A DIGITIMES analyst projects that optical communications will migrate inside AI server racks around 2028, as ballooning cluster scale demands faster, lower-latency, higher-bandwidth links. The framing matters more than the date: it signals that copper is approaching a physical ceiling exactly as model training and inference push interconnect requirements past what electrical signaling can economically sustain.
The strategic story here is a shift in where value accrues. As GPUs saturate available bandwidth, the bottleneck moves from compute to the fabric that stitches accelerators together. Co-packaged optics and optical I/O turn what was a passive commodity, cabling and connectors, into a performance-defining, margin-rich layer. Expect the winners to be firms controlling lasers, optical engines, and the packaging that fuses photonics to silicon, not just the switch and GPU vendors that dominate today's headlines. A 2028 horizon also implies design-in decisions are being made now, which compresses the window for suppliers to qualify parts and lock into hyperscaler roadmaps.
There is a second-order effect worth flagging: optics is partly a power story. Every watt saved moving data is a watt freed for compute, and with community and regulatory resistance to power-hungry AI datacenters intensifying globally, interconnect efficiency becomes a lever for keeping buildouts politically and economically viable, not merely a speeds-and-feeds upgrade.
For Japan, this is a rare case where the country's industrial strengths align with an emerging bottleneck. Japan holds deep positions in optical components, laser diodes, precision connectors, and photonics materials, the unglamorous inputs that a rack-level optical transition consumes at volume. If domestic suppliers can qualify into hyperscaler and merchant-switch supply chains before 2028, they capture durable, high-spec demand rather than commodity volume.
For Japanese SIers and enterprise IT teams, the near-term implication is planning, not procurement. On-premises AI clusters and colocation deployments built in the next 18 to 24 months should be evaluated for a forward migration path, since rack architectures assuming purely electrical interconnects may face earlier-than-expected refresh pressure. The teams that treat interconnect roadmaps as a capacity-planning variable, rather than a vendor detail, will avoid stranded infrastructure as the optical transition arrives.
