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

Summary

This episode centers on Naveen Rao’s thesis that AI scaling is increasingly constrained by energy, not just chips or data-center capacity. He argues current systems waste most power moving data around, while biological systems like brains are far more efficient because they embed computation in physical dynamics. Rao says Unconventional AI is building a "dynamical computer" and calling the approach "4D computing" to unify memory and compute, reduce information movement, and target orders-of-magnitude better efficiency. He claims the team has already taped out a first physical prototype, demonstrated an open-source oscillator-based image model called uno, and measured energy use around 500 nanojoules per image. The long-term product is described as a rack-scale system that can eventually support model migration, distributed data centers, and even robotics, with a full product target in about two years.

Notes

• Rao frames AI scaling as an energy problem, not just a silicon or real-estate problem.
• He claims AI demand is already large enough that token throughput translates into multi-gigawatt power needs.
• He uses the human brain as the benchmark for efficient intelligence.
• He argues the main inefficiency in computing is data movement, not arithmetic.
• He says classic computer architecture has barely changed since the 1940s.
• The company’s core concept is a dynamical computer with unified memory and compute.
• He defines 4D computing as physical 3D plus time, used to exploit dynamics rather than fight them.
• The company is organized as a full-stack effort from theory to chips to systems.
• An open-source oscillator-based model called uno is the team’s early software proof point.
• The model evolves as a state-space trajectory conditioned on the desired output.
• He uses sparsity to address N-squared scaling in highly connected systems.
• He says the first physical dynamical computer was built in five months.
• The prototype is claimed to achieve roughly 500 nanojoules per image.
• He says the energy gain comes from avoiding information movement.
• He explicitly contrasts the new machine with von Neumann systems.
• Rao says the company wants to beat biology on efficiency and spread compute into robots and smaller distributed sites.
• He says the first commercial product will be a rack-scale data-center system.
• The product is positioned as a networked rack rather than a conventional accelerator board.
• He says migration will happen at the model layer, not the operations layer.
• Existing models should work, but transition will require substantial compute.
• The software layer is Python-based and meant to express time-varying stochastic behavior.
• Python libraries are the coordination layer between theory and physical implementation.
• He says the company has already achieved results from first silicon in the lab.
• He sees the architecture as a new category rather than a faster version of CPUs or GPUs.
• He argues the industrial opportunity is broad if energy efficiency improves enough to change deployment economics.
• He implies the product moat may come from system-level efficiency rather than model novelty alone.
• He describes a long-term path to compute everywhere, including robotics.
• The team claims a hardware-software co-design stack spanning theory, model validation, and circuit design.
• Sparsity is presented as both a scaling strategy and a way to preserve or improve behavior.
• The technical thesis is that physics can substitute for expensive information shuttling.
• Rao’s broader investment case is that lowering compute energy could create more demand, not less.
• He says the company is pursuing a new hardware category that could eventually outcompete biology on efficiency.
• He describes a first product that is still compatible with today’s networking and deployment patterns.
• He argues the transition burden should be manageable because the model interface stays familiar.
• He positions the system as a platform, not a single model breakthrough.
• The technology pitch is fundamentally about changing the economics of AI infrastructure.