AI has entered a phase where the limits are no longer set by algorithms or transistor counts but by energy and speed.
For years, the industry relied on a simple strategy: Deploy more Graphics Processing Units (GPUs). That approach is now colliding with physics. Datacenters push megawatts of power, airborne and orbital systems cannot dissipate power, and edge computing devices live under strict energy consumption. Scaling electronic computing alone is no longer feasible.
Researchers at the University of Florida are seeking a new direction. In work published in Photonics Research, UF engineers demonstrate a hybrid photonic–electronic neural network that shifts one of the most energy-intensive stages of AI — feature extraction — into light itself. The result is dramatic: 6.5 times faster training and nearly 1,000 times higher energy efficiency than an all-electronic baseline.
Read more about America’s Next AI Accelerator Constraint: Heat, Power and the End of ‘Just Add GPUs.’
