OpenAI's new Jalapeño chip outperforms Nvidia processors in inference benchmarks, delivering 1.5x-1.9x more AI work per watt and 1.7x-3.6x lower latency across multiple models.
OpenAI has unveiled performance metrics for Jalapeño, its custom AI inference chip, claiming significant advantages over existing hardware solutions.
The Application-Specific Integrated Circuit (ASIC) was tested on Semianalysis's InferenceX benchmark against Nvidia chips running GPT-OSS, DeepSeek R1, and Kimi K2.5 1T models. Results show Jalapeño delivered substantially higher efficiency and speed metrics.
"Jalapeño offers the best of both worlds with lower latency and higher throughput," said Richard Ho, OpenAI's hardware vice president. The achievement addresses a longstanding trade-off in AI systems, which typically force choices between response speed or processing capacity.
The chip registered both more tokens per user and more throughput per kilowatt than current state-of-the-art alternatives, according to benchmark data. This positions Jalapeño as a competitive option for organizations running large-scale AI inference workloads.
OpenAI first introduced Jalapeño in June as part of its broader hardware development strategy. The company positions the chip as enabling customers to choose between models offering lower operational costs or faster response times—previously incompatible objectives.
The performance gains carry significant implications for AI infrastructure costs. Lower latency reduces user wait times, while improved efficiency per watt decreases electricity consumption, a major expense for data centers running continuous AI services.
OpenAI's move into custom silicon reflects broader industry trends. Companies including Google, Amazon, and Meta have developed proprietary chips to optimize their AI operations and reduce reliance on Nvidia's dominant GPU market position.
The company has not announced general availability or pricing for Jalapeño. Details on customer access and deployment timelines remain unclear.
Apple's latest desktop computers are built to support local artificial intelligence work, addressing the growing trend of developers using multiple Macs for AI tasks.
Ukraine has granted British firms exclusive access to Avengers Labs, a platform containing roughly five million annotated battlefield images. The partnership marks the first international sharing of Ukraine's combat dataset for autonomous weapons development.
Accelerated Understanding unveiled an enterprise physics AI model using neural operators that processed 5 trillion data pieces in testing. The startup was founded by researchers previously considered to lead Jeff Bezos's Project Prometheus.