OpenAI has revealed the first detailed performance results for its custom Jalapeño AI chip, showing an ambitious push to make artificial intelligence inference faster and more energy efficient while giving the company greater control over the computing infrastructure behind its models.
The company presented the results at the Hot Chips conference on Tuesday, offering a closer look at the processor it developed with Broadcom. Jalapeño is OpenAI’s first custom inference chip and is designed specifically to run trained AI models rather than train them.
Early benchmarks suggest the strategy could deliver significant performance gains.
Testing using SemiAnalysis’s InferenceX benchmark showed Jalapeño delivering higher peak throughput per kilowatt and lower token latency than commercially available systems included in the comparison. OpenAI said the processor also performed strongly across multiple model families, including GPT-OSS 120B, DeepSeek R1 and Kimi K2.5.
TechCrunch reported that the benchmark comparison included an Nvidia Blackwell system, giving Jalapeño an important early test against hardware from the company that dominates the AI accelerator market.
OpenAI Pushes Deeper Into AI Infrastructure
Jalapeño is designed around inference, the process that occurs when an already trained AI model receives a request and generates a response. As services such as ChatGPT grow, inference becomes an increasingly important part of the cost and energy required to operate AI systems at enormous scale.
OpenAI says Jalapeño addresses a common challenge in inference hardware, where systems can be forced to trade high throughput for low latency. The company says its architecture is designed to improve both simultaneously by reducing unnecessary data movement and communication delays.
More detailed specifications presented at Hot Chips underline the scale of the system. Jalapeño is rated at 700 watts, although measured sustained power remained at or below 550 watts during tested workloads, according to Data Center Dynamics. OpenAI plans racks containing 128 chips, with a full pod consisting of 2,048 ASICs.
The project is also significant because OpenAI developed Jalapeño with Broadcom as part of a broader effort to build hardware specifically around its own AI workloads. OpenAI says future generations of the platform are already under development.
That approach could give the company more control over the economics of operating increasingly powerful AI services. Rather than relying entirely on general-purpose accelerators supplied by outside partners, OpenAI can optimize chips, memory, networking, serving software and models together.
The move does not mean OpenAI is abandoning Nvidia. The company has described its first-party silicon as another path alongside accelerators supplied by its partners. Jalapeño is also focused on inference, meaning OpenAI will continue to rely on external computing infrastructure for other workloads, including frontier model training.
Deployment Begins as the Chip Race Accelerates
OpenAI expects Jalapeño to enter deployment in very small volumes toward the end of 2026, with a more significant rollout planned for 2027, according to TechCrunch.
That timetable leaves room for the competitive landscape to change. Nvidia and other semiconductor companies are continuing to develop new generations of AI hardware, meaning today’s benchmark advantage does not guarantee that Jalapeño will lead once it reaches large-scale deployment.
Still, the processor represents an important strategic shift for OpenAI. The company is increasingly positioning itself not simply as an AI model developer, but as a vertically integrated technology company spanning models, software, data centers and custom silicon.
If Jalapeño’s early performance translates successfully into large-scale deployment, OpenAI could reduce the power and infrastructure required to serve growing AI workloads while gaining more control over one of the most expensive components of the AI business.
The battle to build the world’s leading AI systems is increasingly becoming a battle over the chips underneath them.
