What is the OpenAI Jalapeño chip?

Jalapeño is a custom-designed artificial intelligence chip developed by OpenAI in partnership with Broadcom. It was specifically engineered for large language model (LLM) inference to improve power efficiency and speed compared to current industry standards.

How much better is Jalapeño's energy efficiency compared to other chips?

According to initial benchmarks, Jalapeño delivers 1.5–1.9 times more AI work per watt than existing solutions. This significant improvement in energy efficiency addresses rising power costs and compute demands in the AI industry.

What are the latency and speed improvements of Jalapeño?

The chip reduces end-to-end latency by 1.7–3.6 times, resulting in faster and more responsive user interactions. Additionally, performance on highly interactive workloads increased by 2.1–4.1 times compared to competitors.

Why is OpenAI creating its own chip instead of using NVIDIA?

OpenAI aims to reduce its heavy reliance on NVIDIA GPUs, which have faced supply constraints and escalating costs. By developing custom silicon with Broadcom, OpenAI seeks to secure its supply chain and optimize its full-stack infrastructure.

How long did it take to develop the Jalapeño chip?

Broadcom designed the Jalapeño chip from scratch over a nine-month period. The development process utilized OpenAI’s own models to accelerate design and ensure the hardware was optimized for specific workload patterns.

OpenAI Jalapeño Chip Metrics: Efficiency and Latency Data

TL;DR: OpenAI’s custom Jalapeño chip delivers up to 1.9x better energy efficiency than current solutions. This breakthrough reduces latency significantly, potentially reshaping the economics of large-scale AI inference.

Key facts

OpenAI released initial performance data for its custom LLM inference chip, Jalapeño, developed in partnership with Broadcom.
Jalapeño was announced on June 24, 2026, and first test results were shared on August 25, 2026.
The chip delivers 1.5–1.9x more AI work per watt compared to current industry standards.
End-to-end latency was reduced by 1.7–3.6x, with interactive workload performance increasing by 2.1–4.1x.
Jalapeño was designed from scratch over a nine-month timeline using OpenAI’s models to accelerate the design process.
Engineering samples successfully ran GPT-5.3-Codex-Spark at target production frequency and power levels.
OpenAI partners with Celestica for board, rack system integration, and scalable production systems.

OpenAI Unveils Concrete Metrics for Custom Jalapeño Inference Chip

OpenAI has released its first concrete performance data for Jalapeño, a custom-designed artificial intelligence chip developed in partnership with Broadcom. Announced on June 24, 2026, the processor represents the company’s most significant step toward controlling its own hardware infrastructure [1]. On August 25, 2026, OpenAI shared initial test results showing that Jalapeño delivers substantial gains in power efficiency and speed compared to current industry standards [3].

The new chip is designed specifically for large language model (LLM) inference—the process of running trained AI models to generate responses or analyze data. Early benchmarks indicate it offers significantly better performance per watt than existing state-of-the-art chips, a critical metric as the AI industry grapples with rising energy costs and compute demand [1].

Performance Metrics and Testing Results

The August 25 announcement provided specific numbers from initial testing on InferenceX, a benchmarking platform. Across three public models tested, Jalapeño demonstrated notable improvements in key operational areas compared to existing solutions [3]:

Energy Efficiency: The chip delivered 1.5–1.9× more AI work per watt. This metric measures how much computational output is generated for each unit of electricity consumed.
Latency Reduction: End-to-end latency was reduced by 1.7–3.6×. Latency refers to the delay between a user submitting a request and receiving a response; lower latency results in faster, more responsive interactions.
Interactive Workload Performance: On highly interactive tasks, performance increased by 2.1–4.1× compared to competitors [3]

These figures support earlier claims made by OpenAI that the chip would offer industry-leading efficiency [1]. The data suggests that Jalapeño is not just a incremental upgrade but a fundamental shift in how OpenAI intends to handle high-volume AI applications.

Development Timeline and Engineering Samples

The rapid development of Jalapeño highlights an accelerated engineering timeline. According to Broadcom, the chip was designed from scratch over nine months [2]. The project utilized OpenAI’s own models to help accelerate the design process, ensuring the hardware was optimized for specific workload patterns [2].

Engineering samples are already operational. These early units have successfully run machine learning workloads at their target production frequency and power levels [2]. Notably, tests included running GPT-5.3-Codex-Spark, a specialized coding model, confirming that the chip can handle complex, real-world AI tasks without stability issues [2].

Strategic Shift Away from NVIDIA Dependence

The introduction of Jalapeño reflects a broader strategic pivot within the tech industry. For years, major AI companies have relied heavily on graphics processing units (GPUs) manufactured by NVIDIA, which currently dominate the AI acceleration market [1]. However, intense demand for these chips has led to supply constraints and escalating costs [1].

By developing its own silicon, OpenAI aims to reduce this reliance. The company is partnering with Broadcom for chip implementation and Celestica for board, rack system integration, and scalable production systems [2]. This collaboration allows OpenAI to optimize its full-stack infrastructure, aligning models, software kernels, serving systems, networking, and hardware around shared operational goals [1][2].

This move marks the first generation of a multi-generation compute platform being built by OpenAI and Broadcom [2]. The chip is planned for deployment at a gigawatt scale in partnership with data center providers, signaling long-term industrialization of AI infrastructure [2]. This approach ensures reliability and accessibility as model demands continue to grow [2].

Leadership Handover and Future Outlook

The significance of the milestone was underscored by a formal handover ceremony. OpenAI CEO Sam Altman and President Greg Brockman received the first Jalapeño chips from Broadcom executives Hock Tan and Charlie Kawwas [2]. This event symbolizes the transition from concept to tangible hardware readiness.

While NVIDIA remains a dominant force, the emergence of custom silicon like Jalapeño indicates that leading AI labs are seeking alternatives to manage costs and secure supply chains. OpenAI’s strategy focuses on industrializing its infrastructure to ensure it can meet future demands efficiently [2].

The initial results from Jalapeño provide early evidence that custom-designed hardware can outperform general-purpose accelerators in specific, high-volume tasks. As the company moves toward gigawatt-scale deployments, these efficiency gains could reshape the economics of AI inference, potentially lowering costs for users and reducing the environmental impact of large-scale computing [1].

OpenAI’s ability to deliver working engineering samples within nine months and achieve significant performance jumps in early testing demonstrates the feasibility of rapid hardware iteration. The upcoming multi-generation platform will likely build on these foundations, further integrating software and hardware optimizations to sustain OpenAI’s competitive edge in the AI market.

Sources

OpenAI and Broadcom Unveil Jalapeño AI Chip for LLM Inference (opendatascience.com) — 2026-06-24
🚨 AI News | TestingCatalog (@testingcatalog) on X (x.com) — 2026-08-25
OpenAI and Broadcom Unveil LLM-Optimized Intelligence Processor | Broadcom Inc. (investors.broadcom.com) — 2026-01-01