Squirrel-Brain Chips: Naveen Rao's 4D Computer and AI's Energy Wall

An Anti-Doomer Opening

The All-In Podcast stage was less a corporate demo and more a thesis pitch. Naveen Rao positioned himself squarely against the doomer narrative and framed artificial intelligence as humanity's next evolutionary step. In his telling, software scale alone will not carry that leap — the physical substrate beneath it must be rebuilt. A childhood curiosity sparked by a home computer in 1978 led him through electrical engineering into science-fiction-inspired questions about how to make a machine intelligent, then back to school for a neuroscience doctorate. That personal arc gave the technical argument a human prologue: you claim to understand intelligence only when you can build it.

The entrepreneurial path has tested that claim in practice. In 2014, when the phrase artificial intelligence had barely entered everyday business language, he founded what he describes as the first AI chip company, Nervana Systems, a hardware bet that was hard to sell at the time. The Intel acquisition came early by his own account, followed by building and leading Intel's AI group. After that chapter, the question grew larger: how do you build infrastructure to train big models like today's language systems. Platforming GPUs and making large-scale training usable for others became MosaicML, which accelerated after generative AI broke through in late 2022 and joined Databricks in 2023. His note that roughly a quarter of Databricks revenue now traces to that effort helps explain why the scale story is taken seriously.

Unconventional AI is presented not as a continuation but as a first-principles rethink. The goal is to redesign the computer that has operated on largely the same principles for decades, with a single purpose of radical power efficiency. The original aim of a thousand-fold gain in five years has been pulled forward to three and a half years because progress has been faster than expected, interestingly accelerated by AI itself on deep scientific problems. The company is vertical top to bottom: theorists with math and neuroscience backgrounds propose ideas that cut information movement, modeling teams turn those ideas into systems trained on real data against real criteria, circuit architects and silicon designers embody the physics, and a systems team closes the loop to board and rack.

The Energy Wall and Biology as Proof

Rao answered whether energy is really a problem with a single-company snapshot. Google, cited because it talks publicly, processes about 3.2 quadrillion tokens per month. Even at a low-end assumption of 10 joules per token, that multiplies to roughly 12 gigawatts. The United States devotes about 40 gigawatts to data centers, about half of global capacity, leaving world data-center power below 100 gigawatts. In that arithmetic, one provider's generative services already consume more than a tenth of global data-center power. If model size and demand both keep rising, an encounter with a hard physical ceiling within about three years is the projection he offered. Visualized, a steeply exponential market curve toward a trillion-dollar opportunity in 2030 diverges from a linearized energy curve below, and the widening gap is the problem to solve.

That gap has already reshaped how data centers are thought about. The primary constraint used to be floor space, then networking gear, then GPUs; today it is energy. The first question is the energy contract, and only then how to fill it with infrastructure. About half the cost of serving a token is energy, the rest is hardware capital, floor space and related spend. The operating logic has simplified to securing a power contract and monetizing every watt. Rao's business case sits cleanly in that logic: monetize the same watt a thousand times better than incumbent hardware.

Biology enters as an existence proof. The human brain runs on about 20 watts, a monkey brain near 1 watt on a linear scaling, a squirrel brain around 8 milliwatts. Your phone burns about 1 watt, so you could in theory power more than a hundred squirrel brains on it, and that tiny brain executes an extremely precise behavior — branch to branch jumps with near-perfect accuracy thousands of times. Rao's favorite maxim reappears here: you do not truly understand something until you can create it. Today's intelligent systems produce impressive outputs but via an inefficient path, and the culprit is mostly moving information around.

The numbers make the inefficiency concrete. The human cortex moves about 16 billion bits per second, a high-end graphics processor moves close to 30 trillion bits per second off-chip to memory, and on-chip movement is ten to a hundred times higher. Synthetic systems shuffle orders of magnitude more bits than the brain, and that movement drives the energy bill. Computers have existed for centuries in mechanical, analog and later digital forms, yet the 1945 Eniac — built to calculate artillery trajectories faster than human computers — shares the same basic operation as today: memory on the outside, compute pulling bits back and forth. That design optimized for speed rather than efficiency, and the assumption is now under review.

Running Physics Directly

Transistor counts kept rising while frequency and single-thread gains stalled and efficiency gains tapered, and lithographic shrink as a free lunch has largely ended. Rao argues the fix is not another layer of abstraction but cutting the middlemen. Even digital ones and zeros are an abstraction over physics; a transistor has intermediate states we engineer away to emulate binary. On top we stack neural networks, each layer lossy and incomplete. The proposal is to take an abstraction of semiconductor physics itself and connect it directly to the neural network, echoing how the brain yields intelligence from neuronal physics without linear algebra or floating point as first-class citizens. The same intuition appears in nature, where birds flock and ants solve problems via simple local rules, studied as dynamical systems where collective behavior emerges from component dynamics.

A hands-on intuition comes from metronomes. Place several on a rigid plank that can roll slightly, and their nudges to the plank gradually lock them into the same phase, even when starting out of sync. Hundreds can synchronize this way, and with different couplings you can get clustered patterns where half share one phase and the other half the opposite. It is a physical system whose outcome emerges from interconnection itself. The question becomes whether such a physical system can compute, and Unconventional AI has tried to answer through generative AI. Its oscillator-based image model UNO, released as an open simulation, showed scale, trainability and useful output, with different state-space trajectories for conditions like airplane, car or bird, and delivered real generated images.

The most consequential recent advance layered on top is sparsity. Connecting ten elements fully needs 100 links, a thousand needs a million, and quadratic scaling quickly becomes painful. Sparsity asks whether many links can be dropped while preserving collective behavior, and the answer turned out better than preservation: the sparser system can be more trainable. It is a rare triple win of higher efficiency, better scalability and stronger performance that also holds in real physical hardware, framed as a payoff for posing the problem correctly.

The most assertive moment was the claim of the first physical dynamical computer ever built. The company started in earnest in January, taped out the design meaning it was sent to fabrication at the start of June, and the chip returned to the lab with early results — the first images asserted to be generated directly on such hardware. The scope is not limited to images; sequence modeling and language tasks are in view. Energy per image is the headline: about 500 nanojoules versus millijoules on a graphics processor, a gap measured in orders of magnitude because almost no information is being shuttled. The team presented this as proof the approach works and emphasized it was shared publicly for the first time on this stage.

Architecturally the lineage is clear: CPUs to GPUs to ever more parallel compute and memory subsystems have all remained von Neumann machines that separate memory and compute and shuttle bits. The new system is described as a dynamical computer where compute and memory are the same thing with no memory interface; each computing element is also a memory. The firm calls this 4D computing, with one dimension of time in dynamics and three physical dimensions from die stacking vertically as well as planar. The near-term product aim is a full rack-level data-center system within about two years, appearing from outside as tokens in and tokens out over a network cable while the interior is entirely different. Porting will happen at the model layer rather than the operator layer, with some compute needed to convert existing models, and even core operations like matrix multiply are reinterpreted not as fixed circuitry but as time-varying behavior where each step is analyzable as current state times a transition. Building the team has meant bridging theorists from dynamical systems — a century-old field — with chip builders who historically did not speak the same language, and creating a Python library that lets users express time-varying stochastic elements instead of a CUDA-like interface.

The closing widens the lens to implications. Intelligence per watt is the metric to optimize toward a thermodynamic ceiling that can never be crossed; mammalian brains sit within one or two orders of it while today's computing sits about ten billion times away. The company aims to hit the limits of two-dimensional lithography within three and a half years and, over the next decade, to beat biology and push compute outward into many small sites rather than a few gigawatt campuses, with claims of environmental, local and adaptive benefits, plus billions of robots that dynamically assemble to solve problems. If disruption reaches a thousand-fold cost change, the invoked Jevons paradox suggests consumption will rise by more than a thousand-fold, potentially creating the largest market humanity has seen.