Unconventional AI bets on new hardware to break the energy wall

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

As conventional AI systems demand ever more energy, Naveen Rao argues that changing the underlying computing architecture may be unavoidable.

3 key takeaways
  1. 1Unconventional AI is developing time-varying, dynamical-systems hardware that departs from conventional matrix multiplication.
  2. 2The company aims to turn an early prototype into a full data-center rack within roughly two years.
  3. 3Porting existing models requires substantial model-layer work, making migration effort a central tradeoff alongside performance gains.

Don't miss

Rao explains how Unconventional AI replaces conventional matrix-multiplication framing with time-varying computation based on states and transitions.

The brief

Naveen Rao, the CEO of Unconventional AI, argues that AI’s energy demands call for more than incremental hardware improvements: the computational substrate itself may need to change.

Unconventional AI is targeting a full data-center rack within roughly two years, with tokens entering and leaving over a network while the underlying computation works differently.

Rao describes a time-varying architecture that moves beyond conventional matrix multiplication, treating each step as a current state interacting with a transition matrix.

The approach can support existing models, but migration requires substantial compute and work at the model layer, creating a direct tradeoff between performance gains and compatibility.

The company’s hardest organizational problem may be as unusual as its hardware: bringing dynamical-systems theorists together with experienced chip designers.

What was said on this episode

27 statements · 19 positive · 2 negative · 1 mixed · 5 neutral

  1. AI will drive humanity’s next evolutionary stage

    “I think AI is the next evolution of human humanity.”

    Listen at 0:23

  2. Naveen Raoon AI hardwarePositive0:26

    Hardware innovation is needed to build true intelligence

    “We need innovation on the hardware substrate to actually build true intelligence.”

    Listen at 0:26

  3. AI is highly transformational and will enable further human evolution

    “I think AI is one of the most transformational technologies that humanity's ever created and will enable us to get to that next level of evolution”

    Listen at 0:58

  4. Unconventional AI aims to achieve 1,000x power efficiency within 3.5 years

    “the goal has been within— it was initially within 5 years to get to 1,000x power efficiency. I've actually revised this to 3.5 years because things have gone faster than we anticipated.”

    Listen at 3:41

  5. Naveen Raoon AI data-center energy supplyNegative5:45

    AI data-center demand will exhaust available energy in about three years

    “So we're going to run out of energy pretty fast. In like 3 years or so is my estimate.”

    Listen at 5:45

  6. Naveen Raoon AI token servingNeutral6:38

    Energy accounts for about half of the cost of serving an AI token

    “About 50% of the cost of serving a token, so every time you try something on ChatGPT, 50% of that cost is energy.”

    Listen at 6:38

  7. Naveen Raoon Unconventional AI hardwarePositive7:01

    Unconventional AI hardware will monetize each watt 1,000 times better

    “We're going to monetize that 1,000x better than existing hardware.”

    Listen at 7:01

  8. Naveen Raoon Squirrel brains and smartphonesPositive7:52

    A phone can support over 100 squirrel-brain equivalents with accurate behavior

    “You could run over 100 squirrel brains on your phone, and it has very precise and accurate behavior.”

    Listen at 7:52

  9. Naveen Raoon Biological brainsPositive7:58

    Biological neural systems provide the right substrate for intelligence

    “Biology's created something quite incredible. In fact, it's the right kind of physical substrate for intelligence.”

    Listen at 7:58

  10. Naveen Raoon Moore's lawNegative10:09

    Transistor scaling no longer produces continuing efficiency gains

    “And now we're actually not scaling efficiency any longer. Moore's Law, if you may have heard of this, is like making transistors smaller, has largely ended.”

    Listen at 10:09

  11. Naveen Raoon UNOPositive13:16

    UNO demonstrates oscillator-based image generation

    “we released a model we called UNO, which actually demonstrated this. It's an image generation model that's built on a set of oscillators like that.”

    Listen at 13:16

  12. Naveen Raoon Sparse physical neural systemsPositive14:59

    Sparsity can improve system behavior and trainability

    “you can actually not only throw away some of the connections, but you can actually get better behavior out of the whole system. It actually becomes more trainable.”

    Listen at 14:59

  13. Naveen Raoon Unconventional AI physical dynamical computerPositive15:36

    Unconventional AI built the first physical dynamical computer

    “This is actually the first physical dynamical computer ever built.”

    Listen at 15:36

  14. Naveen Raoon Unconventional AI physical dynamical computerPositive16:18

    The physical dynamical computer uses about 500 nanojoules per image

    “it's only 500 or so nanojoules per image”

    Listen at 16:18

  15. Naveen Raoon Unconventional AI physical dynamical computerPositive16:33

    The dynamical computer is far more energy-efficient than standard computers

    “So it's many orders of magnitude more efficient than a standard computer.”

    Listen at 16:33

  16. Naveen Raoon 4D computingNeutral17:20

    4D computing combines temporal dynamics with three-dimensional die stacking

    “we call this 4D computing, where we use the time dimension and the dynamics, and we actually use the physical 3 dimensions of die stacking”

    Listen at 17:20

  17. Unconventional AI expects to reach 2D lithography limits within 3.5 years

    “We think in that 3.5 years, we can hit the limits of 2D lithography.”

    Listen at 18:13

  18. Unconventional AI aims to exceed biological intelligence efficiency

    “the overarching goal of this company is to beat biology.”

    Listen at 18:18

  19. Naveen Raoon Unconventional AI computingPositive18:21

    Unconventional AI will enable ubiquitous computing and new robotic forms

    “We want to make something better and enable compute everywhere, including compute in new robotic forms and things like this in the next decade or so.”

    Listen at 18:21

  20. Naveen Raoon Data-center infrastructureNeutral18:34

    Data-center infrastructure will shift from large sites to many smaller sites

    “we'll see the shift going from big, big data centers with gigawatts to many small data centers all over the place.”

    Listen at 18:34

  21. Naveen Raoon Dynamically assembling robotsPositive18:50

    The technology will enable billions of dynamically assembling robots

    “enabling the ability to build billions of robots that kind of dynamically assemble to solve problems in our world”

    Listen at 18:50

  22. Naveen Raoon AI computing marketPositive19:27

    A 1,000-fold efficiency improvement could create humanity’s largest market

    “I think this will create the largest market that humanity's ever seen.”

    Listen at 19:27

  23. Unconventional AI is within two years of a full product

    “Yeah, time-wise, we're within 2 years of getting it to a full product.”

    Listen at 20:06

  24. Unconventional AI expects a full product within two years

    “time-wise, we're within 2 years of getting it to a full product.”

    Listen at 20:06

  25. Naveen Raoon Existing AI modelsMixed21:13

    Existing AI models will work after porting but require substantial transition compute

    “So yes, the existing models will work, but there's a fair bit of compute required to make that transition happen.”

    Listen at 21:13

  26. The architecture implements computation as time-varying behavior

    “It implements it as sort of time-varying behavior.”

    Listen at 21:28

  27. Naveen Raoon Unconventional AI Python librariesPositive22:18

    Unconventional AI’s Python libraries express stochastic time-varying elements

    “It's a language of sorts that allows you to kind of express time-varying elements that have stochastic behavior.”

    Listen at 22:18

Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.

Weekly email

State of the Pod, every Monday

The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.

Listen to the full episode and explore every guest, topic, and moment on PodLume.