The wires that carry data between AI chips are running out of road, and the world's biggest chipmakers are betting on light to replace them.
According to Tom's Hardware, TSMC, Intel, Samsung Foundry, and GlobalFoundries are each pursuing a distinct strategy in co-packaged optics, or CPO — the practice of moving optical connections physically closer to the processor instead of leaving them at the edge of the system. The publication's roadmap breakdown frames this as a response to a specific problem: as AI systems grow, they are outgrowing the copper interconnects that traditionally shuttle data between chips.
That matters because modern AI training doesn't happen on one chip. It happens across racks of accelerators that must behave like a single machine, constantly exchanging enormous volumes of data. This is what the industry calls "scale-up" connectivity, and Tom's Hardware identifies it as the target these four foundry roadmaps are aiming at.
Copper is cheap and well understood, but it loses signal quality as speeds climb and distances stretch, and it burns power doing it. Optical links — data as pulses of light rather than electrical current — hold up better over distance. The engineering challenge is integration: fusing photonics and standard silicon manufacturing into one package is difficult, and Tom's Hardware reports the four companies are approaching that problem in four notably different ways.
The source material does not specify technical details of each roadmap beyond the fact that the approaches diverge. That divergence is itself the story — it suggests the industry has not yet converged on a single standard for how optics should meet compute.
Why it matters: whichever approach wins will shape the cost, power draw, and physical scale of the AI data centers being built over the next decade, and could reshuffle competitive standing among the foundries that manufacture nearly all of the world's advanced chips.