The 80-Year Bet: Why Naveen Rao Is Rebuilding the Computer from Scratch

AI data centers now consume 4% of the US energy grid, with brownouts already appearing in the Southwest. The industry needs 400 additional gigawatts over the next decade—far exceeding our ability to build new power generation at ~4 GW/year. This creates an existential need to rethink computing substrates rather than just adding more power.

• •US data centers use 4% of national power grid; first brownouts occurred in Southwest in 2024
• •AI demand requires 400 GW additional capacity over 10 years vs. only ~4 GW/year buildout capacity
• •Squirrel brains run on 0.1 watts; human brains on 20 watts—orders of magnitude more efficient than current AI
• •Computing is now energy-bound at global scale for first time in human history

Digital computers implement numerics with fixed precision, making them general-purpose simulators of any physical process. Analog computers use physics itself to perform computation—like wind tunnels modeling airflow. The shift to digital in the 1940s happened because analog systems couldn't scale due to manufacturing variability, but that constraint may no longer apply.

• •Digital computers simulate physics with numbers; analog computers ARE the physics they're computing
• •ENIAC (1945) had 18,000 vacuum tubes—similar scale to modern GPU clusters
• •Early analog computers were more efficient but couldn't scale due to manufacturing precision limits
• •Wind tunnels are analog computers—still used because computational fluid dynamics has inherent errors

Neural networks are inherently stochastic and distributed, yet we run them on highly precise, deterministic substrates. Brains implement intelligence directly in physics with no abstraction layers—no OS, APIs, or memory hierarchies. This suggests building circuits that recapitulate neural network behaviors physically could be vastly more efficient.

• •Intelligence is stochastic; digital computers are deterministic—fundamental mismatch
• •In brains, 'intelligence IS the physics'—no lossy abstractions between stimulus and response
• •Elite athletes achieve millimeter precision with noisy neural networks through dynamic integration
• •Transformers' parameters are artifacts of GPU architecture, not natural laws of intelligence

Diffusion and flow models are naturally expressed as differential equations with time dynamics, making them ideal candidates for analog implementation. Naveen argues that systems with inherent causality and time evolution—rather than time-reversible math—may be necessary for true AGI because the physical world operates with directionality.

• •Diffusion/flow models are ODEs with inherent dynamics—map naturally to physical systems
• •Causality may be key missing ingredient in current AI; children innately understand cause-effect
• •Physical world has time directionality; reversible math may be wrong basis for intelligence
• •Motor control (moving limbs) demonstrates innate causal understanding in biological systems

TSMC is a critical manufacturing partner for scalability—the company needs to build millions of units to address the global energy problem. Google has everything in-house and focuses on incremental TPU improvements. NVIDIA built the current platform, creating potential for either competition or collaboration depending on how the market evolves.

• •TSMC partnership essential for manufacturing scale—need 10M+ units to impact global energy crisis
• •Five-year goal: prove analogous paradigm works, then ensure manufacturing scalability
• •Google optimizing TPUs for margins with lower-risk incremental improvements
• •NVIDIA relationship unclear—building 'better substrate than matrix multiply' but collaboration possible

Unconventional AI operates as a practical research lab with open-ended exploration in early years, deliberately avoiding premature manufacturing constraints. The first prototype will likely be the largest analog chip ever built—a mixed-signal design requiring expertise across theory, AI systems, architecture, and circuit design.

• •First prototype targeting 'probably the largest analog chip people have ever built'
• •40+ years of academic research provides existence proofs across different technology nodes
• •Team needs: AI systems engineers, energy-based model theorists, dynamical systems experts, analog/digital circuit designers
• •Culture: prove it works first, engineer manufacturing trade-offs later—avoid premature optimization

Startups provide breadth early in careers that pays long-term dividends—Naveen's ability to think across the full stack came from building hardware, software, and applications early on. He prioritizes decisions that increase organizational agency over short-term optimal choices, giving passionate people room to own both successes and failures.

• •Early career breadth (hardware + software + applications) enables cross-stack thinking later
• •Big companies optimize for deep specialization; less valuable when facing future change
• •Leader's role: make decisions that increase team agency, not just short-term optimal outcomes
• •Hard problems attract smart people; giving them ownership creates long-term organizational strength

Naveen views AI as humanity's next evolution, enabling deeper collaboration and understanding rather than replacing human experience. Success here would be written in history books—a rare generational opportunity. He's explicitly anti-doomer, believing positives vastly outweigh negatives, but achieving ubiquity requires fundamentally changing the computer.

• •AI enables humans to 'collaborate and understand each other and the world in much deeper ways'
• •Current digital paradigm won't achieve AI ubiquity—fundamental substrate change required
• •Success would be 'written in history books' as generational impact, not forgotten in 5 years
• •Technology always has negatives, but AI's positives 'so far outweigh' the risks