| | | Patrick Moorhead: Qualcomm just made a huge move into the AI infrastructure space with its new AI200 and AI250 platforms. This is a big step beyond edge, PC, IoT, auto, and mobile—positioning the company as one of the most diversified AI chip players in the market.
The AI200, arriving in 2026, is a rack-scale inference platform focused on performance-per-watt and total cost of ownership. It combines Oryon CPU compute, Hexagon NPU acceleration, and LPDDR memory in a liquid-cooled design. It scales over PCIe and Ethernet, drawing about 160 kW per rack—purpose-built for efficient inference, not brute-force training.
Then in 2027, the AI250 takes it further with "near-memory compute", delivering a reported 10x higher effective memory bandwidth and enabling disaggregated inference for large distributed AI systems.
Looking forward to more technical details as they become available.
Qualcomm also announced a 200 MW deployment partnership with HUMAIN AI, focused on building out Saudi Arabia’s AI infrastructure.
The takeaway: Qualcomm is serious about data center inference efficiency. If it executes, it could evolve from being known for mobile and edge efficiency to becoming a leader in rack-scale AI performance-per-watt—a big shift in how the market sees Qualcomm’s role in the broader AI ecosystem.
$QCOM
x.com
Interesting discussion on this link as he a panelist at the Nvida AI Washington DC conference now. |
|
