TLDR
Naveen Rao, founder of Unconventional AI, argues that current AI hardware is fundamentally limited by energy inefficiency due to the von Neumann architecture's need to move data between memory and compute. He presents a new 'dynamical computer' inspired by biological brains that merges compute and memory using physical dynamics and 3D stacking, demonstrating a prototype chip that generates images at 500 nanojoules — orders of magnitude more efficient than GPUs. Rao aims for 1,000x power efficiency within 3.5 years, enabling ubiquitous compute and a massive market expansion via Jevons Paradox. The company, built in five months, combines theorists, neuroscientists, and chip architects to rethink computing from first principles. A full rack-scale product is targeted within two years. Existing models can run but require porting at the model layer to the new time-varying paradigm. The approach represents a fundamental shift from digital abstractions to physics-based computation. Rao is optimistic about AI as humanity's next evolutionary step, not a doom scenario. The talk includes a live demo of the first physical dynamical computer and discussion of the team's interdisciplinary challenges. The energy crisis in data centers is framed as the primary bottleneck, with power contracts now preceding hardware deployment. The new architecture eliminates the memory interface entirely, using each computing element as memory. The goal is to beat biological efficiency and enable billions of robots. The company uses Python libraries for programming stochastic time-varying elements. The prototype was taped out June 1 and is already producing results. The talk concludes with Q&A on ecosystem, compatibility, and team composition.
For you
PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.
Sign in to unlockSurprising Insights
The first physical dynamical computer was designed, taped out, and returned from fab in just 5 months (January to June) by a team that didn't fully exist at the start.
Sparsity — throwing away connections — not only reduces complexity but actually improves system behavior, trainability, and performance, a rare 'holy grail' in computing.
The chip generates images using only 500 nanojoules (10^-9 J), while GPUs use millijoules (10^-3 J) — a million-fold difference — because it doesn't move information around at all.
The programming model is not CUDA or MatMul; it's Python libraries expressing time-varying stochastic dynamics. Existing models work but must be ported at the model layer, not the operator layer.
Data center planning has inverted: you now secure the energy contract first, then figure out how to fill it with hardware. Power is the new primary constraint.
Key takeaways
Current AI hardware hits an energy wall: data centers are limited by power, not space or GPUs; 50% of token serving cost is energy; Google alone uses ~12 GW for AI — a significant fraction of global data center capacity.
Biological brains prove extreme efficiency is possible: human cortex runs on ~20W moving 16 Gbps; a squirrel brain uses 8 mW for precise motor control — 100x more efficient than a phone. Synthetic systems move 30 Tbps+ off-chip, wasting energy on data movement.
Unconventional AI built the first physical 'dynamical computer' in 5 months: a chip that computes via coupled oscillators (like metronomes synchronizing) with compute and memory unified, no von Neumann bottleneck, using 4D computing (3D stacking + time dynamics).
Prototype achieves 500 nanojoules per generated image — many orders of magnitude below GPU millijoules — by eliminating data movement. Sparsity in dynamical systems enables better scaling, trainability, and performance simultaneously.
Goal: 1,000x power efficiency in 3.5 years, beating biology's thermodynamic limit. This will shift compute from massive centralized data centers to distributed, local, environmentally friendly systems, enabling billions of robots and creating the largest market in history via Jevons Paradox.
Notable quotes
“I'm the opposite of an AI doomer. I think AI is the next evolution of humanity.”
▸ Notable quotes · 0:21“We need innovation on the hardware substrate to actually build true intelligence.”
▸ Notable quotes · 0:26“The human cortex only moves about 16 billion bits per second. A GPU moves nearly 30 trillion bits in and out of memory per second. That's what drives the energy demand.”
▸ Notable quotes · 8:48“This is actually the first physical dynamical computer ever built. We did this in five months.”
▸ Notable quotes · 15:27“If you make something one one-thousandth the price, you'll consume more than one one-thousandth of it. And I think this will create the largest market that humanity's ever seen.”
▸ Notable quotes · 19:19Chapters
0:00 — Introduction and Background: Naveen Rao introduces himself as a lifelong technologist who founded the first AI chip company (Nirvana Systems), led AI at Intel, built MosaicML (acquired by Databricks), and now leads Unconventional AI to rethink computing from first principles.
4:43 — The Energy Crisis in AI: Data centers are hitting a power wall: Google processes 3.2 quadrillion tokens/month requiring ~12 GW; global data center capacity is ~100 GW. Energy is now the primary constraint — 50% of token serving cost — and demand will outstrip supply in ~3 years.
7:07 — Biological Proof of Efficiency: Brains demonstrate extreme efficiency: human cortex 20W, monkey 1W, squirrel 8mW. They move far less data (16 Gbps vs 30 Tbps for GPUs) because compute and memory are unified. This proves the right physical substrate for intelligence exists.
12:06 — Dynamical Systems and the Metronome Analogy: Computing can be done via physical dynamics (coupled oscillators) rather than digital abstractions. Metronomes on a movable plank synchronize through physics alone — a model for how simple local interactions create emergent computation without data movement.
13:16 — Unconventional AI's Dynamical Computer: The company built 'Uno', an image generation model on simulated oscillators, then a physical chip using sparsity to improve scaling and performance. The first prototype taped out in June, generates images at 500 nJ — orders of magnitude more efficient than GPUs.
17:01 — 4D Computing Architecture: The new architecture merges compute and memory (no memory interface), uses 3D die stacking plus time dynamics (4D computing), and eliminates the von Neumann bottleneck. It's programmed via Python libraries for stochastic time-varying systems.
17:43 — Implications and Future Vision: 1,000x efficiency enables compute everywhere: distributed data centers, robots, and a massive market via Jevons Paradox. Goal is to beat biology's thermodynamic limit in 3.5 years. Team spans theorists, neuroscientists, and chip architects — interdisciplinary coordination is the main challenge.
19:34 — Q&A: Ecosystem, Compatibility, and Team: Product in 2 years as a rack-scale system. Existing models work but need model-layer porting; no MatMul, uses time-varying dynamics. Team combines dynamical systems theorists (100-year-old field) with chip builders — bridging them is the hardest part. Software is Python-based.
Speakers
Naveen Rao — Co-founder and CEO of Unconventional AI; former founder of Nirvana Systems, head of AI at Intel, founder of MosaicML
Interviewer/Host — Conference host conducting Q&A after the talk