AI optical networking rarely gets the attention given to a new Nvidia processor. Yet a powerful chip is only useful if the rest of the machine can keep it supplied with information. In a data centre full of accelerators working together, the connections deserve their own place in the story.
On 17 September 2026, Marvell and GlobalFoundries announced a multiyear expansion of manufacturing capacity for silicon-germanium technology at GF’s Burlington, Vermont, facility. The chips support high-speed optical connections used in AI infrastructure. This is an expansion of an existing production relationship, rather than the unveiling of a finished new AI computer. GlobalFoundries announcement
The race to move the answer
Think of a group project in which every participant works exceptionally quickly but the shared documents arrive late. Buying each person a faster calculator would not fix the delay. That is a useful way to understand why computing power and communication capacity have to develop together.
Optical links carry information using light. But they still need electronics at the interfaces. GF says its silicon-germanium platform supports 200 gigabits per second per lane and is intended for pluggable optical modules as well as optics placed near, or alongside, major chips. The announcement does not specify the extra wafer volume, investment value or a single date when all the added capacity will be available. Agreement details
Putting light closer to the processor
A pluggable transceiver is a replaceable module that sends and receives optical signals. Co-packaged optics moves optical functions into the same package as the relevant electronic chip, shortening the electrical journey before data becomes light. Near-packaged optics is another approach to bringing those functions closer together. GF’s technology portfolio supports multiple arrangements rather than requiring an immediate industry-wide switch. GF’s silicon-photonics overview
The trade-off is practical as well as technical. Pluggable modules are familiar and serviceable. Tighter integration can improve bandwidth density and energy efficiency, but a network also needs to be manufactured, tested and repaired. The most elegant laboratory arrangement is not automatically the easiest product to deploy across thousands of racks. GF’s own comparison highlights modularity on one side and shorter electrical paths on the other. Platform trade-offs
A different kind of chip advantage
The interesting implication is that an AI supply chain has several kinds of winners. The company designing an accelerator, the foundry making a specialised interface, and the team assembling the optical package solve different problems. A headline about the smallest transistor does not tell us whether the whole installation will run efficiently.
For operators, the useful questions are now concrete: how much data can the connection move, how much electricity does it consume, and can enough reliable parts be delivered on time? This agreement addresses that final question. It does not, by itself, prove a particular percentage improvement in AI performance.
Our view: watch AI optical networking as part of the system, not as an accessory to the GPU. The next expensive bottleneck may be the journey between two extremely fast chips.
Related reading: why Nvidia and Google want data centres that can adjust their electricity demand.
Featured photograph: optical fibres, a representative image. Denny Müller / Unsplash.
Leave a Reply