OCP Summit Turns AI Racks Into a Memory Problem: SK hynix MARS and Cisco NVIDIA USFP

The OCP Global Summit 2026 put memory and interconnect, not accelerators, at the centre of the AI rack roadmap. SK hynix senior vice president Hoshik Kim introduced MARS, a Memory-Centric AI Rack-Scale System that disaggregates memory from CPU and GPU server constraints and coordinates compute and memory as a shared rack-scale pool. Kim argued processor-centric clusters force operators into GPU over-provisioning because idle memory on one server cannot be shared with another and adding memory means buying a whole new node. Applied to multi-turn agentic AI inference, where the KV cache explodes, MARS delivered a three- to five-fold improvement in time-to-first-token over recomputation, with tail latency dramatically reduced.

Alongside MARS, SK hynix detailed High Bandwidth Flash, which extends HBM-class bandwidth to rack-scale capacity. Kim described HBF as typically half the bandwidth of HBM but roughly 14 times the capacity, framing it as an answer to the industry's capacity constraints rather than a bandwidth play. The OCP HBF specification version 0.7 was released in July, version 1.0 is expected early next year, and HBF Cube samples are planned for early 2028. The pitch lands squarely on the current memory crunch, where capacity rather than compute is emerging as the binding constraint on large-context inference.

Networking moved in step. Ahead of the summit, Cisco briefed analysts on USFP, an ultra-dense small form factor pluggable module developed with NVIDIA for 400G SerDes lanes rather than today's common 200G, enabling higher-density 1.6 Tbps fabrics - 1.6 Tbps across four lanes, scaling to 6.4 T over 16 lanes and 12.8 T over 32. The two are drafting a multi-source specification so other vendors can build to it. Cisco also detailed Silicon One G300 in first systems at 102.4 T of capacity across 1.6 T ports, a 51.2 T G302 top-of-rack switch due in the fourth quarter of 2026, and the A100L for low-cost out-of-band management. Rack-level integration is following: Foxconn used the summit to unveil Vera Rubin AI racks and liquid-cooling systems, while Chenbro showed AI rack-level solutions.