On August 25, OpenAI released initial performance metrics for its custom inference chip, Jalapeño. In the company's tests, the accelerator demonstrated 1.5 to 1.9 times more computational work per watt and 1.7 to 3.6 times lower latency compared to Nvidia's GB200 and GB300 systems.

Since announcing Jalapeño, our first custom inference chip, we’ve been testing it and the system around it.

The results show a major advance: more intelligence from every watt and faster responses, delivering both higher throughput and lower latency in one architecture without… pic.twitter.com/vj7VOrA8pP

— OpenAI (@OpenAI) August 25, 2026

For highly interactive workloads, Jalapeño's performance outstripped competitors by 2.1 to 4.1 times. OpenAI aims to deploy the chip within its infrastructure by the end of 2026.

Jalapeño Tested Against Three AI Models

The company utilized the InferenceX benchmark from SemiAnalysis to evaluate the accelerator, which measures the entire request processing cycle while assessing throughput, power consumption, and latency concurrently.

In tests with GPT-OSS 120B, Jalapeño was compared to Nvidia's GB200. The custom chip achieved approximately 1.9 times greater peak performance per kilowatt: 85,448 versus 44,960 mixed TPS/kW. The total latency was recorded at 1.03 seconds compared to 1.8 seconds.

For DeepSeek R1 670B, the competitor was GB300. Jalapeño delivered 19,641 mixed TPS/kW against 11,781, marking a roughly 1.7 times advantage. Its latency was 3.6 times lower: 1.65 seconds versus 5.99 seconds.

With Kimi K2.5 1T, OpenAI's chip achieved 18,195 mixed TPS/kW compared to GB300's 11,862. The latency was 1.56 seconds versus 5.31 seconds, yielding advantages of approximately 1.5 and 3.4 times, respectively.

Source: OpenAI.

In terms of power consumption, OpenAI used the manufacturers' specified power ratings for the accelerators. Jalapeño is rated at 700 watts, while GB200 is at 1200 watts and GB300 at 1400 watts. The company noted that the actual sustained power consumption of its chip in tests did not exceed 550 watts.

AI Aided in Chip Design

OpenAI employed its AI models in the development of Jalapeño. According to the developers, this approach facilitated a journey from initial design to final production handoff in just nine months. Neural networks were involved in exploring implementation options, validating solutions, and optimizing arithmetic units.

AI tools continued to be utilized post-chip creation. Using Codex based on GPT-Astra, engineers optimized three open-weight models for Jalapeño in under two months, models that were not initially part of the production plan.

For specific attention and mixture-of-experts blocks in GPT-OSS, the AI-generated implementations were found to be 1.5 to 1.8 times faster than those crafted by engineers. OpenAI emphasized that this advantage only applies to selected components, not the entire model.

OpenAI Enhances Its AI Stack

In a separate publication regarding computational strategy, OpenAI's CFO Sarah Friar described Jalapeño as part of a broader vertically integrated infrastructure. This includes data centers, processors, memory, networks, AI models, software, and end products.

Friar stated that the custom chip allows OpenAI greater control over inference costs and enables the selection of hardware for specific tasks. The company plans to direct workloads to where it achieves the best performance-to-cost ratio while maintaining "price discipline" with suppliers.

OpenAI's computational portfolio includes Microsoft, Nvidia, AWS, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy, and SoftBank. The company clarified that developing Jalapeño does not mean abandoning third-party accelerators, as it will continue to utilize partner processors for both training and inference.

The economic goal is to reduce the cost per computation. OpenAI anticipates that improved hardware performance will allow it to handle more requests with the same resources, enhancing the revenue-to-infrastructure cost ratio.

Friar also expects a counter effect, referencing Jevons' Paradox: increased efficiency may not lower overall computational consumption but rather increase it by creating new economically viable AI tasks.

Jalapeño marks the first generation of OpenAI's custom platform. The second-generation processor is already in advanced development, and the company has begun designing the third. The initial Jalapeño chips are expected to be integrated into OpenAI's infrastructure by the end of the year.

It is worth noting that in August, OpenAI temporarily slowed the scaling of new AI models and paused reinforcement learning training for two weeks for its latest systems, which are set for future deployment.