Inside a modern AI training cluster, the GPUs are not the most numerous components. The links are. A cluster of tens of thousands of accelerators is really a fabric of hundreds of thousands of connections, and the speed of the whole system is set by how fast data moves between chips, not how fast any single chip computes. That fabric is in the middle of a materials transition as significant as anything happening in the processors themselves.
The physics is unforgiving. Push an electrical signal through copper faster and it attenuates more over shorter distances, which is why the cables connecting today's racks are thick, short, expensive and hot — some of the most thermally challenging parts of a server rack are the cables.
Why it matters
Two consequences follow. The first is power: at the newest interconnect speeds, moving a bit across a rack electrically can consume as much energy as computing on it, and data center operators increasingly budget interconnect power as a first-class cost. The second is scale: training runs that synchronize across tens of thousands of accelerators stall if the network cannot keep up, so network bandwidth now translates directly into model-training time and therefore into competitive advantage.
The market has noticed. Networking revenue at the AI hardware vendors has grown faster than compute revenue in recent reporting periods, and optical component suppliers — historically a sleepy telecom backwater — have become some of the tightest links in the AI supply chain.
How it works
The transition happens in stages. The first, already mature, is pluggable optical transceivers: modules the size of a thumb that convert electrical signals to laser light at the rack's edge, replacing copper for anything beyond a few meters. The second, now ramping, is co-packaged optics — moving the optical conversion onto the same substrate as the switch chip, eliminating the power-hungry electrical link between chip and module.
The third stage, still emerging, extends the idea to the processors themselves: optical interconnects between chiplets and eventually between boards, using silicon photonics — lasers and waveguides fabricated with semiconductor processes. Each stage trades manufacturing complexity for energy per bit, and each stage moves value from cable and connector makers toward photonics and advanced packaging specialists. Standards bodies are following: the OIF and IEEE have published frameworks for co-packaged optics, and the UALink and Ultra Ethernet consortia are standardizing the scale-up and scale-out protocols that ride on these physical layers.
Evidence
The IEEE 802.3 working group has ratified 800-gigabit Ethernet standards and is deep into 1.6-terabit work, with the technical feasibility documents openly acknowledging that optics dominate beyond short reaches. The Optical Internetworking Forum has published implementation agreements for co-packaged optics architectures. On the protocol side, the Ultra Ethernet Consortium's specification work targets exactly the collective-communication patterns of AI training, and the UALink consortium — whose members include AMD, Broadcom, Cisco, Google, HPE, Intel, Meta and Microsoft — is standardizing the interconnect between accelerators within a pod.
On the supply side, the strain is visible in financial disclosures: optical transceiver suppliers report multi-quarter lead times for their fastest parts, and Nvidia's networking segment — built substantially on its Mellanox acquisition — has grown into one of the company's largest businesses, with its Spectrum-X and NVLink products defining the current generation of AI fabrics.
The competing read
Copper's defenders have a real case for the short term: active electrical cables keep improving, linear pluggable optics preserve the flexibility of replaceable modules, and co-packaged optics has a serviceability problem — when the laser is inside the switch package, a failed laser is a failed switch. Reliability engineers, a conservative tribe by profession, note that pluggables failed gracefully for two decades.
The counterpoint is arithmetic rather than preference: at each doubling of speed, the reach of copper halves, and AI cluster builders have already voted with purchase orders. The debate now is about the pace of co-packaging, not the direction.
What happens next
Watch three indicators. First, whether external laser sources — keeping the delicate lasers in a replaceable module while the optics live in the package — become the accepted compromise on reliability. Second, whether UALink and Ultra Ethernet actually converge the market or split it further, which determines how much of this value stays open to competition. Third, the qualification timelines at the hyperscalers, whose deployment decisions effectively set the industry's roadmap a year before anyone else ships.