Intel, SambaNova, and Foxconn have announced a partnership to build rackscale AI infrastructure based on Intel Xeon processors. The collaboration pairs Xeon chips with SambaNova's SN-50 Reconfigurable Dataflow Units, targeting inference performance per watt and dollar rather than raw training power. The strategic bet is that as AI shifts from training to inference, the GPU-to-CPU ratio moves from 4:1 toward 1:1, restoring CPUs to a central role in data centres. Foxconn handles system integration, including a CPU-dense variant for cost-optimised inference. Intel also unveiled its new Xeon 6+ processor built on the 18A process, capable of 36,864 cores in a single liquid-cooled rack. No financial commitments or volume figures were disclosed — the announcement is a statement of intent whose thesis depends on whether the inference-era CPU-GPU ratio shift materialises.
