On October 27, Qualcomm announced its next-generation data center AI inference chips — Qualcomm AI200 and AI250, along with new accelerator cards and rack-scale systems built around them. These new solutions showcase Qualcomm's deep expertise in NPU technology, offering rack-level performance, exceptional memory capacity, and energy efficiency — all designed to accelerate generative AI workloads at industry-leading cost efficiency.

Following the announcement, Qualcomm's stock surged nearly 22% intraday, closing up 11.09% on Wall Street.

The AI200 is a rack-scale inference solution engineered for large language and multimodal models (LLM, LMM). With 768GB of LPDDR memory per card, it delivers high scalability, flexibility, and low total cost of ownership (TCO) for diverse AI workloads.

The AI250, on the other hand, introduces a near-memory computing architecture, offering over 10x higher effective memory bandwidth and significantly lower power consumption. This breakthrough boosts AI inference efficiency and enables more cost-effective hardware utilization — a leap forward for decomposed inference and scalable AI deployment.

Both systems use direct liquid cooling for thermal efficiency, PCIe for vertical scaling, and Ethernet for horizontal expansion. They support confidential computing for secure AI operations and deliver 160kW rack-level power.

Qualcomm's AI software stack spans from application to system layers, fully optimized for inference. It supports major machine learning frameworks and includes tools such as the Efficient Transformers Library and AI Inference Suite, allowing seamless integration and one-click deployment of Hugging Face models. The ecosystem also includes ready-to-use AI apps, agents, APIs, and developer tools to speed up AI deployment.

While Qualcomm has not yet published the detailed technical documentation for AI200 and AI250, the company confirmed commercial availability in 2026 and 2027, respectively. Qualcomm also plans to update its data center roadmap annually, focusing on performance, energy efficiency, and total cost leadership.

Earlier this year, Qualcomm signed an MoU with Saudi Arabia's AI company HUMAIN to co-develop next-generation AI data centers and edge-to-cloud infrastructure using Qualcomm's in-house CPUs and AI solutions. CEO Cristiano Amon later confirmed the company's ambition to build general-purpose data center CPUs and AI inference racks targeting hyperscale clients.

“If discussions with a leading hyperscaler go well,” Amon said, “we expect to start generating data center revenue in fiscal 2028.”