Qualcomm Unveils AI200 and AI250 Data Center AI Chips

Qualcomm Technologies, Inc. today unveiled two new artificial intelligence chips specifically engineered for data center deployment.

Following the announcement, the company's stock surged by more than 15% intraday, currently trading up approximately 11%.

The newly introduced chips, named AI200 and AI250, will be sold as part of accelerator cards designed to be plugged into servers. Qualcomm stated that these processors are optimized for cost-efficient AI inference workloads, delivering high performance per watt per dollar.

While Qualcomm has not disclosed specific power consumption figures or expected processing speeds for the accelerators, the company confirmed that the chips are built on its Hexagon architecture—the foundation of the neural processing units (NPUs) found in its consumer-grade system-on-chips.

Hexagon-based NPUs are already featured in Qualcomm’s flagship Snapdragon 8 Elite Gen 5 smartphone processor, which is fabricated using a 3-nanometer process. This NPU integrates 20 cores based on three distinct designs and is capable of processing up to 220 tokens per second.

Beyond mobile, Qualcomm has also embedded the Hexagon architecture into its connectivity chips. Last year, the company launched an NPU-equipped processor tailored for Wi-Fi routers. This on-device NPU enables local execution of AI models to enhance functions such as filtering malicious network traffic and optimizing power management.

The core count in the new AI200 and AI250 chips is expected to significantly exceed that of Qualcomm’s consumer-grade NPUs. For context, NVIDIA’s Blackwell Ultra GPU packs over 20,000 cores—roughly 1,000 times more than the NPU in the Snapdragon 8 Elite Gen 5.

The AI200, positioned as the less advanced of Qualcomm’s new AI chips, comes with 768GB of LPDDR memory. LPDDR is a type of RAM commonly used in mobile devices, offering lower power consumption but also reduced memory bandwidth compared to DDR5 memory typically found in servers. Memory bandwidth plays a critical role in determining how quickly data moves between processor cores and attached RAM, directly impacting AI inference performance.

Qualcomm claims the AI250 will deliver over 10 times the memory bandwidth of the AI200. This performance leap is likely achieved by replacing the slower LPDDR memory in the AI200 with a higher-performance alternative—potentially high-bandwidth memory (HBM), which is widely adopted in data center AI accelerators.

Both the AI250 and AI200 will feature confidential computing capabilities—a technology also supported by NVIDIA’s Blackwell Ultra. This feature partitions the chip’s memory into encrypted segments, ensuring that only authorized applications can access the data stored in their designated memory regions.

Qualcomm plans to ship its AI chips as part of liquid-cooled server racks. These systems will use PCIe for internal component connectivity and Ethernet for inter-rack communication.

These racks may also incorporate Qualcomm’s in-development server-grade central processing units (CPUs). In addition to dedicated AI accelerators, such systems require general-purpose CPUs to handle tasks like operating system management. NVIDIA has already begun shipping its rack-scale DGX AI systems with internally developed CPUs.

Qualcomm intends to launch the AI200 in 2026 and the AI250 in 2027, with plans to refresh its data center AI processor portfolio on an annual basis going forward.