The Chip That Could Unlock AGI.
Summary
Naveen Rao’s thesis is that AI’s energy wall requires changing the computer, not merely scaling today’s architecture. U.S. data centers already consume roughly 4% of the national grid, he says, while some estimates call for 400 additional gigawatts over the next decade. The bottleneck is becoming physical: “The current paradigm, as good as it is and as far as it’s taken us, is not going to take us to that level.”
Unconventional AI is starting with a theory of physical learning, not a conventional chip roadmap. Rao wants to use circuits whose own dynamics perform useful intelligence rather than relying on layers of digital abstraction: Bornstein summarizes the idea as “Intelligence is the physics,” and Rao agrees. Human brains operate around 20 watts, while a squirrel or cat may use roughly a tenth of a watt—an efficiency gap large enough to justify revisiting first principles.
Analog is not meant to replace digital computing; it targets stochastic, time-dependent workloads that may benefit from physical dynamics. Rao’s candidates include diffusion, flow, and energy-based models because their dynamics can be written as ordinary differential equations and potentially mapped onto physical systems. His goal is an “intelligent substrate” alongside conventional computation, preserving digital precision for problems that require it.
Rao thinks dynamics might supply the causality missing from current AI, but he repeatedly labels the AGI argument “hand wavy.” His intuition is that systems built from elements with real time evolution will provide a better basis for understanding cause and effect than systems whose basis does not include such dynamics. Today’s models contain intelligence and are highly useful, he says, but remain “nowhere close to AGI” because they still make elementary errors and do not feel like working with a person.
The commercialization test is whether Unconventional can find an intelligence-like physical paradigm within five years and make it manufacturable at massive scale. Rao sees TSMC as a necessary partner for prototyping and eventual scale; he characterizes Google as having everything internally and pursuing lower-risk, continual TPU improvements, while Nvidia has built the dominant programming platform. Unconventional is trying to build “a better substrate than matrix multiply,” though Rao leaves open collaboration rather than assuming direct conflict.
Execution risk is extreme, but Rao sees several partial proofs rather than a blind leap. The brain provides an existence proof, more than 40 years of academic work supplies evidence of promise, and dynamical-systems and neuroscience theory offer pieces that engineers can combine and “sand down.” The first prototype may be one of the larger, possibly the largest, analog chips ever built, making talent across theory, systems, algorithms, analog circuits, and digital circuits central to the thesis.
Deep dive
1. The company begins with physical learning, not a chip specification
Rao corrects the premise immediately: Unconventional is “not a chip company per se.” Its opening work is theoretical—asking from first principles how learning operates in a physical system—because he believes computing’s largely unchanged, 80-year-old architecture can be redesigned.
His path from hardware for wireless and real-time video compression, through a neuroscience PhD, Nervana, MosaicML, and Databricks made crossing boundaries feel natural. In Rao’s older definition, a full-stack engineer understood silicon devices, logic, architecture, low-level software, operating systems, and applications—not merely JavaScript and Python.
Hardware and software are therefore not a natural boundary for him; they are where people choose to draw the line over what they configure. The governing question is where a capability will be consumed, followed by right-sizing the solution to fit the problem.
2. Digital won on scalability, while analog kept the efficiency advantage
Digital computers represent numbers with fixed bits, trading precision error for a general machine capable of simulating anything expressible through arithmetic. Early analog systems were efficient but could not scale because of manufacturing variability; vacuum tubes could reliably represent high or low even when their intermediate behavior was hard to characterize.
Bornstein compares ENIAC’s 18,000 vacuum tubes in 1945 with the number of GPUs used in some large-scale training systems.
A wind tunnel captures the alternative. Instead of numerically approximating fluid dynamics and always being somewhat off, engineers construct a physical analogue whose underlying physics models the process directly.
Intelligence may fit that model unusually well because neural networks are stochastic and distributed, yet today run on precise, deterministic substrates. The brain also shows a striking efficiency target: the human brain uses about 20 watts, while a squirrel or cat may use roughly a tenth of a watt.
Bornstein frames the biological case by saying that “Intelligence is the physics”; Rao agrees, explaining that neural dynamics are mediated directly by chemical diffusion and the physical properties of neurons, without an operating system or API separating computation from matter.
3. Energy scarcity turns architectural efficiency into the binding constraint
Rao says the U.S. holds about 50% of global data-center capacity and commits roughly 4% of its grid to data centers. In 2025, he adds, news articles about summer brownouts began appearing in the Southwest; moving toward 8% or 10% would make the constraint substantially worse.
Power generation can expand, but infrastructure is expensive and slow. Rao cites estimates requiring 400 gigawatts of additional capacity over ten years, versus an expansion rate “on the order of 4 gigawatts per year”; Bornstein adds that even sufficient generation could overwhelm a transmission grid built largely in the 1970s.
Bornstein frames the effort as humanity mobilizing “species-scale resources” to invent the future. Rao says the resulting shortfall means the problem must be rethought, while also arguing that more power generation should be built.
Rao rejects a digital-versus-analog binary. Digital remains appropriate for deterministic numerical problems; analog dynamics may suit retrieval and summarization across multiple inputs, creating an intelligent substrate that complements conventional computation.
4. Physical dynamics could preserve precision while integrating messy reality
Bornstein cites a story that Steph Curry set up a special tracking system so he could ensure the ball was hitting the middle of the rim, not merely going through. In a game, however, position, defenders, shoes, surface, ball tackiness, and sweaty hands make every input unique.
Brains integrate those variables while producing exceptionally accurate behavior. That combination—fuzzy, distributed inputs yielding precise action—is the problem class Rao wants an intelligent physical substrate to address.
Unconventional will begin from current model families rather than discard them. Diffusion, flow, and energy-based models are especially interesting because they contain dynamics, sometimes expressed as ordinary differential equations that might map onto the time evolution of a physical circuit.
Transformers remain valuable because they made GPU constructs work exceptionally well, but Rao sees “no natural law” in their parameterization. He expects mappings between transformer and alternative parameter spaces, arguing that transformers may simply use “lots of parameters” to achieve their results.
5. Time and causality are the speculative bridge toward AGI
Asked whether this path advances AGI, Rao’s answer is deliberately hedged: “Honestly, I do,” followed immediately by “this is hand wavy.” His intuition is that a basis containing time and causality could be better than one that lacks those dynamics or represents time only numerically.
Bornstein notes that mathematical systems can often be reversible in time, whereas the physical world, at least as humans perceive it, generally is not. Rao argues that building from primitives with genuine time evolution might produce systems that understand causation.
Young children provide Rao’s existence hint: they appear to understand that events unfold causally, and people know that sending a particular command to an arm will produce a particular kind of movement. He suspects brains are innately built from causal primitives, though he does not claim to know the mechanism.
Current machines possess intelligence and deliver useful tools, but Rao says they are “nowhere close to AGI.” They still make “stupid errors,” and interacting with them is not yet like working with another person.
6. Manufacturing scale and organizational breadth decide whether the theory matters
Rao sets two milestones: find a paradigm analogous to intelligence within five years, then make it scalable from a manufacturing standpoint by that point. Without the ability to build 10 million devices, the technology cannot address the global energy problem.
TSMC is therefore “absolutely going to be a partner.” Rao says Google has everything internally and, based on what he can see publicly, is pursuing lower-risk, continual TPU improvements for its business. Nvidia has built the platform everyone programs on; Rao cannot say whether it becomes a competitor or collaborator, only that Unconventional seeks “a better substrate than matrix multiply.”
His confidence rests on the brain as an existence proof, more than 40 years of academic work, proof-of-concept devices built by researchers, and developing theory from neuroscience and dynamical systems. Great engineering then combines imperfect pieces: as Bornstein puts it, “That thing doesn’t quite fit—sand it down and make it right.”
The first years will operate as a practical research lab: prove that an idea works before allowing manufacturing objections to close doors. Rao expects a mixed-signal team spanning theorists, model experts, system architects, and analog and digital circuit engineers. The first prototype may be one of the larger, possibly the largest, analog chips ever built.
Rao favors early-career startup breadth and an agency-heavy culture because narrow specialization may adapt poorly to change. Leaders should increase organizational agency and get out of the way when people are passionate about an approach; people should own both the good and the bad, including admitting, “Okay, I screwed up.” His motivation is the belief that changing the computer can make AI ubiquitous: he is “the opposite of an AI doomer” and sees AI as the next evolution of humanity. If Unconventional succeeds, he says, “the world will not forget this for a very long time.”