Scale-up networking drives AI ROI

Connections among GPUs, CPUs, storage, and switches within racks must be extremely fast and error-free. Without this, GPUs can't work as a single unit, causing costly AI equipment to fail.

NVIDIA controls the rack today

NVIDIA dominates in-rack networking via NVLink and NVSwitch, technologies that create a highly scalable mesh of GPU connectivity that improves with each new generation.

Hyperscalers build their own alternatives

AWS, Microsoft, Google, and Meta have created custom interconnect technologies to work with their proprietary chips, challenging NVIDIA's dominance.

Two industry standards are emerging

Ultra Accelerator Link and Ethernet for Scale-Up Networking are gaining momentum as alternatives to NVIDIA's proprietary technologies.

ESUN leads the standards race

NVIDIA's NVLink controls over 90% of in-rack networking, but ESUN, promoted heavily by Broadcom, has emerged as the near-term specification for alternatives.

“NVIDIA's NVLink and NVSwitch connectivity control over 90% of in-rack networking”

Optical networking is coming soon

All leading hyperscalers and infrastructure vendors plan a gradual but definite shift to optical networking within the next few years.

Co-packaged optics will be standard

Nearly all companies involved in scale-up networking, from component suppliers to switch makers, plan to eventually standardize on CPO technology.