OpenAI wants to claim more of the AI stack with Jalapeño, its first custom chip

OpenAI on Wednesday announced Jalapeño, its first custom inference accelerator, co-developed with Broadcom and supported by Canadian electronics manufacturer Celestica, and the first step in its multi-generation compute platform. 

The AI company says Jalapeño was designed to work with all large language models (LLMs) and will help make AI faster, better, and cheaper. Behind that rosy mission, OpenAI isn’t shy about its desire to own the full AI stack, something more AI giants are already leaning into.  

“Those serious about platforms should be serious about silicon.”  

As Ben Bajarin, CEO and principal analyst at consumer technology research firm Creative Strategies, posted on X: “Those serious about platforms should be serious about silicon.”  

But with few technical details released, developers are left wondering if OpenAI’s widening footprint will be empowering or restrictive. 

Get in, Big Tech. We’re all building in-house chips now. 

OpenAI isn’t the only Big Tech name to mint its own AI chips. 

Way back in 2016, Google designed and built its own custom hardware for TensorFlow, its machine learning software, the Tensor Processing Unit (TPU). A couple of years later, Amazon debuted AWS Inferentia, its first purpose-built chip for AI and ML. Trainium then hit the scene in 2022, shortly followed by Microsoft’s Azure Maia AI Accelerator in 2023. And nothing is certain yet, but in April, Reuters reported Anthropic is contemplating designing its own chips, though the AI company remains noncommittal for now, at least publicly. 

Why is everyone jumping on the custom-chip bandwagon? 

Blame the compute gold rush, as AI companies increasingly clamor for compute power — ”a compute-powered economy,” as Greg Brockman, president, chairman, and co-founder, OpenAI, puts it. And the numbers surely back it; Stanford’s 2025 AI Index Report says, “training compute doubles every five months.”

While building custom AI chips in-house doesn’t completely alleviate compute pressures, it is one way for OpenAI and its Big Tech brethren to expand compute capacity while potentially lowering costs and reducing reliance on third-party suppliers. 

Exciting claims, no proof

In its announcement blog post, OpenAI describes its new chip as “designed to be the best inference platform for LLMs.”

Specifically, Richard Ho, head of hardware at OpenAI, states: 

“We optimized the architecture around the kernels, memory movement, networking, and serving patterns that matter most for frontier AI models. Based on early testing, Jalapeño will efficiently execute our most important workloads close to the hardware’s theoretical limits.”

But the AI company remains tight-lipped on any real technical details. 

While it claims current tests put Jalapeño’s performance “substantially better than current state-of-the-art,” it doesn’t provide benchmarks to back that up. Instead, it tells developers to expect a detailed technical report “in the coming months.”

What OpenAI does divulge is that engineering samples of the chip are currently running on ML workloads in its lab, including GPT-5.3-Codex-Spark.

Will Jalapeño serve developers, or is OpenAI’s desire to own the AI stack? 

OpenAI makes no qualms about its quest for full-stack control. In doing so, the AI company claims it will make its models “faster, more reliable, and more affordable for users.”

Its logic goes a little something like this: Better infrastructure means more efficient compute, which means better training, which means better models, which means better products, which means more revenue. Then, it explains, it can reinvest that revenue in its infrastructure to make intelligence better for everyone.

But given how little the AI company has revealed about the chip’s specs, it seems developers will have to sit back and watch where the chips fall. 

Jalapeño, then, is simply the next move in OpenAI’s quest to control the whole AI chessboard, moving beyond models and products to the underlying infrastructure itself. 

For developers, OpenAI seems adamant on insisting its full-stack strategy will lead to better performance and pricing for everyone and ultimately empower “anyone trying to learn, create, or solve hard problems.” Still, it’s worth considering: As OpenAI’s grip tightens, will developers become beholden to its ecosystem? 

Several times in its announcement, OpenAI reiterates that it designed Jalapeño for current and future LLMs — all of them. But given how little the AI company has revealed about the chip’s specs, it seems developers will have to sit back and watch where the chips fall. 

Built fast with a long roadmap ahead

The few behind-the-scenes details OpenAI does choose to share boast about its development speed, stating it brought Jalapeño from design to manufacturing tape-out in nine months — “what we believe to be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors.”

The AI company chalks up that fast timeline, in part, to its own models accelerating parts of the design and optimization processes. 

Looking ahead, Jalapeño is slated for deployment at a gigawatt scale in Microsoft’s and other partners’ data centers by the end of the year. 

That’s just the beginning. OpenAI hints at an upcoming multi-generation roadmap, posing the question: What will it seek to control next?