AI's energy wall: Naveen Rao's dynamical computer and 1,000x efficiency bet

Chapters

1. 0:00 AI and tech

   Naveen Rao's anti-doomer intro and founder track record

   Hosts introduce Naveen Rao, co-founder and CEO of Unconventional AI, touting his record founding Nervana Systems and later selling to Intel. Rao immediately stakes out an anti-doomer position, calling AI the next evolution of humanity. He traces his origin story: programming a home computer in 1978, becoming an electrical engineer, then earning a neuroscience PhD to understand how to make machines intelligent. He explains he founded what he calls the first AI chip company in 2014, sold it to Intel, ran Intel's AI group, and later built GPU platform infrastructure that joined Databricks.

2. 3:07 AI and tech

   Unconventional AI's 1,000x power-efficiency mission

   Rao describes Unconventional AI as a top-to-bottom company rethinking computing from first principles. The goal is 1,000x power efficiency, revised from five years to three and a half because progress has been faster than expected. The company starts with theorists who devise ways to move less information, translates ideas into models trained on real data, then architects and models physical circuits, and finally builds boards and products. Rao frames the bet as removing abstractions, connecting semiconductor physics directly to neural networks, rather than layering digital logic and software on top.

3. 4:37 AI and tech

   AI's energy wall: Google, gigawatts, and the cliff

   To show energy is the real constraint, Rao cites Google's 3.2 quadrillion tokens processed per month. Using a conservative 10 joules per token, that implies 12 gigawatts of continuous power for one company's AI services, versus roughly 40 gigawatts for all US data centers and under 100 gigawatts globally. He estimates the industry runs out of energy in about three years. Graphically, exponential AI market growth meets near-linear available energy, creating a gap he argues only new hardware can close. Rao frames this as the central problem that technology must solve.

4. 6:13 Science

   Biology's proof and the cost of moving bits

   Rao explains data center thinking has shifted from floor space to networking to GPUs to energy contracts; he says about 50% of the cost of serving a token is energy. Biology is his proof that intelligence can be cheap: the human brain runs on 20 watts, a monkey brain on 1 watt, a squirrel brain on 8 milliwatts. He contrasts bits moved: human cortex moves ~16 billion bits per second while a GPU moves ~30 trillion bits per second in and out of memory, making data movement the main energy drain. He notes today's computers still use the 1940s von Neumann model.

5. 9:18 Science

   Rethinking computation with dynamical systems

   Rao argues Moore's law is no longer delivering efficiency gains, so the industry must cut out the middleman. Digital one/zero abstraction is lossy; each layer of abstraction throws away physics and adds inefficiency. He points to computation in nature: flocking birds, ant colonies, and the brain itself are dynamical systems where simple elements produce emergent behavior. As an example, metronomes on a movable plank synchronize through physical interaction alone. The idea is to build circuits whose physics, not software algorithms, perform computation, using time and coupling instead of stored-program fetch-execute.

6. 12:59 AI and tech

   Uno: scaling image generation with sparsity

   Rao introduces Uno, an open-source image generation model built from coupled oscillators, the first demonstration that a dynamical system can be scaled, trained, and produce useful output. He shows state-space trajectories where conditioning on a target like 'airplane,' 'car,' or 'bird' sends the system down different paths. The team then found sparsity: eliminating most pairwise connections does not hurt and actually improves trainability, scalability, and performance. He calls this a rare efficiency holy grail, and frames it as a problem that required the right framing to solve.

7. 15:33 AI and tech

   First physical dynamical computer and 4D computing

   Rao reveals the first physical dynamical computer, built in five months. The company sent the tape-out to the fab on June 1 and now has the chip back in the lab with results. It generated the first images ever produced by such a computer at roughly 500 nanojoules per image, versus millijoules for a GPU — many orders of magnitude more efficient because it does not move information across a memory interface. He calls this 4D computing: three physical dimensions from die stacking plus time as the fourth dimension, with compute and memory unified in each element.

8. 17:37 AI and tech

   Beating biology and Jevons paradox

   Rao lays out implications of 1,000x efficiency: intelligence per watt can be pushed toward a thermodynamic limit that mammalian brains sit within one or two orders of magnitude of, while today's systems are 10 billion times away. He expects to hit the limits of 2D lithography in three and a half years. The broader vision shifts from gigawatt-scale data centers to many small, local, environmentally friendly ones, and eventually billions of robots that dynamically assemble. He invokes Jevons paradox: dropping the cost of AI 1,000x will expand consumption beyond the price drop and create humanity's largest market.

9. 19:41 AI and tech

   Q&A: path to product and porting models

   In Q&A, the host asks what it takes to scale from prototype to deployed hardware. Rao says Unconventional AI is within two years of a full product: a rack-scale system sold as a data center product, accepting tokens over a network cable but with completely different internals. Existing models will work, but porting happens at the model layer rather than the operations layer, requiring significant compute. He describes the team as theorists from dynamical systems research plus chip builders who had to learn to talk to each other, and says the software stack is Python-based libraries that express time-varying stochastic elements.