OpenAI CFO Sarah Friar has detailed the company strategy for scaling intelligence through an integrated system of data centers, chips, frontier models, and products. According to OpenAI, progress in artificial intelligence compounds fastest when hardware, software, and models improve together. A central announcement is the release of measured performance results for Jalapeño, OpenAI's first custom inference chip. Tested on InferenceX using GPT-OSS 120B, Jalapeño delivered higher peak throughput per kilowatt and lower token latency than competing commercial systems while also performing strongly on DeepSeek R1 and Kimi K2. The custom silicon provides OpenAI with a first-party path alongside external accelerators, improving latency, throughput, energy efficiency, and serving economics. To meet varying workload demands, OpenAI manages a diverse compute portfolio involving Microsoft and NVIDIA as foundational partners alongside AWS, AMD, Broadcom, Cerebras, CoreWeave, Oracle, SB Energy, and SoftBank. This multi-provider approach allows the company to optimize for capability and cost while maintaining pricing discipline. Furthermore, data center initiatives like Project Camellia in Georgia showcase facility design centered on workload efficiency, local infrastructure support, and closed-loop water conservation. OpenAI argues that these system-wide efficiencies result in more useful intelligence per unit of compute, citing GPT-5.6 Sol reaching a high on the Artificial Analysis Coding Agent Index while using 54 percent fewer output tokens than another leading model. Friar notes that as intelligence becomes more capable and affordable, it triggers Jevons paradox, where greater efficiency expands consumption and drives new economic activity. Ultimately, OpenAI aims to leverage lower costs and higher productivity to fund continued research, infrastructure, and safety advancements.
The full stack behind abundant intelligence
OpenAI CFO Sarah Friar outlines how the company is scaling its AI capabilities through an integrated approach spanning custom silicon, software, and data centers. The company shared initial performance results for its first custom inference chip, Jalapeño. OpenAI also maintains a broad compute portfolio partnering with companies like Microsoft and NVIDIA.
Why it matters
Developing first-party silicon alongside partnerships gives AI companies greater control over hardware economics, performance, and workload optimization. This integration drives down the cost of intelligence, fueling broader enterprise adoption and sustained industry growth.