Qualcomm unveils AI200 and AI250 data-centre inference chips

Qualcomm announced two rack-scale accelerator cards for AI inference, the AI200 and AI250, marking its entry into the data-centre chip market long dominated by Nvidia. The AI200 is built around the Hexagon neural processing architecture that also powers Qualcomm’s Snapdragon phone chips, carries 768GB of LPDDR memory per card, and is due in 2026; the AI250 follows in 2027 with a near-memory computing design Qualcomm said would deliver more than ten times the effective memory bandwidth of the AI200 at lower power. Both are water-cooled, rack-scale, PCIe- and Ethernet-connected systems with confidential-computing features, and Qualcomm said it plans an annual refresh cadence going forward.

Qualcomm pitched the chips on performance-per-dollar-per-watt for inference workloads rather than on raw throughput, positioning them as a lower-cost alternative to Nvidia’s GPUs for running rather than training models. The company’s shares rose sharply on the announcement, closing up roughly 11% after an intraday gain of more than 15%.

The launch came with a customer already attached: Saudi Arabia’s Humain, the state-backed AI company, said it would deploy 200 megawatts of AI200 and AI250 capacity starting in 2026 to deliver inference services regionally and globally, with Qualcomm opening an AI engineering centre in Riyadh to support the rollout.

Qualcomm’s move added a further challenger — alongside AMD, Amazon’s Trainium and Google’s TPUs — to a market where Nvidia has for several years captured the large majority of AI accelerator revenue, though Qualcomm’s chips target inference rather than the training workloads where Nvidia’s lead has been most pronounced.