GlobalFoundries Gets $300 Million to Fix AI’s Data-Movement Bottleneck

AI does not only have a chip problem. It has a data-movement problem.

The U.S. government is awarding GlobalFoundries $300 million to develop silicon-photonics and co-packaged-optics technology at its facilities in Malta, New York, and Burlington, Vermont. The objective is straightforward: move more data between AI processors without burning as much power.

Silicon photonics replaces some electrical connections with optical links. Co-packaged optics moves those optical components closer to the processor. Both matter because an expensive accelerator creates no value while it is waiting for memory or another chip to send data.

GlobalFoundries is targeting 400-gigabit-per-second links and as much as five times the energy efficiency of current implementations. Those are development targets, not verified commercial benchmarks, and the $300 million is research funding rather than immediate product revenue.

The commercial opportunity is still significant. GlobalFoundries does not need to beat Nvidia or AMD in accelerators to benefit from AI growth. It can manufacture the photonics, packaging and connectivity around those processors, where demand rises regardless of which GPU platform wins.

There is also a supply-chain angle. Much of the silicon-photonics and advanced optical-packaging capacity sits outside the United States. GlobalFoundries is positioning itself as an end-to-end domestic supplier, from the silicon wafer through a tested optical module.

The next AI bottleneck will not be raw compute alone. It will be moving data between chips fast enough, cheaply enough and with power the grid can support.