Qualcomm unveiled a full‑stack data‑center roadmap at its Investor Day, introducing the Dragonfly C1000 CPU, High Bandwidth Compute modules, and the AI300 inference accelerator. The lineup targets agentic AI workloads, promising higher token‑per‑watt efficiency and lower total‑cost‑of‑ownership. CEO Cristiano Amon said multi‑year deals with top customers position Qualcomm to challenge traditional server vendors.

Dragonfly’s C1000 chiplet combines custom Oryon cores clocking above 5 GHz with a 250‑plus core count, delivering more than twice the performance‑per‑watt of current server CPUs. It supports PCIe Gen 7, CXL and up to 2 TB/s bandwidth, while offering both air and liquid cooling for OCP‑compliant racks. Qualcomm expects first silicon shipments in 2028.

The AI300 accelerator builds on the earlier AI200/AI250 cards, integrating HBC Gen 2 to achieve 4‑8× better performance‑per‑watt than GPU‑based rivals and up to 133 TB/s effective memory bandwidth. HBC Gen 1 AI250 sampling begins mid‑2027, giving customers early access before the AI300 rollout in 2028. Qualcomm frames the portfolio as an alternative for hyperscalers seeking disaggregated inference platforms.

Qualcomm also announced bespoke custom silicon for agentic AI, leveraging its 40‑billion‑component design experience to deliver modular packaging and rapid time‑to‑market. The company targets commercial sampling of these solutions by 2028.