OpenAI unveils custom AI chip Jalapeño

OpenAI’s latest hardware, the Jalapeño chip, is a custom inference processor developed with Broadcom to support its growing AI infrastructure. Though not yet widely deployed, Broadcom’s CEO Hock Tan has drawn parallels between it and Nvidia’s Blackwell chips alongside Google’s tensor processing units.

The push for in-house silicon has gained momentum. Google, Meta, and Amazon have each created their own chips to better manage AI workloads. OpenAI’s decision reflects a wider industry shift from generic hardware to systems designed for specific needs.

“AI applications have become so resource-intensive that companies are now prioritizing customization and deeper integration,” said Alexander Harrowell, senior principal analyst at Omdia. Analysts at the firm have observed this pattern since 2022, describing it as Makimoto’s Wave—a recurring cycle where standardized products yield to specialized designs.

The partnership between OpenAI and Broadcom had been discussed for months before Tan confirmed it during a 2025 earnings call.

Jalapeño is an application-specific integrated circuit (ASIC) built for AI inference, the process of executing trained models to produce responses. Unlike Nvidia’s chips, which handle both training and inference, this processor focuses exclusively on the latter, particularly for prompts sent to ChatGPT.

Richard Simon, CTO at T-Systems International, explained the benefits: lower cost per inference token, improved energy efficiency, reduced latency, and quicker API responses. For OpenAI, the financial upside stands out. Each user prompt carries computational expenses, much of which currently goes to providers like Microsoft, AWS, and Nvidia.

“NVIDIA’s gross margin ranges from 75% to 78%, and that directly affects your bottom line,” Harrowell noted. “Replacing it with Broadcom’s 30%-35% margin for ASIC production cuts those costs nearly in half.”

Energy use also plays a key role. Data center expansions in the U.S. won’t arrive for years, leaving AI providers searching for ways to improve efficiency. Custom chips help by lowering power consumption—Jalapeño operates below 800W, avoiding the need for liquid cooling.

“Tailoring hardware helps control power demands, which drive data center costs,” Harrowell added.

Enterprise users won’t notice immediate changes, as most access AI through APIs and platforms rather than direct hardware. However, these infrastructure choices could influence pricing, performance, and availability over time. Omdia expects ASICs to gain substantial market share by 2027, driven by volume due to their lower costs.

Simon suggested IT leaders would see advantages: “Inference costs will drop as scale increases.” Users may experience better performance at the same or lower prices, thanks to OpenAI’s savings.

Quentin Reul, director of global AI strategy at expert.ai, pointed to another benefit: security. “By creating its own chip and dedicated data centers, OpenAI can minimize the risk of data exposure when sharing information across cloud infrastructure.”

Independent tests of Jalapeño haven’t been conducted, so its actual performance remains unclear. Harrowell mentioned that OpenAI uses the same ASIC manufacturer and server supplier (Celestica) as Google, whose tensor processing units rival Nvidia’s Blackwell chips. Still, the comparison may not hold much weight, since Jalapeño isn’t intended for training.

Reul emphasized the chip’s alignment with OpenAI’s architecture. With Jalapeño, the company has taken a step toward hardware optimized for its needs.