Qualcomm unveils AI200 and AI250 data centre chips
Qualcomm has introduced its AI200 and AI250 inference accelerators and rack systems in a bid to expand beyond smartphones into data centre infrastructure.
Qualcomm has unveiled two artificial intelligence chips aimed at data centre inference, stepping up its push beyond smartphones and sending its shares sharply higher.
The AI200 and AI250 accelerators have been introduced alongside full rack‑scale systems and are slated for commercial availability from 2026 and 2027 respectively.
The new chips are inference‑optimised accelerators that will ship as add‑in cards and as part of direct liquid‑cooled, Ethernet‑connected racks.
Qualcomm said the systems support leading AI frameworks and focus on lowering total cost of ownership for running large language and multimodal models.
Key technical details have included support for up to 768 GB of LPDDR memory per accelerator card (AI200) and a near‑memory computing design in AI250 that has been described as delivering more than 10× effective memory bandwidth, aimed at scaling up inference while curbing power use.
Qualcomm also said the accelerators are built on its Hexagon neural processing architecture adapted for data‑centre workloads, with features such as low‑precision INT formats and confidential computing support.
Saudi‑backed AI startup Humain was named as the first customer, with plans to deploy 200 megawatts of Qualcomm rack systems beginning in 2026.
Strategy beyond smartphones
The launch formed part of a wider diversification drive that has included an agreement in June to acquire U.K. connectivity specialist Alphawave for about $2.4 billion to bolster data‑centre interconnect and custom silicon capabilities.
Separately, chief executive Cristiano Amon signalled data‑centre ambitions this year, saying Qualcomm has planned a custom CPU that links to Nvidia’s platform — a step that aims to ease adoption in heterogeneous AI servers.
Alongside the silicon, Qualcomm unveiled full racks rated around 160 kW with direct liquid cooling, designed to simplify deployment at scale. The company highlighted an end‑to‑end software stack to onboard models from popular frameworks and to streamline inference operations across its cards and racks.