Artificial intelligence is pushing data centers toward an energy crisis. Projections suggest that by 2028, powering global data-center infrastructure will require more than 70 new nuclear power plants, a staggering figure given that the entire United States currently operates just 53. To cut back on energy consumption, attention is turning to an overlooked culprit: the interconnects responsible for transporting data between chips. Copper interconnects dissipate energy as heat and laser-based optical links offer some improvement, but both technologies are hitting fundamental limits in bandwidth and power efficiency.

Lighthouse investigates a new approach; replacing conventional interconnects with links driven by energy-efficient Gallium Nitride microLEDs that transmit data as pulses of visible light. Because GaN microLEDs can be packed into dense two-dimensional arrays, they unlock a level of parallelism that copper interconnects and laser-based systems simply cannot match – delivering greater bandwidth within the same physical footprint while drawing significantly less power. The project spans the full hardware stack, from fabricating the microLEDs themselves to developing the high-speed electronics needed to drive them, culminating in a functional free-space optical link that can transmit data. Lighthouse offers a concrete demonstration that visible light communication is a viable, energy-efficient foundation for the data infrastructure that next-generation AI will demand.