Deep|Optics: Addressing Optical Transceivers Oversupply Concern

Recently, we have been receiving an increasing number of investor questions on the potential overbuild in transceivers, which has partially weighed on the shares of US-listed optics names. In short, the argument is that there are around 90-100m 400G+ transceiver shipments this year, versus only around 14-15m XPUs, i.e., a ~6:1 attach rate, which is much higher than the attach rate implied by GPU and TPU network architectures. We believe the market has overlooked the details below:

AI compute clusters require a very high optical transceiver attach rate, with the industry average at roughly 1:3.5

Taking the NVIDIA GPU cluster architecture as an example, the ratio of GPUs to 400G-and-above optical transceivers typically falls within the 1:3 to 1:4 range. Breaking this down at the hardware-link level: each GPU requires one set of optical transceivers to connect upstream to the first-layer switch, while the switch also requires another set on the downlink side to connect back to the GPU. This already consumes two optical transceivers per GPU. In today's mainstream deployment architecture, which generally uses a two-tier switching fabric, the interconnect between the first-layer and second-layer switches consumes another two sets of optical transceivers, bringing the overall ratio close to 1:4.

Some deployments apply bandwidth convergence at the second-layer switch, while some large-scale compute campuses adopt a three-tier switching architecture; others may simplify the design to a single-tier switching structure. Taken together, these different deployment models suggest an industry-average assumption of approximately 1:3.5.