The chip, called Jalapeño, was developed with Broadcom as part of OpenAI’s push to gain greater control over the computing infrastructure powering ChatGPT and its other artificial intelligence services. Tests released on Tuesday showed the processor delivering between 1.5 and 1.9 times more AI inference work per watt than systems based on Nvidia’s GB200 and GB300 technology.
OpenAI also said Jalapeño reduced end-to-end response latency by between 1.7 and 3.6 times across several large language model workloads. The comparison covered models including GPT-OSS 120B, DeepSeek R1 and Kimi K2.5, with testing focused on the process of running already-trained models rather than training them.
The results mark the first detailed performance disclosure since OpenAI unveiled Jalapeño in June. At that time, the company said laboratory testing indicated substantially better performance per watt than state-of-the-art alternatives but did not publish direct comparisons with Nvidia hardware.
Jalapeño is designed specifically for inference, the computing stage during which a trained AI model generates answers, images, code or other outputs for users. That makes it different from the broader range of graphics processors commonly used both to train models and operate them after training.
The distinction is important because OpenAI is not claiming that Jalapeño can replace Nvidia processors across its entire computing operation. The company continues to depend heavily on Nvidia hardware for training large frontier models, while its own chip is intended to reduce the cost and power requirements of serving growing volumes of AI requests.
OpenAI Hardware Vice-President Richard Ho said the architecture was built to combine high throughput with low latency, two performance goals that often require compromises in conventional accelerator designs. The company has also emphasised reduced movement of data between computing and memory components as a central part of the design.
The processor has a rated thermal design power of 700 watts. During workloads demonstrated at the Hot Chips industry conference, sustained measured power consumption remained at or below about 550 watts, strengthening OpenAI’s argument that specialised hardware could improve the economics of large-scale AI inference.
Jalapeño was developed from initial design to manufacturing tape-out in about nine months, with OpenAI saying its own AI models helped engineers accelerate parts of the chip development and optimisation process. Broadcom supplied semiconductor implementation expertise and networking technology, while Celestica has been involved in board, rack and system integration.
OpenAI and Broadcom previously announced plans to deploy 10 gigawatts of OpenAI-designed accelerators and associated networking systems over several years. Initial Jalapeño deployments are expected by the end of 2026, although volumes are likely to remain limited at first before production expands during 2027.
The company does not plan to sell the chip commercially. Instead, Jalapeño will be deployed within OpenAI’s infrastructure, where specialised processors could lower the cost of answering user queries and reduce pressure on electricity and data-centre capacity.
The move reflects a broader shift among major artificial intelligence companies towards custom silicon. Google has developed Tensor Processing Units, Amazon operates its Trainium and Inferentia chips, and Microsoft has built its Maia accelerators. Custom processors allow technology groups to optimise hardware around specific models and workloads rather than rely entirely on general-purpose accelerators.
Nvidia nevertheless remains the dominant supplier of advanced AI processors and continues to benefit from an entrenched software ecosystem, particularly its CUDA platform. Its Blackwell generation has become a core component of major AI data centres, while the company is preparing broader deployment of its next-generation Vera Rubin systems.
OpenAI’s benchmark therefore represents a challenge to Nvidia in a specific part of the AI computing market rather than evidence that the company has displaced Nvidia across training and inference. Benchmark outcomes can also vary significantly depending on model architecture, batch sizes, software optimisation, memory configuration and the way latency is measured.
The Jalapeño tests used workloads designed to resemble large-scale production inference and included comparisons involving Nvidia GB200 and GB300 systems. Independent benchmarking work has increasingly focused on measures such as tokens generated per second, response latency, power consumption and the cost of operating large models under different traffic levels.
OpenAI’s growing hardware effort also reflects the enormous computing requirements created by expanding use of ChatGPT, coding agents and other AI products. The company has been building partnerships around data centres, networking, power supply and semiconductor manufacturing as it seeks enough capacity to support more capable models and higher user demand.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
