Qualcomm on Monday announced its entry into the data center AI chip market with the AI200 and AI250, two inference accelerators based on its Hexagon neural processing units. The AI200 will ship in 2026 and the AI250 in 2027, both as rack-scale systems with liquid cooling. The AI200 supports 768GB of LPDDR memory onboard and a rack power envelope of 160 kilowatts, comparable to some high-power Nvidia GPU racks.

The AI250, arriving a year later, adds a near-memory computing architecture that Qualcomm says boosts effective memory bandwidth by more than 10 times. The AI200 supports confidential computing, and the AI250 preserves the same security characteristics. They are designed to run large transformer models for inference, not training.

Saudi Arabia’s Humain, which signed a partnership with Qualcomm in May, is the first announced customer, committing to deploy up to 200 megawatts of Qualcomm’s AI systems as the systems become available in 2026. Qualcomm said it will also sell components separately, allowing hyperscalers to mix and match, and even suggested that Nvidia or AMD could become customers for its data center CPUs.

Qualcomm’s stock rose 11% on the news, as investors bet that the company can carve out a slice of the fast-growing AI infrastructure market currently dominated by Nvidia and AMD. The announcement marks a strategic shift for Qualcomm, which has primarily focused on wireless connectivity and mobile devices.