrastructure from partners like Amazon, Google and Microsoft. By designing its own silicon, OpenAI is following a path of vertical integration, wherein it wants to control every level of the technology stack from the application that users see down to the microscopic circuits in the data centre.

This strategy allows the ChatGPT-maker to reduce its dependence on firms like NVIDIA which currently dominates the market for AI chips.

Speed through collaboration

OpenAI and its partner Broadcom said they managed to go from the initial design to the tape-out stage in just nine months. In the semiconductor industry, a tape-out is the final blueprinting stage before a chip is physically manufactured in a factory.

This is believed to be one of the fastest development cycles ever seen for high-performance hardware. Interestingly, OpenAI used its own AI models to help speed up the design and optimisation of the chip, creating a sort of digital loop where current AI is being used to build the better hardware that will run the next generation of AI.

The partnership behind this chip is a complex web of industry leaders.

While OpenAI provided the vision and the specific requirements for how the chip should handle language, Broadcom provided the technical expertise to turn those ideas into silicon. In the custom chip world, Broadcom helps major tech firms build their own hardware without them needing to start a chip company from scratch.

Once the design was finished, the manufacturing was handled by TSMC which is a specialist factory in Taiwan that produces the world’s most advanced processors. To turn these chips into working servers, a company called Celestica was brought in to build the physical boards and racks.

Finally, these systems will be deployed in data centres managed by Microsoft and other partners starting in late 2026.

Efficiency and global scale

From a technical perspective, the main goal of Jalapeño is efficiency. Early lab tests suggest that the chip will offer a performance-per-watt that is substantially better than the current top-of-the-line hardware.

Performance-per-watt means how much work a computer can do for every unit of electricity it consumes. Because massive AI data centres use a staggering amount of power, even small improvements in efficiency can save billions of dollars.

In fact, some early estimates suggest that Jalapeño could lead to a cost reduction of roughly 50% compared to using standard off-the-shelf graphics chips. The architecture of the chip was designed to reduce how much data has to move around inside the processor which is often a major bottleneck that slows down AI responses.

OpenAI is not the first company to realise that custom hardware is a strategic necessity. The landscape for AI chips is already crowded with major players who have similar programmes. Google is perhaps the veteran in this field as it has been developing its own Tensor Processing Units (TPUs), for nearly a decade. Amazon has its own lines called Trainium and Inferentia while Microsoft recently introduced its Maia chips. Meta, the parent company of Facebook, is also working on its own hardware known as the Meta Training and Inference Accelerator (MTIA).

When compared to its direct rival Anthropic, OpenAI appears to be taking a more hands-on approach to hardware. While OpenAI is designing its own proprietary silicon, Anthropic has focused on securing massive amounts of existing compute power through partnerships. For example, Anthropic committed $21 billion to a deal that gives it access to Google’s TPU units. While there are reports that Anthropic might be exploring its own designs, OpenAI is the one that has now publicly announced a finished and physical processor.

Scaling intelligence

The timing of the Jalapeño announcement is also important when looking at the broader financial state of OpenAI. In the first half of 2026, the company underwent a period of massive growth and fundraising. It raised approximately $122 billion in capital which gave it a valuation of about $852 billion. The company has also confidentially filed for an initial public offering to become a publicly traded firm.

All of this money is being funneled into a long-term plan to build a massive infrastructure for AI. Jalapeño is a key part of this future because it provides the efficient engine that these massive data centres will need.

The ultimate goal of this hardware push is to make advanced AI more accessible to the average person. As OpenAI co-founder and president Greg Brockman, notes, “The world is moving to a compute-powered economy.”

“Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant, resulting in AI which is faster, more reliable, more affordable for people and businesses, and can be used to solve more important problems. By designing more of the stack ourselves, we can serve more intelligence with greater efficiency,” he adds.

If inference becomes cheaper and faster, then the tools built on top of it can become more capable. This could lead to ChatGPT evolving into a superapp that can handle complex tasks, act as a personal agent, and even help with professional software engineering through platforms like Codex.

In the long run, OpenAI does not view Jalapeño as a one-time project. It has described the chip as the first step in a multi-generation platform. This means that as AI models continue to change and become more complex, the hardware running them will be updated in lockstep. By owning the design of the chip, OpenAI can in a way ensure that its future models, such as the upcoming GPT-5.3-Codex-Spark which is already running on Jalapeño samples in the lab, have exactly the hardware they need to perform at their peak.

While the technical details of the chip’s performance will be released in a more formal report in the coming months, the strategic message is already clear.

The Sam Altman-led firm that once only aimed to build safe and beneficial AI is now aiming to build the very machines that make that intelligence possible. If the early testing holds true and the chip delivers on its promise of greater efficiency, the result for the end user will be a faster and more reliable experience.

The move to create Jalapeño emphasises that in the technology industry, the most powerful and efficient systems will be the ones where the software is married most closely to its hardware, something that we have seen tech giants like Apple do for decades now.