AI training is not one GPU working; it is thousands, even tens of thousands, of GPUs working together. These GPUs have to keep exchanging gradients, parameters, and activations, and a single training iteration may involve moving terabytes of data over the network. No matter how powerful the GPUs are, if the network can’t keep up, the whole cluster just sits there waiting.
So in an AI cluster, beyond computing power, there is another layer that is just as critical but overlooked by ordinary investors — the interconnect layer. The core of this layer is two things: switching chips and optical modules.
First, let’s talk about the layering of interconnects.
The network in an AI cluster is divided into two parts: Scale-up and Scale-out.
Scale-up is the connection between GPUs inside a rack. Nvidia uses its own NVLink, and the fifth-generation NVLink on Blackwell has reached 1.8TB/s of bidirectional bandwidth per GPU. The next-generation Rubin platform will double that again to 3.6TB/s.
Scale-out is the connection between racks and between data centers. It runs on InfiniBand or Ethernet, and the speed per port is shifting wholesale from 400G to 800G, with 1.6T already on the way.
In essence, these two layers of networking determine whether you can assemble tens of thousands of GPUs into a “logical giant GPU.”
NVLink is currently Nvidia’s exclusive weapon, but in 2026 Nvidia opened NVLink up as NVLink Fusion, allowing custom chip makers like Marvell to build NVLink-compatible XPUs and switching chips. The two sides also signed a cooperation agreement worth as much as $2 billion.
The significance of this is that Nvidia is gradually turning the Scale-up ecosystem from a closed loop of its own into an open standard it dominates, while also bringing old switching-chip players like Marvell and Broadcom into its camp.
On the Scale-out side, InfiniBand used to hold roughly 90% of the AI cluster market, but Broadcom-led Ultra Ethernet Alliance is catching up quickly. Hyperscalers prefer Ethernet because they can control the ecosystem themselves. This is a quiet battle over technical direction beyond the chip layer.
Now let’s talk about optical modules.
Connections between GPUs over distances of a few hundred meters have to go through optical fiber, and the device that performs the electro-optical conversion is the optical module.
In AI clusters, each GPU is generally attached to two or three 800G optical modules, and a cluster with several thousand cards will use more than ten thousand modules.
TrendForce’s numbers show that the global AI optical module market will surge to $26 billion in 2026, up about 60% year over year, with 800G and 1.6T as the main force. Global shipments of 800G modules already exceeded 24 million units in 2025 and are expected to reach 63 million units in 2026.
At this scale, this is no longer a “component” business. It is an independent, tens-of-billions-of-dollars track.
The industrial structure of optical modules is also very interesting.
Upstream optical chips (EML, DSP) are concentrated in the hands of a few American and Japanese companies. Midstream module assembly is largely controlled by Chinese manufacturers. Companies like InnoLight and Eoptolink have seen share price gains over the past two years that are no less impressive than those of any semiconductor leader. Downstream switches are dominated by Nvidia’s own Spectrum-X, Broadcom’s Tomahawk, and switching chips from Jais.
The entire interconnect layer is a complete chain from optical chips to DSPs to modules to switching chips and then to switches, and every layer has its own scarcity point.
So when looking at AI clusters, you can’t just look at GPUs. GPUs are the heart, and the interconnect layer is the blood vessels. If the blood vessels are too narrow, even the strongest heart can’t supply enough blood. Today Nvidia’s single-card FP4 computing power has increased several times over, but if the bandwidth of optical modules and switching chips can’t keep up, the computing power of the new-generation GPUs will be trapped in the network and unable to run out.
The interconnect layer is not a supporting role in the industrial chain; it is the gate that determines whether “computing power” can become “effective computing power.”