Qualcomm introduced AI200 and AI250, new data center solutions optimized for generative AI
By ForGeeks
Qualcomm Technologies, Inc. announced a new generation of AI-optimized data center solutions: gas pedal cards and racks based on Qualcomm AI200 and AI250 chips. These products, based on the company’s advanced NPU technology, deliver scalable performance, high energy efficiency, and an optimal power-to-cost ratio, paving the way for more efficient deployment of generative artificial intelligence across a variety of industries.
The Qualcomm AI200 solution is a purpose-built AI inference rack focused on low total cost of ownership (TCO) and high performance for large language and multimodal models (LLM, LMM) as well as other AI tasks. Each card supports up to 768 GB of LPDDR memory for greater capacity and flexibility when working with large-scale models.
Qualcomm AI250 utilizes a new memory architecture with near-memory computing that delivers ten times the effective bandwidth and significantly reduced power consumption compared to previous generations. This solution enables efficient allocation of compute resources, balancing performance and cost for customers.
The two racks feature direct liquid cooling for efficient heat dissipation, support PCIe for scaling «up» and Ethernet for scaling «wide». In addition, support for confidential computing is provided to ensure the security of AI workloads. The power consumption of a single rack is 160 kW.
«With the Qualcomm AI200 and AI250 solutions, we are redefining AI-inference capabilities at the rack level. These innovative infrastructure products enable companies to deploy generative AI with unprecedented efficiency at minimal cost, while maintaining the flexibility and security that today’s data centers require», — said Durga Malladi, senior vice president and general manager, Technology Planning, Edge Solutions & Data Center at Qualcomm Technologies.
According to Durga Malladi, senior vice president and general manager, Technology Planning, Edge Solutions & Data Center at Qualcomm Technologies.
The new solutions are backed by an enterprise-grade software stack optimized for AI inference — from the application to the system level. It supports all leading machine learning frameworks, inference engines and LLM/LMM optimization, including disaggregated serving techniques. Developers will be able to easily integrate and run models from the Hugging Face library at the click of a button through the Qualcomm Efficient Transformers Library and Qualcomm AI Inference Suite.
The software suite also includes out-of-the-box AI applications and agents, libraries, APIs and tools for deploying and scaling artificial intelligence in work environments.
The Qualcomm AI200 is expected to launch commercially in 2026 and the AI250 — in 2027. The company intends to adhere to an annual cycle of updates to its data center solutions, focusing on improving AI inference performance, energy efficiency and lowering TCO.
