Qualcomm at its Investor Day on June 24 unveiled the Dragonfly data center chip family, entering the AI infrastructure market with custom silicon built for agentic workloads — and secured Meta as its first major customer.
What Qualcomm announced
The Dragonfly portfolio includes three products:
Dragonfly C1000 CPU: A 250-plus-core server processor built on Qualcomm’s custom Oryon architecture, running above 5 GHz. The company claims more than 2x better performance per watt versus competing server CPUs.
Dragonfly AI300 inference accelerator: A third-generation AI card rated at 4x to 8x better performance per watt than current GPU-based systems on memory-bandwidth metrics. Commercial sampling expected in 2028.
High Bandwidth Compute (HBC): 3D-stacked near-memory computing that delivers 133 terabytes per second per card with the AI250 accelerator — an 18x bandwidth improvement over the prior generation — due for mid-2027.
Meta signs on
Meta committed to a multi-year, multi-generation agreement to deploy the Dragonfly C1000 in its next-generation server fleet, with production planned to begin in the second half of 2028. More than 35 hardware and cloud partners joined the ecosystem, including Supermicro, Lenovo, GIGABYTE, and SK hynix.
Agentic AI drives the strategy
“Agentic AI is driving significant increase in demand for AI inference,” said CEO Cristiano Amon. Qualcomm’s pitch is that purpose-built inference hardware can serve large-scale AI deployments more efficiently than repurposed training accelerators.
The announcement comes alongside the simultaneous close of Qualcomm’s $3.9 billion acquisition of AI software startup Modular, signaling an intent to offer a complete hardware-and-software stack for AI data centers — a direct challenge to Nvidia’s dominant position.