• Qualcomm AI200 and AI250 deliver rack-scale performance and high memory capacity with industry-leading total cost of ownership (TCO), enabling faster generative AI inference in data centers
• Qualcomm AI250 introduces a new near-memory computing architecture that significantly boosts memory bandwidth and efficiency for AI workloads
• Both solutions are compatible with major AI frameworks and include a full software stack for secure, scalable deployment across data centers
• AI200 and AI250 are part of Qualcomm’s annually refreshed, multi-generation AI inference roadmap
October 28, 2025 — Qualcomm Technologies, Inc. announced new datacenter-focused AI inference solutions built on the Qualcomm® AI200 and Qualcomm® AI250 chips, including accelerator cards and rack-scale systems. Based on the company’s NPU technology leadership, the new solutions are designed to deliver superior rack-scale performance and memory capacity while maximizing performance per watt and per dollar, positioning them as a key step toward flexible and scalable generative AI infrastructure across industries.
The Qualcomm AI200 is a rack-level solution optimized for a wide range of generative AI inference workloads including large language models (LLMs) and multimodal models (LMMs). It supports up to 768 GB of LPDDR memory per card, offering high memory capacity and cost-efficiency to improve scalability and flexibility in large-scale deployments.
Meanwhile, the Qualcomm AI250 adopts an innovative near-memory computing architecture, enabling more than a 10-fold increase in memory bandwidth and significant reductions in power consumption. This generational improvement is designed to enhance the efficiency and performance of AI inference workloads. The AI250 also supports a disaggregated compute architecture to meet customer performance and TCO objectives.
Both AI200 and AI250 solutions support direct liquid cooling for thermal efficiency, as well as PCIe and Ethernet connectivity for scale-out architectures. Confidential computing enables secure handling of sensitive AI workloads, and system-level power consumption at the rack level is approximately 160 kW.
Durga Malladi, Senior Vice President and General Manager of Technology Planning, Edge Solutions, and Data Centers at Qualcomm Technologies, said,
“With the Qualcomm AI200 and AI250, we are redefining what rack-scale AI inference can achieve. These solutions allow enterprises to deploy generative AI with improved TCO while delivering the security and flexibility data centers require.”
He added,
“With robust software support and an open ecosystem, developers and organizations can easily integrate, manage and scale pretrained models on our inference platforms. Full compatibility with major frameworks and one-click deployment provide an accelerated path to innovation.”
Qualcomm’s hyperscaler-grade AI software stack covers the entire infrastructure from applications to system software and is optimized for AI inference. It supports leading machine learning frameworks, inference engines, generative AI frameworks and optimizations for LLM/LMM workloads including disaggregated architectures. Developers can deploy Hugging Face models with one click using Qualcomm’s Efficient Transformers Library and Qualcomm AI Inference Suite. The software offering also includes ready-to-use AI applications and agents, as well as tools, libraries, APIs and services designed to bring AI into production environments.
