20VC: Why Robotics Will be 100x Bigger than Coding Agents | Inference Will Be a $10TRN Market: What No One Sees | Why Software Investors are F***** Today with Naveen Rao, Unconventional

What It Covers

Naveen Rao, founder of Unconventional AI (valued near $5B) and former Mosaic ML CEO, argues that hardware and energy infrastructure will define AI's next decade, the inference market will reach $10 trillion within ten years, robotics will dwarf agentic coding as a market, and pure software investors face structural obsolescence.

Key Questions Answered

• •Energy as the binding constraint: US power grid operates at roughly 900 gigawatts but peaks near 1,300 gigawatts — that excess capacity is nearly exhausted. Building one gigawatt costs approximately $50B, more than a semiconductor fab. Realistic new construction adds perhaps 10 gigawatts annually, not the 40 gigawatts commonly cited. Investors and operators should price energy scarcity into infrastructure assumptions now, not after the surplus disappears within two to three years.
• •Hardware investment thesis over software: Efficiency gains from physical transistor scaling have largely stopped. Remaining alpha sits in co-designing models and chip architecture together, optimizing for energy-per-token rather than raw speed. Rao positions this as the dominant investment theme for the next decade. Pure SaaS investors face margin compression as application-layer value migrates toward whoever most efficiently converts energy into intelligence across the full hardware stack.
• •Inference market sizing: Rao projects the inference market reaches $10 trillion within ten years, driven by Jevons paradox — each order-of-magnitude drop in token cost expands the addressable use-case base by more than tenfold. Investors should evaluate inference infrastructure plays not on current token pricing but on the compounding demand unlocked as costs fall, similar to how cheap electricity created entirely new industrial categories rather than simply reducing existing bills.
• •Robotics will be 100x larger than agentic coding: Industrial robotics today is rote execution, not genuine intelligence. Rao expects working prototypes for intelligent physical robots within two to three years and a real market emerging in five to seven years. He invests in the model and data stack powering robotics rather than application-layer companies. The physical world represents an entirely untapped surface area for intelligence deployment, dwarfing the virtual software automation market in total addressable value.
• •Open models will carry 80% of token volume, but revenue splits 50/50: Within three years, Rao estimates open models handle roughly 80% of token traffic while closed frontier models retain 20%. However, dollar spend splits closer to 50/50 because frontier models command premium pricing on high-stakes tasks. The analogy is iPhone versus Android — 20% unit share, 80% of profit. Investors should distinguish between token-volume leadership and margin capture when evaluating model companies.
• •Post-transformer architecture is inevitable: Transformers implement reasoning inefficiently through large context retrieval and multi-sample comparison rather than the structured, sparse, low-energy reasoning biological systems use. Rao expects a post-transformer paradigm within the foreseeable future, driven by co-design between model architecture and physical hardware. Startups pursuing genuinely new architectural paradigms — rather than incremental scaling — represent the higher-risk, higher-return bet, analogous to biotech startups taking drug-discovery risk that large incumbents cannot afford.