Silicon photonics work for data centers speeds up through investments

In short: Silicon photonics is being pushed forward for AI data centers mainly through company investments and partnerships, not one formal group effort.

What's going on

Silicon photonics is a way to move data using light on and around computer chips (like swapping some copper wires for tiny optical lanes). It is getting more attention because AI data centers need to move huge amounts of data between chips and servers, and today’s connections are becoming a bottleneck.

There is no clear public evidence, based on the available information, that multiple semiconductor groups have formally “joined forces” in one specific silicon photonics consortium aimed at speeding up data center systems. Instead, progress appears to be coming from companies competing, expanding production, and signing one to one partnerships.

One example is Tower Semiconductor. It plans to sharply increase silicon photonics manufacturing capacity, including a reported $300 million investment across factories in the US, Israel, and Japan. Tower has said it expects to double capacity by the end of 2025 and triple it by mid 2026.

Another big push is co packaged optics, often shortened to CPO. This means putting optical connections closer to the computing chips (like moving a router closer to the device so the cable run is shorter). Reports point to TSMC leading work on CPO, with partnerships involving Nvidia and Broadcom. Industry forecasts also suggest silicon photonics will take a growing share of optical transceivers, which are the parts that send and receive data between machines.

What to watch

Watch for when CPO moves from demos into regular data center products, and whether higher speed links like 1.6T and 3.2T become common. Also watch which manufacturers can scale production reliably, since making these parts in large volumes is one of the hardest steps.

Source: Financial Times