Optical PCIe extends AI rack design flexibility - ET Datacenters

Optical PCIe is not a material swap—it is a paradigm rewrite of AI interconnect architecture, shifting from copper-bound closed topologies to light-based open designs. When optical signals replace copper as the intra-rack backbone, transmission range jumps from roughly one meter to ten-plus, fundamentally liberating server topology constraints. Upstream, the structural demand shift hits EML/DFB laser and silicon photonics engine supply chains hardest. Chinese module makers will see their order mix pivot from discrete 800G/1.6T transceivers toward on-chip optical engines. The deeper disruption: NVIDIA's NVLink proprietary moat gets materially diluted, as open PCIe plus optical interconnect grants AMD, Intel, and custom ASIC players equivalent intra-rack bandwidth. Supply-chain concentration remains a vulnerability—EML laser capacity is heavily anchored at Sumitomo Electric and Mitsubishi Electric in Japan, with Lumentum lines in Taiwan, China holding critical share. Tightening export controls will accelerate domestic InP epitaxy and optical engine substitution, but full decoupling still requires a two-to-three-year window. Broadcom and Marvell are racing to embed CPO into switch silicon to capture standard-setting leverage. NVIDIA counters with Spectrum-X plus NVLink, but the gravitational pull of open ecosystems is irreversible. Within eighteen months, Optical PCIe will enter hyperscaler pilot deployments, and intra-rack interconnect cost share will climb from under five percent to fifteen-plus, making the optical engine the second value anchor in AI infrastructure after the GPU.