TSMC, Intel, Samsung, GlobalFoundries diverge on AI optical chip links

As AI accelerators outgrow copper interconnects, the four major foundries are pursuing distinct co-packaged optics (CPO) strategies to move optical connectivity closer to compute. Their approaches differ on integration method and timeline.

Why it matters: Interconnect bandwidth, not just raw compute, is becoming the bottleneck for scaling AI training and inference clusters, so whoever solves optical I/O most efficiently gains a real system-level edge. This foundry divergence signals co-packaged optics moving from lab research toward a near-term manufacturing race, tied to the same power and scaling pressures behind GlobalFoundries' recent $300M silicon-photonics grant.