Multi-Core Fiber - A solution for Scale-up, Out and Across

Author: Taylor Bremekamp, Fiber Network Design Engineer, Sumitomo Electric Lightwave Collection: 2026-OCP-EMEA-Summit-Executive-Sessions

Summary

Taylor Bremekamp of Sumitomo Electric Lightwave presents multi-core fiber (MCF) as the next-generation fiber infrastructure solution for AI data center networking. The presentation addresses the exponential fiber density crisis driven by GPU generational progression: from Hopper (32 GPU dies/rack, ~64 FE fibers/rack) to Blackwell NVL72 (72 GPU dies/rack, ~288 FE fibers/rack) to Rubin NVL144 (~576 fibers/rack) to Feynman NVL288 (~1,152 fibers/rack on scale-out, plus ~4,608 fibers for optical scale-up). Each generation roughly doubles fiber count, and Nvidia has officially announced that optical scale-up will be required in upcoming generations, adding a massive new fiber density demand beyond existing scale-out connectivity.

The existing solutions — parallel optics (adding more fibers/transceivers, approaching practical limits at 50,000–100,000 fibers per row) and wave division multiplexing (WDM, power and cost-intensive, currently limited to scale-across links >500m) — are both hitting their limits for AI backend applications. MCF, classified as spatial division multiplexing (SDM), places 4 cores inside a single cladding with the same 125 µm outer diameter as a standard single-core fiber. This achieves 4x fiber density while maintaining similar insertion loss and attenuation characteristics. The result: 60–80% space savings in conduit (single-core OD ~37mm, area 1,075 mm²; 4-core MCF OD ~13mm, area 133 mm² for equivalent capacity), 40–50% weight reduction, and proportional reductions in connector count, trunk count, transceiver count, and installation labor.

An MSA (Multi-Source Agreement) among major MCF manufacturers was reached in April 2026, standardizing 4-core MCF design requirements, with formal standard publication expected June–July 2026. The near-term deployment model is hybrid: MCF trunks between consolidation points with FIFO (fan-in/fan-out) devices converting to single-core at transceivers. The target-state is end-to-end MCF with MCF-native transceivers. Future directions include MCF direct-to-chip integration with CPO (co-packaged optics) solutions.