Jefferies’ new forecasts illustrate the potential scale of growth in optical connectivity within AI infrastructure.

Jefferies projects that the AI ​​transceiver market will approach 180 million units by 2027.

Of this total, approximately 90 million units are expected to be 800G, while over 80 million will be 1.6T.

By 2028, the total number of transceivers is expected to reach 245 million units.

The key takeaway here is this:

AI data centers require more than just additional GPUs; these GPUs must constantly communicate with one another and with other systems. As the number of GPUs and the volume of data traffic rise, the demand for high-speed optical connectivity grows exponentially.

Consequently, the expansion of AI infrastructure generates significant demand not only for GPU manufacturers but also for companies within the optics and networking ecosystem.

This is one of the reasons why I am closely monitoring $LITE, $COHR, and $AAOI.

There is also the matter of CPO (Co-Packaged Optics).

In simplest terms, CPO involves positioning optical connections much closer to the switch or processor chip—or even within the same package—rather than having them located outside the system as separate transceivers.

The goal is to achieve higher speeds while reducing power consumption and signal loss.

Jefferies highlights a crucial point: the ramp-up of CPO to significant production volumes may be delayed until 2029.

This is significant because CPO is not expected to immediately displace the existing 800G and 1.6T transceiver market. There is still ample room for these technologies to achieve substantial volumes over the next few years. In summary:

2027 → ~180M transceivers
2028 → ~245M transceivers
2029 → CPO is expected to achieve significant scale

As the number of GPUs increases, the infrastructure connecting them must also grow.

Compute grows → data traffic grows → networking grows → the need for optical connectivity grows.