Marvell is demonstrating 2nm-based optical interconnect technologies for AI data centers at ECOC 2026, spanning 400G-per-lane PAM4, 800G and 1.6T pluggables, and a 102.4T co-packaged optics platform. The headline is the process node, but the real story is where value in AI infrastructure is migrating.
For two years the industry narrative centered on GPUs and HBM. That framing is now incomplete. As clusters grow into tens of thousands of accelerators, throughput is increasingly gated not by raw compute but by how quickly data crosses the fabric between chips, racks, and buildings. Copper is running out of reach at these speeds and power budgets, which pushes optics from the network edge toward the package itself. Co-packaged optics matters because it attacks the energy-per-bit problem that determines how large a coherent training cluster can physically get before power and heat cap it. Whoever controls the interconnect layer captures margin that the market has so far assigned mostly to accelerator vendors.
Strategically, this deepens the moat around a small set of connectivity suppliers and reprices the AI supply chain. Hyperscalers designing custom silicon still need someone to solve serialization, coherent optics, and MACsec-grade security at scale. Expect interconnect IP and DSPs to become a contested procurement category, and expect optics roadmaps to dictate the cadence of the next data center generation as much as GPU roadmaps do.
For Japan, this is squarely in a zone of historical strength. Japanese firms hold deep positions in optical components, lasers, connectors, precision materials, and photonics manufacturing that feed exactly this transition. The opportunity is real but narrow: leadership in discrete components does not automatically translate into control of the integrated modules and DSP-driven systems where value is concentrating. The question is whether Japanese suppliers move up the stack toward co-packaged assemblies rather than remaining upstream part vendors.
For domestic SIers and enterprise IT teams, the implication is planning, not product. As sovereign and enterprise AI data center projects advance in Japan, interconnect architecture, power density, and thermal design become first-order decisions, not afterthoughts. SIers that can specify optics-aware fabric designs, and RPA and automation teams that account for the latency and bandwidth realities of large clusters, will be far better positioned than those still scoping AI infrastructure as a GPU-count exercise.
