오픈에이아이(OpenAI)는 구글(Google), 마이크로소프트(Microsoft), 아마존(Amazon)과 함께 맞춤형 할라페뇨(Jalapeno) 칩을 도입하며, 학습 워크로드에서 표준 엔비디아(NVIDIA) 가속기 의존도를 한층 낮췄다.

OpenAI recently introduced its first custom artificial intelligence processor, dubbed Jalapeño, marking a significant step toward in-house silicon development and intensifying pressure on Nvidia’s near-monopoly over advanced AI computing hardware. The custom-built inference semiconductor is designed to deliver industry-leading speed and operational efficiency, positioning OpenAI alongside tech giants such as Google, Amazon Web Services (AWS), Microsoft, and Meta in engineering proprietary chips to power large-scale AI systems. Developed in collaboration with semiconductor manufacturer Broadcom, the Jalapeño processor is specifically optimized for inference—the phase in which trained models execute tasks and process user queries.

OpenAI plans to deploy Jalapeño across its computing infrastructure before the end of the year. The company confirmed that engineering teams are already developing second- and third-generation versions of the chip.

Nvidia’s market valuation has surged throughout the global data center expansion, driven by mounting demand for its GPUs in both model training and daily inference. Nevertheless, market analysts warn that the rapid proliferation of hyperscaler-designed silicon presents a growing threat to Nvidia’s long-term dominance, particularly in the inference segment. Evaluating the chip’s performance inside OpenAI’s development facilities, independent research firm SemiAnalysis found that Jalapeño surpassed Nvidia’s Blackwell architecture in performance per watt across nearly every testing benchmark. Analysts caution, however, that the direct comparison is somewhat asymmetric because Jalapeño utilizes next-generation HBM4 memory, making Nvidia’s upcoming Rubin platform a more accurate peer. TrendForce analyst Fion Chiu observed that while Jalapeño will decrease OpenAI’s reliance on Nvidia for daily inference tasks, Nvidia’s GPUs will remain indispensable for large-scale model training and frontier AI workloads due to their flexible programmability and established CUDA software ecosystem.

OpenAI’s silicon debut reflects an accelerating industry-wide transition toward custom application-specific integrated circuits (ASICs). Google pioneered this shift with its Tensor Processing Unit (TPU) expansion and continues to deploy next-generation TPUs across its cloud network for both model training and live inference. Meta followed suit after formalizing agreements to deploy one gigawatt of Broadcom-engineered custom AI processors as part of a multi-GW infrastructure plan. Anthropic has committed more than $100 billion over the next decade toward AWS infrastructure, which includes Amazon’s proprietary Trainium processors. Meanwhile, emerging chipmakers such as Cerebras, SambaNova, D-Matrix, Etched, and Fractile are similarly advancing specialized AI accelerators.