Jensen Huang has a habit of making big predictions, and his latest one involves the pipes, not the chips. At Nvidia’s GTC Berlin event, the CEO said that optical networking in AI data centers could drive demand for optical components up by 15 times, a figure that reflects just how severely current interconnect technology is straining under the weight of modern AI workloads.
The number is striking, but the logic behind it is straightforward. As AI models grow larger and data centers pack in more and more accelerators, the cables and connectors linking those chips together become a genuine bottleneck. Copper, which has handled that job for decades, is running into physical limits around power consumption, signal degradation, and bandwidth density. Optical interconnects, which use light instead of electrical signals, sidestep many of those constraints.
The plan: 15x by mid-2027, 30x by end of 2028
Huang’s comments came alongside a concrete manufacturing roadmap. Nvidia, working with packaging partner Siliconware Precision Industries, plans to expand co-packaged optics optical-engine capacity by 15 times by mid-2027. That expansion is then expected to double again, reaching a total increase of 30 times by the end of 2028.
Co-packaged optics, often shortened to CPO, is the technology at the center of this push. Rather than connecting optical transceivers externally to a switch chip, CPO integrates the optical engines directly alongside the silicon, cutting the distance signals need to travel and improving both power efficiency and signal integrity in the process.
Nvidia is introducing its Spectrum-6 CPO switches as the flagship product built around this architecture. These switches are designed specifically for what Nvidia calls its Vera Rubin AI factories, the next generation of large-scale AI infrastructure the company is building toward. The Spectrum-6 line promises higher bandwidth density and reduced energy consumption compared to existing networking gear.
Huang framed this as the beginning of one of the largest infrastructure buildouts in history, driven by demand that is extraordinary even by the standards of an industry accustomed to exponential growth curves.
The broader manufacturing investment behind the switch
The optical capacity expansion is one piece of a larger investment picture Huang laid out at GTC Berlin. Nvidia has committed $1.1 billion to advance US semiconductor testing capacity through a partnership with KYEC, with production targeted to begin in April 2027.
Separately, Amkor, another key packaging partner, plans to grow its interposer wafer production capacity by over 300% by the end of 2028. Interposers are the substrates that allow multiple chiplets or dies to connect to each other at high speed within a single package. A more-than-tripling of that capacity signals that Nvidia’s partners are betting heavily on continued demand growth.
Why optical, and why now
The shift from copper to optical interconnects in data centers has been discussed in the networking industry for years, but the economics and engineering have only recently aligned in a way that makes large-scale deployment practical. AI training clusters, which link thousands of accelerators in tightly coordinated networks, have accelerated that timeline considerably.
The move toward CPO specifically represents a further refinement of that logic. By integrating the optical engine directly into the switch package rather than treating it as an external module, CPO reduces the electrical path between the chip and the fiber, cuts latency, and improves yield on power consumption. The bottleneck in CPO deployment has historically been the ability to produce optical engines at scale, not the demand for them.
Huang’s prediction that volume could increase 15 times is therefore less a forecast about the distant future and more a statement about what Nvidia believes its own infrastructure roadmap requires. The company is not predicting that some abstract market will grow; it is describing the scale it intends to build toward, and announcing the manufacturing partnerships it has lined up to get there.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.