What happened in this episode?
Naveen Rao, co-founder and CEO of Unconventional AI and a former Intel AI executive, outlines a hardware-led attempt to push generative AI beyond what he describes as an impending energy constraint. He argues that electricity has become the dominant bottleneck for data centres, offering a back-of-the-envelope estimate that one major company’s AI services could already be consuming gigawatts of power and predicting the industry could hit an “energy wall” within roughly three years if model size and demand keep rising.
Rao’s core claim is that his company has built what he calls a “dynamical computer” that co-locates compute and memory, avoiding much of the energy spent moving data in conventional von Neumann architectures. Drawing on dynamical systems concepts (including synchronisation examples), he describes an oscillator-based generative model the company open-sourced as a demonstration.
He then says, for the first time publicly, that Unconventional AI has taped out a chip and has it back in the lab, showing “first ever images” generated from the physical system. Rao claims the prototype operates at “500 or so nano per image”, asserting orders-of-magnitude better energy efficiency than GPU-based generation, while also describing sparsity techniques intended to improve scalability and trainability.
In brief Q&A, Rao says the path to a product is a rack-scale data-centre system managed like a VM, with an estimated two-year timeline, and that existing models would require work to port at the model layer rather than the operations layer.
