Unfortunately, this isn’t about capacity being sold out, but rather that the coherent transceiver market is still very limited mainly to frontend connections (inter-data center traffic) and only for those customers whose stack and software support coherent optics. Ciena, like Nokia, only has a coherent optics portfolio. For example, Coherent (the company) and a million other transceiver manufacturers also have traditional lasers in their portfolios.
Nokia’s (Infinera’s) ICE-D, on the other hand, is a coherent transceiver family optimized for the backend—intra-data center, short-range, ultra-low latency data transfer.
So, while the optical transceiver business is currently growing in data centers, it is primarily still through traditional bi-directional 800G and 1.6T OSFPs, which are for backend connections, inside the rack / inside the data center. The coherent optics business is growing mainly in the frontend, i.e., external data center connections—of which Nokia’s massive Google deal is a good example.
For instance, NVIDIA has now made Vera Rubin compatible with coherent backend pluggables, but CPO (Co-Packaged Optics) anywhere other than switching front-panel connections is still in the future. Google is not even moving toward CPO technology in its own system but uses OCS “mirroring.”
If and when NVIDIA moves fully to coherent lasers in both the backend and frontend, the entire interconnection business will shift from copper to optics, and from traditional connector firms like Amphenol, Molex, and TE to optical firms—but only to those who have CPO coherent optics for the entire system, from the GPU to inter-data center connections. This equals massive energy savings (25% lower power consumption than current levels), and fiber as a material is light and cheap compared to copper and doesn’t require, for example, retimer chips on the transmission line.
Nokia is the only one capable of offering the entire system. But will there be others in a year or two when the successor to Vera Rubin is announced?
A small calculation of the business case with NVIDIA, utilizing AI. Also, a bit of background on the technology whose business Nokia could capture—even as a single source if all the stars align. I know the products and the technology, so the analysis below is based on factual background information regarding what would happen if all high-speed copper were replaced by fiber.
If Nvidia successfully transitions to a fully optical GPU rack architecture, the value shift from traditional copper and electrical networking to optical interconnection content will be astronomical.
For a fully optical rack architecture to work, copper trace lines on the printed circuit board (PCB) are replaced by optical waveguides, and standard pluggable transceivers at the faceplate are replaced by co-packaged optics (CPO) right next to the ASIC.
By anchoring themselves to Nvidia’s next-generation AI stacks (like the liquid-cooled Blackwell NVL72 and the upcoming Vera Rubin architectures), Nokia stands to capture massive value across three distinct layers of the physical AI stack.
In current architectures like the Blackwell NVL72, Nvidia uses a massive copper backplane (Co-Packaged Copper) to link 72 GPUs into a single logical “Super-GPU.” This is the physical limit of copper; the cables are already incredibly stiff, heavy, and limited to less than 2 meters. For Vera Rubin, Nvidia intends to expand this single-image compute pool across multiple racks. To do this, the NVLink fabric must become fully optical.
Nokia’s Value Capture: Nvidia needs ultra-dense, ultra-low-latency optical distribution frames and short-reach switching matrices. Nokia’s expertise in high-density fiber management, automated optical patching, and ICE-D (intra-data center coherent/linear optics) gives them the perfect product portfolio to act as the “inter-rack glue.”
The Content Value: In a standard data center, networking represents roughly 10–15% of the total infrastructure cost. In an all-optical, multi-rack AI cluster, the cost of the optical interconnect fabric (the fibers, CPO connectors, and optical routing) jumps to 20–25% of the multi-billion-dollar cluster cost.
What does this look like in dollars per rack?
To put the financial scale into perspective, let’s look at how the bill of materials (BOM) changes for a next-generation AI cluster deployment:
| Era | Architecture | Networking Component Profile | Estimated Optical BOM per Rack |
|---|---|---|---|
| Hopper (H100/H200) | Copper inside rack, Pluggable Optics outside. | Traditional 800G optical transceivers, standard InfiniBand switches. | ~$50,000 – $100,000 |
| Blackwell (NVL72) | Co-Packaged Copper backplane, LPO/Linear optics for scale-out. | Ultra-dense copper wiring, high-speed 800G/1.6T OSFP pluggables. | ~$150,000 – $250,000 |
| Vera Rubin (All-Optical) | Full CPO Rack Architecture + Optical NVLink. | Integrated InP laser arrays, external laser sources (ELS), co-packaged optics, optical switch fabrics. | $500,000+ |
It’s worth considering that NVIDIA is currently manufacturing one Blackwell rack every two hours…