Naveen Rao: 4D Computing, AI's Energy Wall & Beating Biology

Naveen Rao discusses his career trajectory from founding Nirvana Systems (the first AI chip company in 2014) and scaling AI infrastructure at Intel, to his current venture Unconventional AI. He frames the core problem confronting AI: Google alone consumes 12 gigawatts of energy for AI services monthly, representing nearly 12% of global data center energy capacity, and at current growth rates, AI energy demand will exhaust available power in approximately 3 years.

Rao argues that the fundamental issue stems from how computers have been designed since the 1940s—they move vast amounts of information between separated memory and compute units. The human brain, by contrast, operates on only 20 watts and moves approximately 16 billion bits per second, while modern GPUs move 30 trillion bits per second in and out of memory. This architectural inefficiency is the primary energy consumer.

To solve this, Unconventional AI is rebuilding computers from first principles using dynamical systems theory—using physics-based oscillatory systems rather than traditional digital abstractions. Rao unveiled Uno, an image generation model built on oscillators, and more significantly, announced the fabrication of the first physical dynamical computer chip, completed in just five months from company founding. This prototype generates images at 500 nanojoules per image compared to millijoules for conventional GPUs—a difference of several orders of magnitude.

The new architecture, termed '4D computing,' integrates compute and memory into unified dynamical elements using three physical dimensions (die stacking, planar layout) plus the time dimension. Rather than moving bits between separated components, the system's physics itself encodes computation and memory simultaneously. Rao positions this as a completely new paradigm beyond CPUs, GPUs, and compute-in-memory architectures.

Regarding commercialization, Rao projects a full product within two years—initially as a data center rack system serving tokens in/out over network. Models must be ported at the model layer rather than operation layer, requiring significant computational transition work but enabling existing model families to run on the new architecture. He recruited an interdisciplinary team spanning dynamical systems theorists, physicists, biologists, and chip architects, coordinated through Python-based libraries that express time-varying stochastic elements.

Rao frames the broader implications through the lens of reaching thermodynamic limits of intelligence per watt. Mammalian brains operate within 1-2 orders of magnitude of thermodynamic limits; current AI is 10 billion times away. He believes achieving 1000x efficiency over 3.5 years will enable ubiquitous distributed computing, mass robotics, and trigger a Jevons Paradox effect creating the largest market humanity has seen.